Who Owns the Decision?

Editorial and evidentiary note

This report develops the preceding article, When the AI Scribe Starts Choosing the Diagnosis, into an industry-wide analysis. The earlier article is retained separately in the companion preservation appendix; its presence is not an assertion that every earlier statement is independently established. The expanded report distinguishes the physician’s reported experience, visible interface evidence, vendor representations, research findings, existing law, pending or vetoed legislation, judicial holdings, allegations in litigation, and policy proposals. Those distinctions are necessary to make a forceful argument that can withstand scrutiny.

The record reviewed includes the author’s own contemporaneous account and notes, the publication’s earlier analysis, and the primary authorities in the source register. Patient-identifying clinical information is not evidence for this public-policy article and is not reproduced; no image of a patient record is published. First-person clinical passages are Dr. Gill’s account. This is a policy analysis, not a legal opinion.

The central proposition

Healthcare AI is not one technology and does not have one economic purpose. A system can save a physician time while increasing the employer’s revenue. It can improve a patient’s access while displacing an administrative job. It can reduce a payer’s expense while increasing the provider’s unpaid work. It can generate accurate information while making disagreement operationally difficult. These possibilities can coexist. The proper inquiry is therefore not simply whether AI is beneficial, but which function is being automated, who controls it, who captures the benefit, who absorbs the new work, and who is answerable when it fails.

Current product descriptions document a transition beyond transcription. Major vendors describe extracting diagnoses, drafting orders, supporting coding, coordinating workflows, and connecting clinical records with financial operations. Those descriptions establish commercial direction; they do not establish that every advertised feature is deployed, autonomous, clinically safe, or unlawful. [S001–S010]

The policy position developed here is that accountability should follow actual control and foreseeable risk. The physician’s signature should not erase responsibility for negligent development, integration, contracting, staffing, or implementation. Equally, the involvement of software should not excuse unreasonable clinical conduct. Neither mandatory adoption nor voluntary enthusiasm is a substitute for evidence.

Part I — From a human encounter to an automated clinical transaction

1. A patient encounter is not a command line

A patient may describe symptoms while also describing fear, an argument at home, an unaffordable prescription, a recent death, or the practical impossibility of attending an appointment. The conversation can move between medicine and ordinary life without announcing the transition. A clinician may repeat a concern to show that it has been heard, discuss a dangerous possibility to explain why testing is necessary, or describe a treatment that the patient ultimately declines. These examples illustrate why clinical meaning cannot be reconstructed safely by treating every medically recognizable phrase as an instruction.

The ethical starting point is that the patient is not present to produce clean training data. The clinician is not present to perform a script for an extraction engine. Clear communication matters, but the obligation to communicate clearly does not justify making patients suppress uncertainty, family context, distress, or disagreement so that software performs better. A well-designed tool should accommodate reasonable variation in human conversation and openly signal what it cannot resolve.

There is also a limit to the criticism. No scribe, human or artificial, can reliably document an unspoken decision merely because it exists in the physician’s mind. The clinician must express or enter the actual assessment and plan. The concern is not that AI fails to read minds. It is that the system may convert conversation into apparent decisions that were never made, then place the burden of reversing that conversion on the clinician.

The distinction can be illustrated by three sentences. “We should check whether your existing referral was scheduled” concerns the status of an earlier action. “We will place a new referral today” authorizes a prospective action. “A referral might be appropriate after the test results” is conditional. A safe workflow needs to preserve all three meanings rather than collapse them into the same order candidate.

Human-centered design also means preserving the ability to pause. A patient discussing trauma, reproductive concerns, addiction, or family violence may want part of a conversation off the recording. The clinical task is then to provide appropriate care and document what is necessary, not to insist that the technology receive an uninterrupted stream. NHS ambient-scribing guidance expressly recognizes explaining the technology and respecting objections. California recording law presents its own consent questions; the legal foundations are discussed later. [S068, S078]

The measure of success should be faithful support for the encounter, not how efficiently the encounter can be made to fit the product.

2. The initiating case: what is established, reported, and still unknown

On October 27–28, 2026, I encountered problems after expanded functionality appeared in my Epic–DAX workflow. My account is that an existing-referral discussion produced a proposed new referral, a unilateral finding became a bilateral diagnosis, and differential possibilities appeared as encounter diagnoses. I also encountered the removal of associated narrative when rejecting a proposed diagnosis. These are reports of specific discrepancies, not a claim that patient injury has been established.

The interface itself provided unusually concrete evidence. It displayed the “Heard in Visit Recording” panel, a proposed encounter diagnosis, several infectious-disease referral matches, and a follow-up-related suggestion. Next to “Don’t use,” the interface states: “Associated ambient A&P note will be deleted.” That text supports the concern that rejection of a structured diagnosis is coupled to removal of its linked assessment-and-plan content. It does not prove that the raw audio or transcript is permanently erased from every backend system.

The same screen also showed selection and signing controls. It therefore does not establish that an order was autonomously signed, transmitted to a specialist, or acted upon. These distinctions strengthen rather than weaken the case: the observed problem is a clinically meaningful error-prone review pathway, even where a final human approval remains necessary.

My understanding is that the organization requires recording and has not offered a usable alternative charting route. That is a material allegation about deployment, but it remains an account to be checked against written policy, permissions, configuration, training, and actual available workflows. A requirement to use Epic is not necessarily a requirement to accept AI diagnoses. A requirement to record is not necessarily proof that manual diagnosis entry is blocked. The investigation should identify exactly which restriction exists.

A contemporaneous internal report of such problems is notice; it is not a regulatory finding, an adjudication of negligence, or immunity for future mistakes.

This distinction between evidence and inference is crucial for public testimony. The strongest statement is not “the computer practiced medicine illegally.” It is: the deployed workflow produced discrepancies in clinical meaning, complicated rejection of incorrect content, and warrants a documented safety and function-specific regulatory review.

FIGURE 01 · From conversation to consequential action

3. The transition did not begin yesterday

In March 2026 I noted a separate concern about shared-office DAX use. The reported situation involved overlapping conversations, telemedicine speakerphones, and multiple clinicians using recording tools in the same area. That is a different failure mechanism from a mistaken diagnosis suggestion: it begins before clinical interpretation, when the wrong speech can enter the system. The earlier account is contextual evidence, not an independently audited incident report.

I have also held a favorable view of AI-assisted documentation as a way to finish charts more efficiently. This matters. The position developed here is not that physicians should never use AI, nor that every earlier positive expectation was unreasonable. It is that the balance can change when a documentation aid gains additional authority, when implementation conditions change, or when an optional aid becomes a mandatory workflow.

The same standard must apply to clinician-selected tools. A physician who voluntarily uses a general-purpose model to improve a draft note still needs to preserve source fidelity, prevent unsupported diagnoses and examination findings, protect patient information, and review the result. Voluntariness can improve practical control; it does not make a tool clinically reliable or legally exempt. A critique aimed only at employer-selected AI, while excusing identical problems in personally selected AI, would be vulnerable and inconsistent.

Corporate-practice analysis teaches that the question is who actually controls clinical decisions, not which corporate label appears on the record. A screen can display the physician’s name while the practical choices are constrained by defaults, mandatory pathways, or institutional performance rules. Conversely, the mere presence of software does not establish that professional control has been displaced.

The longitudinal question is whether each expansion in capability was accompanied by a corresponding expansion in validation, consent, explanation, staffing support, and accountability. A product that was acceptable as a draft generator last year should not receive automatic approval for order generation this year merely because it retains the same brand.

What connects these earlier concerns is a set of specific themes: cross-capture, documentation fidelity, workload, independent judgment, and the need to distinguish the entity with formal responsibility from the party with operational control.

4. The technical chain: where a correct sentence can become a wrong action

Ambient clinical systems can involve several separable operations: acquiring sound; transcribing words; distinguishing speakers; summarizing content; identifying medical entities; resolving chronology, negation, and uncertainty; linking statements to diagnoses; selecting EHR catalog entries; and presenting or releasing downstream actions. This is an analytical map of possible functions, not a claim that every vendor uses the same internal architecture.

A failure at the beginning can propagate. If another patient’s conversation is captured, an excellent summarizer can faithfully summarize the wrong encounter. If the words are correct but speaker attribution is wrong, a family member’s history can become the patient’s diagnosis. If attribution is correct but uncertainty is lost, a rule-out discussion can become an established condition. If the clinical concept is correct but the catalog match is wrong, the wrong side, dose, specialty, or destination can be selected.

The integration layer is therefore as important as the language model. A proposed “infectious disease referral” may match several orderable items with different routing, locations, prerequisites, or insurance workflows. A technically valid order is not necessarily the clinically intended order. The incident should be traced across the whole chain rather than reflexively assigned to “the AI” or “user error.” Microsoft itself describes local EHR configuration as relevant to the order-management workflow. [S001]

A source-linked excerpt can help identify why a suggestion appeared, but it is not always an explanation of the reasoning. Showing that the word “cancer” was spoken does not establish that cancer was diagnosed. Showing a referral-related sentence does not establish that a new referral was intended. Meaningful review needs the relevant context, not merely a keyword match.

The system should preserve distinct states: proposed, selected, accepted, pending, signed, transmitted, canceled, and completed. An audit event showing that a clinician opened a panel should not be treated as acceptance of every item in it. Similarly, a signed encounter note is not necessarily evidence that every pending referral was intentionally released.

The practical remedy is a traceable chain of provenance: what input was used, what the software proposed, what the clinician changed, what the EHR committed, and what downstream recipient received. This is a proposed investigation and design standard. It is more useful than a generic assurance that a person remained “in the loop.”

5. What the principal vendors actually say

Microsoft’s DAX documentation describes extracting verbal instructions into potential medication, testing, procedure, referral, and follow-up orders. Its diagnosis-retrieval documentation separately describes generating assessment-and-plan diagnoses from ambient recordings and comparing them with Epic diagnoses. Its release material describes diagnosis-aware note behavior and configuration requirements. These are different functions and should not be treated as interchangeable proof of what was enabled in one clinic. [S001, S003, S004]

Epic publicly describes Art as capturing conversation, drafting notes, and extracting diagnoses and orders. Abridge’s October 2026 release describes outpatient orders appearing in Epic’s Heard in Visit Recording workflow for review and signature. Abridge also describes safeguards, including handling changing or retracted order intentions. Those statements deserve consideration; the case is not strengthened by pretending vendors advertise completely uncontrolled prescribing. [S005, S006]

Oracle’s February 2026 announcement describes ambient order drafting using information such as earlier orders, history, physician favorites, and organizational preferences. Its broader AI-agent page spans clinical, administrative, and financial operations, but explicitly labels some capabilities as planned. A roadmap is evidence of commercial strategy, not proof of present availability. [S007, S008]

The distinction between clerical extraction and clinical inference is fundamental. Turning an explicit physician instruction into a draft order may be a useful reduction in manual entry. Inferring a diagnosis or selecting a plan that was never adopted is a different function. A sensible evaluation tests both whether the system can help and whether it can avoid going beyond the clinician’s intent.

The vendor’s strongest defense is that outputs are drafts, the clinician retains control, and review is required. That defense has real force where the interface makes uncertainty visible, preserves source material, supports easy rejection, and prevents unintended release. It becomes less persuasive as a complete answer where rejection destroys useful documentation, errors are difficult to detect, or the employer reduces the time available for review.

No product should be classified merely by a reassuring name such as “scribe,” “assistant,” or “copilot.” Nor should it be classified merely by an alarming description such as “robot doctor.” Function, intended use, implementation, and consequences are the relevant objects of examination.

6. Four clinical-intent failures, not four cosmetic defects

The reported laterality error changes the anatomical meaning of a record. Its actual severity depends on what was documented, whether the physician corrected it, and whether anyone relied upon it. It is not appropriate to assign a universal “high-risk” score without that context. Nevertheless, unilateral-to-bilateral conversion is categorically different from a punctuation error because it can alter what a future reader believes was examined or diagnosed.

The conversion of a differential into a visit diagnosis changes diagnostic certainty. A patient may leave believing a disease has been confirmed when it was only being considered. A future clinician may treat the label as established. A coding workflow may carry it into a claim. Current outpatient ICD-10-CM guidance does not permit coding uncertain diagnoses as established in the same way allowed in specified inpatient circumstances. The setting therefore matters to both clinical and coding interpretation. [S086]

The referral-status error changes the temporal and action meaning of the conversation. “Has it happened?” becomes “make it happen again.” A draft that is rejected may be a near miss or a workflow hazard, depending on the institution’s definitions. An unintended referral that is transmitted and acted upon requires a different review. The report should identify the actual stage rather than use “ordered” ambiguously.

The linked-deletion issue changes the cost of correction. It can require the clinician to preserve text manually before rejecting a wrong diagnosis. This creates a second opportunity for error: the original inference may be corrected while counseling, reasoning, or follow-up instructions are lost. The on-screen warning substantiates the interface behavior, not the precise backend retention behavior.

Together these examples suggest testing four distinct properties: anatomical fidelity, uncertainty preservation, action-intent fidelity, and correction integrity. They do not prove a single common algorithmic cause. Each may arise from the model, structured mapping, template logic, or local configuration. A meaningful vendor response should investigate those possibilities instead of assuming that more training or clearer speech will necessarily solve every defect.

The clinical-safety case is strongest when it remains specific. A missed finding, a false finding, and an incorrectly activated order demand different controls. The phrase “AI made an error” is too broad to guide corrective action or allocate responsibility fairly.

7. Meaningful override is more than a visible button

A clinician can possess formal authority while lacking practical control. Consider an interface in which accepting suggestions completes the encounter in one step, while rejecting them requires copying text, locating another field, rebuilding the assessment, re-associating orders, and resolving warnings. The clinician still has a choice in a literal sense. But the software has changed the effort required to exercise that choice.

That asymmetry is a governance issue even when nobody deliberately designed it to coerce. Default acceptance can become an operational norm through time pressure rather than an explicit command. The proper question is not only whether rejection is permitted, but whether it is understandable, reversible, and realistically usable during ordinary care.

A meaningful override should preserve the valid parts of the encounter, clearly distinguish rejected content from signed content, avoid silent regeneration of rejected suggestions, and permit clinically appropriate completion without adopting an unwanted plan. These are proposed usability standards, not a claim that current California law specifies each interface control.

The independent-review concept in federal CDS law and FDA guidance is relevant but not identical to interface usability. Regulatory classification asks whether the applicable statutory conditions are met. Human-factors evaluation asks whether clinicians can exercise control safely in actual practice. A product may fall outside FDA device oversight and still have a poor workflow; a regulated product may still require local usability assessment. [S025–S027]

An employer should be asked to demonstrate the alternative in a test encounter: reject the diagnosis, preserve the narrative, enter a different assessment, remove an unwanted referral, and close the encounter. Demonstration is more informative than an abstract promise that “the physician is always in charge.” It also permits fair identification of a training gap where a safe route genuinely exists.

The same exercise should be repeated after substantial updates. A one-time training session cannot establish that later releases retain equivalent control. The acceptance criterion should be the clinician’s demonstrated ability to complete an accurate record without unintended clinical actions, not merely completion of a tutorial.

FIGURE 02 · A meaningful override pathway

8. Patient consent, shared rooms, and the right not to be optimized

Recording introduces a relationship question before it introduces a model-performance question. Patients should understand whether a device is listening, whether the recording is retained, whether a third party processes it, and whether use extends beyond creating their record. Consent to treatment, agreement to a privacy notice, consent to recording, and authorization for unrelated secondary uses are not automatically the same event.

California’s confidentiality and recording rules require fact-specific analysis. In Gruber v. Yelp, the appellate court rejected an attempt to avoid the consent requirement by recording only one side of a communication. It was not a medical-AI case, but it illustrates why recording design cannot be assumed to eliminate all consent issues. HIPAA business-associate arrangements address a different legal layer. [S068, S054]

Shared-office use raises additional practical questions. A clinician may obtain consent from the intended patient while inadvertently capturing another clinician’s conversation. An interpreter or family member may become part of the recorded encounter. A telemedicine speakerphone can make boundaries less visible. The technical remedy may involve room design, headsets, microphone directionality, scheduling, or different recording practices; the legal and clinical remedy begins with recognizing that the problem is not solved merely by obtaining one signature.

A patient-facing explanation should be intelligible rather than promotional: what the system does, what it does not do, who reviews the result, and how to pause or decline where applicable. It should not promise perfect privacy, accuracy, or confidentiality beyond actual contractual and technical protections.

The proposed policy is a workable non-recorded route that does not penalize patients or leave clinicians unable to document care. The legal entitlement to a particular alternative may vary by setting, so the policy argument should not be disguised as a universal existing statutory command. It is nevertheless difficult to defend a patient-centered system that makes a sensitive disclosure conditional on continuous recording.

A human encounter is not an inefficient data pipeline to be corrected. Its complexity is part of the clinical work. The software’s limitations should be managed around that fact, not used to redefine what patients are allowed to say.

Part II — An industry-wide map of automation

9. The business of medicine begins before the patient is seen

The commercial reach of healthcare AI is broader than diagnosis. The same ecosystem can touch advertising, patient acquisition, call answering, eligibility checks, scheduling, intake, documentation, testing, care coordination, coding, claims, collections, utilization review, quality reporting, staffing, and research. Oracle’s product and roadmap descriptions explicitly connect clinical and financial functions; specialist vendors concentrate on coding, revenue integrity, and patient communication. These are documented areas of commercialization, although individual capabilities vary in maturity and availability. [S008–S014]

An industry map should therefore follow the patient’s journey rather than the boundaries of a single software product. A scheduling algorithm may influence when the patient obtains care. An intake classifier may influence the queue in which the complaint appears. A chart summary may influence what the clinician notices. A coding engine may determine how the encounter becomes a claim. A payer’s system may determine whether that claim is paid or whether the next service is authorized.

No one model needs to be “in charge” for the combined process to constrain care. Several small automated decisions can create a narrow path that becomes difficult for the patient or clinician to change. This is a systems hypothesis to test through workflow mapping, not proof that every institution has adopted such a structure.

Administrative automation can genuinely help. A reliable reminder may prevent a missed appointment. Correct eligibility information can reduce financial surprises. Better referral tracking can identify a patient lost to follow-up. The policy goal should preserve these benefits while requiring a human route for exceptions, especially where an apparently administrative decision has clinical consequences.

A useful classification has three dimensions. First, what is the output: information, recommendation, or action? Second, what is the consequence: clerical inconvenience, financial burden, delayed access, or direct treatment? Third, who can contest it: a staff member, a licensed clinician, the patient, or nobody visible? A tool labeled “administrative” can be consequential if it closes the only route to care.

The proposed governance unit is therefore the decision pathway, not merely the product license. A hospital may buy five separate products, but the patient experiences one pathway. Accountability should be capable of following the pathway across corporate and technical boundaries.

FIGURE 03 · The healthcare AI value chain

10. Primary care, internal medicine, and the longitudinal record

Primary care exposes a particular challenge: one encounter can concern several active problems, old diagnoses, preventive issues, specialist plans, and social barriers. A discussion of a historical medication reaction must not become a new allergy without review. A resolved condition must not silently regain active status. A specialist’s provisional conclusion should not automatically become the primary clinician’s confirmed diagnosis. These are illustrative test cases, not reports of additional events in the initiating incident.

The longitudinal record can amplify both benefit and error. A good summary can help a clinician orient quickly. A wrong summary can repeatedly make the same inaccurate assertion appear familiar. When later notes reuse earlier AI-generated text, repetition can be mistaken for corroboration. The design question is whether the source and date remain visible enough for a clinician to distinguish a new finding from a repeated assumption.

The chart also contains disagreement. One specialist may favor one diagnosis while another favors a different explanation. A generative system may be tempted to resolve that disagreement into a coherent narrative, but coherence is not always accuracy. The appropriate note may need to preserve uncertainty rather than select a single story. A safe summary should identify conflicting information instead of concealing it behind fluent prose.

For older adults and patients with multimorbidity, the relevant question is often not which action can be generated, but which action fits the patient’s goals and treatment burden. An additional test or referral may be technically appropriate in isolation while poorly aligned with the larger plan. These are examples of why the clinician’s contextual judgment cannot be reduced to catalog matching.

AI can assist with reconciliation, but assistance should make discrepancies easier to review rather than make the record appear artificially settled. The clinician should be able to retain “history unclear,” “patient declines,” “awaiting records,” or “diagnosis not established” as legitimate outcomes, not errors to be eliminated.

A useful local evaluation would sample complex encounters, multilingual visits, medication changes, follow-up visits with no new orders, and encounters involving multiple historians. Success should be measured by fidelity to the clinician’s intended assessment and the patient’s actual plan, not just whether a complete-looking note was produced.

11. Emergency care, critical care, surgery, and time-sensitive decisions

In emergency and critical care, a system may encounter rapidly changing information. A diagnosis considered at triage may later be excluded. A drug may be discussed, prepared, withheld, or administered; those states are not interchangeable. A care plan can change between the beginning and end of an encounter. The proposed test is whether the system represents the final clinically relevant state while preserving the sequence needed to understand the decision.

Time sensitivity changes the feasibility of review. A physician may have an opportunity to review a draft outpatient note later in the day. A real-time alert or proposed intervention may require action before detailed verification is possible. FDA guidance recognizes the significance of whether the professional can independently review a recommendation’s basis. Not every emergency-department function is inherently time-critical, but some uses plainly place different demands on oversight. [S026, S027]

Surgical and procedural workflows add side, site, technique, consent, and scheduling risks. A discussion of possible surgery must not be confused with a decision to proceed. An alternative procedure mentioned during informed consent must not become the scheduled operation. A canceled intervention should not remain active because an earlier segment of the recording was interpreted without the later correction.

Anesthesia and medication-administration settings require particular care in separating a suggestion from a documented act. An AI-generated statement that a medication was given can be dangerous even if no electronic prescription was issued. The safety question is therefore wider than order transmission: does the record falsely represent what actually happened?

These examples do not prove that ambient AI is unsuitable for acute care. They show that testing only straightforward outpatient conversations is insufficient to validate every setting. Emergency, procedural, and inpatient implementations need their own scenarios, users, and release controls.

The proposed threshold for deployment should increase with consequence and reduce with reversibility. A harmless formatting suggestion and a time-sensitive treatment recommendation should not be subject to the same acceptance process merely because both appear in the same EHR. A unified interface must not obscure differences in clinical authority.

12. Oncology, psychiatry, obstetrics, pediatrics, and sensitive context

Some specialties make uncertainty especially consequential. In oncology, a suspected lesion, a pathological diagnosis, a stage, a recurrence, and a treatment response are different assertions. In psychiatry, a patient’s intrusive thought, a hypothetical discussion, a historical event, and a current intention are different assertions. In obstetrics, the same conversation may concern the pregnant patient, fetus, partner, or prior pregnancy. In pediatrics, caregivers may provide contradictory histories. These are illustrative distinctions that validation should preserve.

The clinical encounter may also include decisions that are not reducible to a guideline’s preferred option. A patient may choose surveillance rather than intervention, decline a medication because of earlier adverse effects, or prioritize quality of life. A generated plan should not treat such choices as incomplete documentation that must be repaired into a more conventional treatment pathway.

Behavioral-health recording deserves a separate assessment of trust and confidentiality. A patient’s willingness to disclose may change when a recording device is present. This is a concern to evaluate, not a quantified finding about every patient. The alternative should be a deliberate consent and documentation policy rather than assuming that universal recording is neutral because it is technically convenient.

Palliative care illustrates the limits of measuring productivity by completed tasks. An encounter that appropriately results in fewer interventions may represent excellent care. A system optimized to produce diagnoses, orders, and follow-up actions should be tested against a valid plan of supportive care, watchful waiting, or no additional treatment.

Similarly, specialty templates can support quality while also flattening unusual circumstances. A standardized note may look complete while missing the reason the clinician departed from a usual pathway. Preserving that reasoning is particularly important when future reviewers may compare the physician’s action with an AI-generated alternative.

The proposed safeguard is not that every discussion requires lengthy defensive documentation. It is that clinically material uncertainty, patient preference, contraindications, and reasons for disagreement must remain easy to record and easy to find. The system should support a reasoned exception as a normal feature of medicine, not as a suspicious deviation requiring unnecessary justification.

13. Radiology, pathology, and the difference between assistance and substitution

It is inaccurate to treat “AI in radiology” as one intervention. Worklist prioritization, image reconstruction, measurement, abnormality detection, comparison with earlier studies, draft report generation, and autonomous screening have different intended uses and regulatory implications. A report assembled from a radiologist’s dictated findings is not the same function as software identifying a finding in an image. [S026, S084]

The claim that radiologists generally do nothing more than sign computer-generated reports is not established by the evidence reviewed. The stronger issue is whether particular implementations create conditions in which substantive interpretation becomes perfunctory. That question requires actual workflow evidence: case volume, available review time, discrepancy rates, escalation practices, and the extent to which AI output is visible before independent interpretation.

Imaging also creates a special form of attention risk. A highlighted lesion may be useful, but it can direct attention away from other findings. A negative AI result may be reassuring without being conclusive. These are failure modes to test, not a claim that the presence of an overlay necessarily worsens performance. Performance should be evaluated for the human–software combination, not only for the algorithm in isolation.

Pathology raises parallel questions about specimen quality, sampling, staining, scanner differences, and interpretation. An image classifier may recognize a pattern while lacking information that changes its clinical significance. A draft report can be accurate in wording yet overstate certainty or omit limitations. The proposed validation standard includes preservation of uncertainty and the conditions under which a specialist must review the primary material.

Bounded autonomous systems can be legitimate. FDA’s IDx-DR authorization concerns a defined retinal-screening use, not general ophthalmic practice. That example supports a function-specific approach rather than the proposition that every autonomous result is unlawful. [S085]

Professional governance programs such as ACR’s ARCH-AI emphasize institutional oversight. Their existence is not proof that all deployments are safe; it is evidence that safe deployment requires more than purchasing software. The question for radiology and pathology is not whether human labor can ever be reduced, but whether any reduction is supported by validated scope, maintained expertise, and a dependable route for cases outside that scope. [S084]

14. Dentistry: diagnosis, treatment planning, and the insurer’s second interpretation

Dental AI can be used at several points: image analysis, caries or periodontal assessment support, measurement, treatment communication, charting, coding, and payer review. The ADA’s standards work explicitly addresses both clinical and administrative uses and calls for independent validation datasets. Expert annotation is therefore a legitimate development activity; a dentist helping label images is not, by that fact alone, participating in an improper denial system. [S081]

The relevant concern is how the trained system is later deployed. A radiograph is one component of a clinical assessment. Symptoms, examination, history, restorability, functional needs, and patient preferences may matter. A model’s image-based output should not be represented as a complete determination of treatment necessity when the function was not validated to resolve that question.

There are three separate propositions: a feature is visible on an image; a treatment is clinically indicated; and a benefit plan pays for that treatment. An insurer may apply a coverage exclusion even where treatment is clinically appropriate. Conversely, a covered procedure is not necessarily needed by every patient. Software that collapses these questions can produce a superficially objective decision that is conceptually wrong.

FDA clearance of a specific dental image-analysis product does not authorize every later insurance use. Overjet Dental Assist has a device-specific 510(k) record. Overjet also publicly describes a Review Pass workflow that confirms criteria or routes cases for review and says it does not deny claims. Fair analysis should acknowledge that representation instead of imputing a denial function merely because AI is involved. [S082, S083]

A meaningful payer audit would ask what inputs were reviewed, whether the treating dentist’s explanation was considered, whether image quality was adequate, whether a qualified reviewer could disagree, and whether the patient received a comprehensible reason and appeal route. It would also examine whether the same image is interpreted differently depending on whether the customer is a practice seeking treatment acceptance or a payer seeking expenditure control.

That last question is a proposed conflict-of-incentives inquiry, not an allegation that a named vendor manipulates findings for different buyers. The appropriate evidence would be contracts, model specifications, validation protocols, thresholds, and actual case comparisons—not a generalized suspicion that all insurance AI is designed to deny care.

15. Nursing, pharmacy, rehabilitation, podiatry, and chiropractic care

AI-related labor debates often focus on physicians while overlooking the people who notice deterioration, reconcile discrepancies, explain instructions, coordinate care, and identify that a patient has not understood. A voice agent that makes follow-up calls may automate one nursing-related task. It does not thereby demonstrate equivalence to bedside assessment, professional accountability, medication administration, or the full range of nursing judgment.

Hippocratic AI markets non-diagnostic voice applications at prices as low as nine dollars per hour and describes escalation safeguards. That is a commercial comparison point, not a conclusion that a complete nursing role can safely be replaced at that price. The appropriate unit of comparison is the bounded task, including human escalation and supervision costs. [S011, S012]

Pharmacy workflows distinguish prescription drafting, authorization, dispensing, counseling, and administration. A medication mentioned in a visit is not automatically a prescription. An algorithmic refill pathway needs defined eligibility, exclusions, identity checks, clinical review triggers, and lawful authority. Utah’s narrow Doctronic pilot demonstrates the importance of specifying phases and safeguards rather than speaking as though every AI system can prescribe generally. [S071]

Physical, occupational, and speech-language therapy involve goals, observed function, adherence, and patient response over time. A standardized home-exercise recommendation is not the same as a patient-specific decision about safety or progression. Podiatry adds site, laterality, wound, vascular, neurologic, and procedural context. Chiropractic care raises scope, contraindication, and referral questions. These are examples of domain-specific requirements, not a claim that one technology is currently making every listed decision.

For every profession, the system should identify the user’s lawful role rather than treating any person with access to an interface as authorized to approve any proposed action. A medical assistant’s acknowledgment should not be mistaken for prescribing authority. A nurse’s completion of a structured assessment should not silently become a physician diagnosis. A therapist’s functional observation should not be converted into a disease label outside the encounter’s purpose.

A responsible policy must protect the entire care team’s judgment, not merely the physician’s final signature. It should also recognize that removing administrative staff can push their work onto licensed professionals, producing nominal labor savings while worsening clinical capacity.

16. Research, pharmaceuticals, devices, and the hidden labor behind AI

AI is also being developed for research tasks, trial recruitment, evidence extraction, safety monitoring, manufacturing, and drug-development decisions. FDA’s January 2025 draft guidance proposes a context-of-use approach to evaluating the credibility of AI-generated information for drug and biologic submissions. Joint FDA–EMA principles issued in January 2026 emphasize human-centered, risk-based lifecycle management. Neither document means that an AI-generated hypothesis is itself an approved treatment. [S030, S031]

The economic question here differs from the clinic’s scheduling question. Faster discovery could reduce some development costs, but lower development cost does not logically guarantee lower patient prices. Price, reimbursement, exclusivity, competition, manufacturing, and commercial strategy are separate variables. The report does not estimate an industry-wide price effect because the sources reviewed do not establish one.

There is also human work inside apparently automated systems. Data selection, annotation, quality review, adjudication of disagreements, software integration, safety evaluation, and ongoing monitoring must be performed somewhere. The labor may move from a hospital payroll to a vendor or contractor rather than disappear. An honest jobs analysis should track that relocation and the working conditions of the people performing it.

Synthetic data and simulated patients can be useful development tools, but they are not automatically substitutes for representative clinical validation. A system can perform well on generated examples that reflect its designers’ assumptions while failing on real exceptions. The proposed standard is to document where evidence comes from and what it can support, rather than presenting every benchmark as a patient-outcome study.

Research governance must also distinguish care from experimentation. A clinician’s ordinary use of a tool, a quality-improvement evaluation, and a study designed to generate generalizable knowledge may involve different institutional processes. The exact requirements depend on the activity and jurisdiction; simply calling a rollout a “pilot” does not resolve them.

The industry-wide lesson is that automation can accelerate production of evidence as well as care. That increases the importance of provenance, independent evaluation, and clear separation between a model’s plausible output and a verified result. Speed should make high-quality investigation easier, not make the threshold for proof disappear.

17. What employment evidence can—and cannot—establish

The strongest defensible claim is that healthcare AI is designed to automate identifiable tasks, and some of those tasks are central to existing jobs. It is not defensible, on the evidence reviewed, to assign a precise number of healthcare jobs already eliminated by ambient AI or to claim that every exposed occupation will disappear.

The Bureau of Labor Statistics currently projects medical-transcriptionist employment to decline by four percent from 2025 to 2035 and identifies technological change as relevant. This is a projection for an occupation, not a count of layoffs attributable to one product. The ILO’s 2025 analysis similarly distinguishes exposure from redundancy and concludes that transformation is more likely than total replacement for many jobs. [S021, S022]

Several paths can follow task automation. An employer can redeploy staff to unresolved needs, leave vacancies unfilled, reduce overtime, eliminate positions, outsource remaining exceptions, or increase output with the same workforce. The same product can support different choices in different organizations. Adoption alone therefore does not identify the employment outcome.

Job quality can deteriorate even where headcount remains stable. Remaining staff may receive more cases, fewer straightforward tasks, and a higher concentration of exceptions. The work can become continuous correction and escalation. Conversely, removing repetitive work can improve job quality where staff retain autonomy, adequate time, and training. These are plausible organizational outcomes that should be measured rather than assumed.

Radiology illustrates why productivity and employment can move differently. A tool can increase throughput per specialist while rising demand sustains or increases total employment. A profession’s aggregate growth does not mean no individual workers are displaced; an individual layoff does not establish that the profession is disappearing. BLS analyses emphasize the interaction of technology and demand. [S023]

The proposed workforce-impact statement should identify which tasks will be automated, which roles are expected to change, what retraining is funded, how exceptions will be staffed, and whether expected savings depend on layoffs, attrition, or increased volume. It should not hide a staffing decision inside a software procurement announcement framed only as clinician wellness.

FIGURE 04 · Task automation has more than one employment outcome

18. Deskilling, surveillance, and the risk of turning expertise into exception handling

A profession can lose practical authority without losing its title. If the system routinely proposes the diagnosis, plan, and order set, the clinician’s work may shift toward confirming or rejecting a preconstructed account. Whether that shift improves or degrades practice depends on the quality of the suggestions, the clinician’s expertise, and the conditions of review. It should not be treated as inherently harmless merely because a license remains required.

Training deserves particular attention. A novice who repeatedly accepts a generated assessment may have fewer opportunities to construct a differential independently. An experienced clinician may become less practiced at tasks delegated over time. These are plausible deskilling mechanisms, not quantified outcomes established by the sources in this report. They justify evaluation of independent performance, supervised training, and periodic work without the tool where appropriate.

Workplace surveillance is a separate use of the same data. Encounter duration, editing patterns, acceptance rates, speech, and after-hours activity could be used to evaluate employees. A low AI-acceptance rate might reflect difficult cases, careful review, an inaccurate model, or resistance to a useful tool. Treating it as a direct productivity measure would collapse several explanations into one score.

California’s civil-rights regulations address automated systems used in employment decisions and do not excuse discrimination because software produced the recommendation. The relevant questions include selection criteria, disability-related inquiries, disparate impact, and recordkeeping. A healthcare employer should not assume that an internal analytics dashboard lies outside employment law because its data originated in an EHR. [S048]

The proposed safeguard is separation of safety reporting from performance punishment. A clinician who rejects inaccurate suggestions should not automatically look less productive than a clinician who accepts them uncritically. Nor should every deviation be exempt from review. The organization should evaluate reasons, case mix, error correction, and actual outcomes rather than reward raw acceptance rates.

A humane transition protects more than jobs. It protects the conditions in which skill is developed and exercised: time to examine, space to reason, the ability to question, and a credible route to challenge a system. Those conditions are part of the infrastructure of safe care, even when they do not appear as a software line item.

Part III — Follow the work, the money, and the power

19. Who benefits economically—and who may pay for the benefit

The economic beneficiaries of healthcare AI are not a single coordinated group. Providers, health plans, EHR suppliers, cloud providers, software developers, investors, clinicians, patients, and taxpayers can have different interests. A technology may create genuine social value while distributing that value unequally. It may also generate private savings largely by moving costs to someone else. Distinguishing value creation from cost shifting is central to evaluating the business of medicine.

A health system may gain additional completed visits, faster billing, better capture of legitimately documented complexity, lower vacancy costs, or improved retention. A payer may gain lower processing costs, better fraud detection, or reduced spending on unnecessary services. A vendor may gain subscription revenue, usage-based charges, integration fees, renewal leverage, and opportunities to develop related products. Patients may gain access and clearer records—or absorb more out-of-pocket costs, more appeals, and less human attention. These are possible economic channels, not findings that all are present at the author’s employer.

Commercial claims make the revenue dimension explicit. Ambience markets coding support, additional reimbursement capture, and documentation that can offset its cost. Fathom markets touchless coding. AKASA describes an autonomous coding and documentation-integrity platform. These representations show that healthcare AI is sold to financial and operational leadership as well as to clinicians seeking relief. They do not prove fraudulent coding or actual employee displacement. [S009, S013, S014]

The proper evaluation uses several ledgers. The provider ledger measures collections and costs. The payer ledger measures medical spending and administration. The clinician ledger measures hours, judgment, compensation, and liability exposure. The patient ledger measures access, outcomes, autonomy, privacy, and financial burden. The public ledger measures system costs, equity, workforce resilience, and trust. A gain in one ledger cannot simply be labeled a gain in all five.

A particularly important distinction is between revenue and margin. An additional encounter may generate a payment while also requiring nursing support, room capacity, supplies, interpretation, laboratory follow-up, and physician inbox work. Counting gross reimbursement as net savings exaggerates return. Conversely, an intervention that reduces burnout without adding billable visits may still have economic value if it reduces recruitment costs or preserves access. A serious analysis should not count only what is easy to invoice.

FIGURE 05 · Five economic ledgers

20. The payer-provider algorithmic loop

The ambient scribe can be the first stage of a chain that ends in a payment decision. A transcript becomes a note; a note becomes diagnoses and codes; a coded claim and selected records become inputs to utilization review, risk adjustment, payment edits, or retrospective review. A subsequent appeal may itself be drafted by another AI tool. The result can be software generating the clinical representation, software evaluating that representation, and software arguing about the evaluation—with patients and clinicians left to resolve exceptions.

This is an architectural risk model, not proof that every payer receives every recording or that all claims are decided by generative AI. Payers may receive structured claims, requested records, authorization submissions, or other permitted information rather than the entire EHR. Interfaces and contracts determine what travels. A clinician’s unaccepted on-screen suggestion is not automatically a submitted diagnosis code. The error has to cross additional workflow gates before that consequence follows.

Consider a hypothetical outpatient encounter in which the physician discusses a disease only to explain why it remains unconfirmed. If an AI draft incorrectly states the disease as established, the clinician fails to correct it, a coding process carries it into a claim, and a payer uses the code in a coverage rule, the original error may acquire financial consequences. Depending on the rule, it could cause inappropriate approval, increased payment, denial, or later recoupment. The direction is not necessarily denial; the central problem is propagation of unsupported certainty.

The outpatient coding rule is a useful checkpoint: uncertain diagnoses are not coded as established merely because they were discussed. The current ICD-10-CM guidelines distinguish outpatient practice from specified inpatient uncertain-diagnosis rules. That setting-specific distinction should survive every transformation. [S086]

The feedback loop can also distort incentives without a factual error. Providers can buy systems that strengthen claims while payers buy systems that challenge them. Each organization may report a favorable return while the combined administrative contest grows more elaborate. Whether automation reduces total friction or accelerates an arms race requires measurement of total handling, appeal volume, reversals, delays, and patient outcomes—not simply each side’s savings.

A correction should therefore be able to travel downstream. Correcting a note without addressing an associated problem-list entry, pending order, claim, authorization request, or patient message may leave the error alive elsewhere. The precise correction duties depend on the record and legal context; the governance recommendation is a traceable propagation map and a responsible person for each affected system.

The litigation parallel must remain precise. Estate of Lokken concerns alleged algorithmically influenced coverage denials and contractual claims; it is not itself an antitrust judgment establishing collusion between documentation vendors and insurers. The March 2026 discovery order underscores the relevance of implementation and business records, not proof of the hypothesized loop in this report. [S065]

FIGURE 06 · A documentation error can enter a payment-and-appeal loop

21. Why an FQHC cannot import every hospital or fee-for-service ROI claim

An FQHC’s financial structure makes generic productivity marketing especially unreliable. Medicaid law provides a prospective-payment framework and permits qualifying alternative payment methodologies. Actual payment depends on applicable federal and state requirements, payer arrangements, qualifying encounters, and local operations. More diagnoses, a higher E/M level, or multiple same-day activities do not automatically generate proportionate additional payment. [S118]

This does not eliminate an incentive to increase capacity. More completed qualifying visits can matter, and accurate documentation can support quality reporting, contract performance, appropriate risk adjustment, continuity, and compliance. But the organization should demonstrate the specific pathway by which a proposed feature improves its economics. A claim that coding support will “pay for itself” should identify which payer, which payment rule, which validated change, and which incremental costs are included.

Safety-net care also involves work that commercial efficiency metrics can undervalue. Interpretation, transportation problems, unstable housing, medication affordability, behavioral-health needs, and fragmented outside records can make a short appointment unusually complex. This is an explanation of relevant workload variables, not a claim that a particular patient population uniformly requires a defined number of extra minutes.

The public-purpose question is how any savings are used. They could support longer complex visits, additional nurses, interpreters, care coordination, clinician retention, reduced waiting, or organizational reserves. They could also be absorbed by licensing fees or higher throughput expectations. None of those outcomes should be inferred merely from nonprofit status. A nonprofit organization still has a budget, vendors still have contracts, and productivity decisions still affect workers and patients.

An appropriate FQHC evaluation would compare actual collections and total cost under the relevant contracts, alongside measures of access and clinical quality. It would not substitute hospital RVU examples for the center’s own payment model. Nor would it assume that a physician’s saved typing minute can be converted one-for-one into another minute of safe appointment capacity.

22. An illustrative break-even model: why the denominator matters

The following model is deliberately hypothetical. It is not a quotation of DAX pricing, an estimate of any health center’s contract, or a representation of FQHC reimbursement. Its purpose is to expose assumptions that a procurement proposal should make visible.

Assume a combined monthly software and allocated implementation cost of $800 per clinician. Assume each additional completed, payable encounter contributes $80 after incremental nonphysician operating costs but before accounting for any additional physician time. On these invented assumptions, ten additional encounters per month cover the $800. That looks attractive on a narrow financial spreadsheet. It says nothing yet about whether those visits can be completed safely or whether the clinician actually saves time.

Now assume 400 monthly encounters. If documentation generation saves two minutes per encounter, gross time saved is 800 minutes. If verification and correction consume one minute per encounter, the net is 400 minutes. If they consume three minutes, the net becomes negative 400 minutes. The same subscription can therefore produce apparent financial payback under a volume plan while imposing a net time cost on the clinician. These are arithmetic scenarios, not observed performance estimates.

Additional visits also produce follow-up work. If ten extra encounters each create an illustrative ten minutes of later result review, messages, and coordination, that is another 100 minutes. An analysis that counts appointment revenue but excludes this work is incomplete. The number is only an example; an actual organization should measure it rather than borrow it.

A useful equation is: net organizational benefit equals validated incremental contribution plus genuinely avoided costs and retention value, minus licensing, implementation, review, remediation, security, and downstream costs. A corresponding clinician equation must include paid and after-hours time, cognitive workload, and the distribution of responsibility. Neither equation is satisfied by a faster first draft alone.

The financial review should also test sensitivity. What happens when encounter volume is lower than expected, a model update increases editing time, a payer changes a rule, or a clinician caring for complex patients receives less benefit? A procurement case that survives only optimistic assumptions is not a reliable foundation for mandatory staffing changes.

FIGURE 07 · Illustrative time-savings sensitivity

23. De facto mandates: when an optional button is not a meaningful choice

A written mandate is not the only way to compel adoption. A clinician can technically retain manual charting while facing appointment templates, chart-closure deadlines, compensation tiers, and performance flags that make the alternative impracticable. The relevant policy question is not only whether a manual route exists, but whether it is reasonably usable without disproportionate penalty.

A hypothetical example makes the distinction clear. An organization calls AI optional but expands a schedule on the assumption that every clinician will use it. A physician who declines because of documented errors must then complete the original clinical work and the additional visits through a slower pathway. When the resulting delays are scored as noncompliance, the formal option has been economically neutralized. This is a testable operational hypothesis, not a finding that this sequence occurred at the author’s employer.

Twenty-minute appointments, RVU targets, or prompt chart completion are not automatically unlawful. Their permissibility depends on contracts, professional standards, patient complexity, reasonable accommodations where applicable, and how the requirements are implemented. Nor does every unverified draft make deployment illegal; drafts exist precisely because review is necessary. The stronger objection is to requirements that foreseeably make necessary review or correction impracticable, particularly after specific hazards have been documented.

California’s Medical Board identifies limits on lay control of clinical decisions, including certain decisions about how many patients a physician must see or hours worked where they affect care. The analysis must also account for statutory exceptions permitting some entities to employ physicians; an FQHC or nonprofit clinic should not be declared unlawful merely because it employs them. Corporate-practice analysis turns on entity authority and actual interference, not labels alone. [S042, S043]

AB 1979’s forthcoming requirement to preserve independent judgment provides a closely related policy question: can an institution reasonably claim to preserve that judgment while making its exercise operationally punitive? The statute does not expressly resolve every productivity dispute or grant an unconditional right to a preferred EHR. Its application should be evaluated once effective, using the actual facts rather than treating this question as already adjudicated. [S032]

The evidence needed is practical: scheduling templates before and after rollout; written and oral instructions; chart-delay notices; compensation formulas; permissions; training; override steps; protected review time; and what happened to clinicians who sought alternatives. Comparative data should account for case mix and duties. A single irritated supervisor’s remark is different from an organization-wide performance system.

The proposed rule is functional voluntariness: an alternative is meaningful only when it is technically available, reasonably usable, and not burdened by penalties that defeat the safety reason for using it. This is an advocacy standard, not a quoted existing statutory test. A defensible implementation could permit temporary documentation-only use, adjust workload while defects are investigated, and measure actual performance before tightening expectations.

After-hours charting also needs precise legal language. For a properly exempt physician, added charting time is not automatically a federal overtime violation. It can still matter to contract terms, workplace safety, retention, accommodation, or the feasibility of care. Nonexempt workers present different wage-and-hour questions. [S121]

FIGURE 08 · Operational pressure and the de facto mandate

24. Billing integrity, fraud risk, and the danger of optimizing the wrong objective

Accurate capture of work actually performed is legitimate. A system that helps an organization avoid undercoding is not, for that reason, fraudulent. The legal problem arises when documentation or claims represent unsupported diagnoses, services, time, complexity, or medical necessity, and the applicable knowledge and materiality requirements are met.

The federal False Claims Act addresses knowingly false claims and material false records, including certain reckless or deliberately ignorant conduct; ordinary mistakes do not automatically constitute fraud. United States ex rel. Schutte v. SuperValu emphasizes the relevance of what the defendant actually knew and believed rather than an after-the-fact objectively reasonable interpretation. This is a billing-law principle, not a holding about DAX or ambient scribes. [S091, S063]

For clinical AI, a warning pattern could become important. If audits show systematic unsupported diagnostic inflation and leadership continues a configuration to increase reimbursement, the facts would differ from a promptly corrected isolated error. The inquiry would need actual claims, knowledge, actions, financial consequences, and contractual roles. It would not be enough to show that marketing uses the phrase “revenue integrity.”

The EHR has already been the setting for a serious commercial-influence enforcement case. In the Practice Fusion resolution, the Department of Justice described a scheme involving an opioid manufacturer’s payments and clinical decision-support alerts designed to increase prescribing. The matter involved criminal and civil resolutions; it was not a generative-AI case and does not establish misconduct by current ambient vendors. Its relevance is structural: clinical software can become a channel through which financial incentives affect care. [S064]

Referral algorithms and preferred order matches deserve the same scrutiny. A default may reflect formulary requirements, availability, evidence, or convenience. It may also create a financial advantage for an affiliated service. The Anti-Kickback Statute and physician self-referral law require their own elements, exceptions, and safe-harbor analyses. A commercially favorable default is not automatically a kickback, and a referral suggestion is not by itself a prohibited financial relationship. [S092, S093]

The proposed governance safeguard is separation between clinical truth and financial optimization. A coding module may identify a query for clarification; it should not silently rewrite uncertain clinical reasoning into certainty to improve payment. Clinical leaders should see audits of both undercapture and unsupported capture. Compliance should not be defined solely as successful claim submission.

25. The economics of denials, appeals, and administrative displacement

Automation can reduce the cost of reviewing requests. That can improve timely approvals, identify missing information, and make routine decisions more consistent. It can also enable a large number of adverse decisions to be produced with limited individual attention. Which outcome occurs depends on design, incentives, rules, and human review—not the word “AI” alone.

CMS’s Medicare Advantage guidance states that algorithms may assist but cannot substitute for the required assessment of individual circumstances and applicable coverage criteria. Predictions about average recovery do not themselves establish that a particular beneficiary no longer needs covered care. [S049]

The WISeR model adds a concrete economic example. CMS is testing technology-supported review in selected Original Medicare settings and describes compensation tied to reducing inappropriate spending alongside quality safeguards and human clinical review of adverse recommendations. These features should be analyzed together. It would be misleading to describe the model simply as unrestricted payment for AI denials, just as it would be incomplete to ignore its financial incentives. [S050, S051]

A strong evaluation asks whether the organization counts wrongful denials, reversals, abandonment, delayed care, and the time required of patients and treating professionals. A low internal processing cost can conceal a high external appeal cost. If a request takes seconds to reject but hours to contest, the parties face asymmetrical burdens even without proof of intentional bad faith.

In Kisting-Leung v. Cigna, the October 2026 order reviewed for this report enlarged briefing limits relating to class certification. It did not adjudicate that the challenged process was unlawful or establish every allegation as fact. In Lokken, the March 2026 order was a discovery decision. These cases illustrate contested accountability questions, not a universal judicial conclusion that all automated utilization review is illegal. [S065, S066]

The proposed reporting standard should disclose authorization and denial rates with meaningful denominators, reasons, review time, specialty qualifications, appeal rates, reversal rates, and time to eventual care. An organization should not present a high automation rate as inherently successful without examining whether the correct services were approved or denied.

26. Market concentration, lock-in, and why the antitrust argument is separate

An integrated AI platform can become harder to replace as it connects to documentation, orders, coding, staffing, patient messages, and payer processes. Switching can require retraining, remapping interfaces, migrating data, and revalidating workflows. These are plausible switching costs. They are not, by themselves, proof of monopoly or unlawful conduct.

The strategic issue is dependence. An organization may find that removing a troublesome feature also disrupts unrelated functions, or that contract terms make selective withdrawal expensive. A clinician’s request for documentation-only use then depends on choices made upstream during procurement. The legal and economic analysis should inspect modularity, export rights, interoperability, term lengths, and whether essential evidence remains accessible after termination.

Antitrust requires its own proof. Sherman Act §1 addresses anticompetitive agreements; §2 addresses monopolization and related conduct. Common use of similar AI, a large market share, or aggressive efficiency marketing does not automatically satisfy the elements of either claim. A denial class action should not be relabeled an antitrust case merely because it involves a powerful insurer. [S119, S120]

Potentially relevant questions include whether competitors coordinate through a common system, whether access is conditioned on exclusionary terms, whether tying or foreclosure affects a defined market, and whether claimed efficiencies are real. Those questions require market evidence, conduct evidence, and legally cognizable harm. This report does not assert a proven healthcare-AI antitrust violation.

Procurement can reduce avoidable dependence without waiting for litigation. Organizations can negotiate function-level suspension, readable exports, evidence access, clear exit terms, limits on unilateral material changes, and defined responsibilities during outages. A hospital should not discover during an incident that neither its clinicians nor its lawyers can retrieve the version of the system that generated the disputed recommendation.

27. A task-by-task workforce impact statement

A workforce assessment should name jobs and tasks rather than speak vaguely about efficiency. Transcriptionists and scribes may face reduced demand for routine drafting. Coders may move toward complex exceptions and auditing. Reception and call-center staff may handle fewer routine questions but more distressed or unusual callers. Nurses may receive fewer standardized outreach calls and more escalations. Clinicians may spend less time composing notes but more time verifying machine-generated material. None of these changes guarantees a net reduction in total employment.

The report’s proposed workforce statement would identify expected staffing changes, attrition assumptions, outsourcing, retraining, new safety roles, workload distribution, and the treatment of workers whose tasks are automated. It would compare promises made during procurement with actual staffing and overtime after deployment. It would also identify where work has moved rather than vanished: to patients, family caregivers, offshore annotators, vendor staff, or unpaid clinician time.

Quality of work matters as much as headcount. If the straightforward cases are automated, remaining workers may face a more difficult case mix. Maintaining the old cases-per-hour target can then become unreasonable even if the raw number of tasks falls. An appropriate metric would account for complexity and escalation burden rather than assume that every remaining item is equivalent to the items that disappeared.

Patients can also become unpaid quality-control workers. They may have to correct a portal diagnosis, explain that a referral was duplicated, repeat a history to a human after a failed bot interaction, or appeal an algorithmically influenced denial. These are potential effects to measure. A system that reduces a call center’s workload by transferring the same work to vulnerable patients has not necessarily improved healthcare productivity.

The positive alternative is intentional redeployment. Time saved can support better access, more complete follow-up, protected training, reduced after-hours work, and human contact where it has the most value. Those outcomes require decisions about allocation. They do not follow automatically from installing software.

Part IV — Provenance, evidence, privacy, and the architecture of accountability

28. The provenance gap: a signature is not a complete history of the text

Authorship, review, approval, and causation are different concepts. A signed note may contain human-typed text, direct dictation, AI-generated sentences, imported history, templates, and later amendments. Knowing who signed it is important, but that event alone does not reconstruct how every statement came to exist.

It would be too broad to assert that all current EHRs record only signatures or lack granular provenance. Capabilities vary, and some systems retain versions, source links, editor events, or other metadata. The correct demand is to inspect the actual implementation and demonstrate what it can reconstruct. Absence of visible lineage on the clinician’s screen is not proof that no backend evidence exists.

HIPAA requires audit controls and protections against improper alteration or destruction. ONC certification criteria address audit logging and reporting. Neither should be casually represented as a universal requirement to preserve sentence-by-sentence AI authorship, every model response, and every intermediate prompt. Baseline logging and clinically useful provenance are related but not identical. [S098, S099]

HL7 FHIR provides a useful technical distinction. AuditEvent records events involving information; Provenance describes the entities and processes involved in producing or transforming it. The HL7 AI Transparency continuous-build guide demonstrates work toward AI-specific representation, but the page reviewed explicitly says it is not an authorized publication. It is a technical direction, not a binding law or proof of deployment in Epic. [S100–S102]

The proposed requirement is selective, clinically material provenance: identify which diagnoses, orders, assessments, and patient instructions were generated or materially modified by AI; link them to relevant sources and versions; and preserve the sequence through clinician review and downstream action. The goal is not an unreadable chart covered in technical labels. It is an accessible explanation backed by an auditable record.

A physician signature should remain meaningful. The proposal is not to let clinicians disown everything in a signed note. It is to avoid confusing an attestation event with proof that the clinician originated every sentence or that software, integration, and institutional decisions played no causal role. California’s electronic-signature statute protects legal effect against objections based solely on electronic form; it is not a complete clinical-authorship rule. [S115]

29. A forensic architecture that could answer the real questions

The minimum useful record should identify the encounter and source recording; the relevant transcript segment; the system and model version; the template and configuration; the output as first presented; edits, rejection, regeneration, and acceptance events; the user and service identities involved; and when a diagnosis or order became active, signed, or transmitted. These are proposed evidence fields, not assertions that all are presently required or available.

Three snapshots are particularly important: what the clinician was shown, what the clinician approved, and what the downstream system ultimately stored or acted upon. A mismatch can reveal a transformation after review. That possibility must be tested rather than alleged. The interface evidence available to the author does not establish that DAX inserted content after signature.

A useful design would bind the approval event to a specific content version. If material clinical content changes afterward, the system should record the change and invoke an appropriate re-review process rather than silently preserving the appearance of unchanged approval. Separate service-account identity can help distinguish human edits from automated transformations. A single shared account can make reconstruction much harder.

Tamper-evident logs, signed event records, version hashes, synchronized timestamps, access controls, and secure exports can improve evidentiary reliability. None proves that a statement was medically correct or that the signer understood every word. A hash proves correspondence to a particular byte sequence when properly implemented, not the truth of the clinical content. “Non-repudiation” should therefore not be marketed as a substitute for clinical accountability.

The system should preserve uncertainty and source type as structured attributes where feasible: patient report, examination finding, historical record, differential possibility, confirmed assessment, or proposed action. This can reduce the chance that a later module treats every sentence as equally authoritative. Provenance also needs to identify corrections and superseded versions so that an obsolete AI suggestion is not rediscovered as current truth.

Clinicians and patients need usable views; investigators need authorized access to deeper evidence. These audiences should not receive identical interfaces. A patient-facing explanation may say that a draft was AI-assisted and clinician-reviewed. An investigator may need the transformation history and configuration. Access must remain controlled to prevent the evidence system itself from becoming a privacy hazard.

FIGURE 09 · Three snapshots for a traceable clinical record

30. Preservation is not the same as keeping every recording forever

A defensible retention policy balances clinical need, privacy, contractual obligations, regulatory requirements, and reasonably anticipated litigation. Indefinite storage of every intimate conversation is not the necessary answer to every audit problem. Nor is rapid deletion always acceptable once a specific incident requires investigation.

An organization should distinguish user-visible availability from backend retention, research-environment processing, backup retention, and legal-hold capability. Commonly cited 30-day DAX retention windows should not be treated as a universal deadline for all Microsoft products. Current Dragon Copilot documentation describes up to 90 days for specified data, while legacy DAX documentation and customer arrangements may differ. The actual product, terms, configuration, and encounter dates must be confirmed. [S002, S104]

For the reported events, the practical request is targeted preservation through authorized personnel: relevant recordings or transcripts where retained, note versions, order states, audit events, configuration changes, support tickets, and applicable vendor communications. The request should identify the affected encounters securely rather than circulate patient details in public advocacy material.

Federal Rule of Civil Procedure 37(e) addresses loss of electronically stored information that should have been preserved in anticipation or conduct of litigation and the conditions for remedial measures or sanctions. It does not make every routine deletion spoliation. State proceedings and institutional investigations have their own applicable requirements. [S097]

The screen-capture restriction shown earlier is an information-security control. It is not, without more evidence, proof of concealment or destruction. The appropriate response is to request an authorized demonstration, secure capture, export, or preservation—not to bypass security controls or create a private archive of patient records.

A quality report is not automatically privileged merely because it is called a safety report. Peer-review and patient-safety protections depend on their legal requirements and the process used. Original clinical records and independently existing facts do not become immune from discovery simply by being copied into an incident system. Counsel should guide the preservation and disclosure plan without allowing privilege concerns to delay correction of patient-care errors.

31. Consent is layered: recording, care, processing, training, and disclosure

“Consent” is often used as if it were a single permission. In ambient healthcare AI, at least five questions should be separated: consent to record a confidential conversation; permission to use information for care and permitted operations; disclosure to service providers; secondary use for development or research; and later disclosure to other recipients. A general treatment form does not necessarily answer all five.

California Penal Code §632 generally addresses recording confidential communications without all parties’ consent, subject to its terms and exceptions. The statute does not impose one universal separate written opt-in form for every clinical AI interaction. Whether valid consent exists depends on the facts, notice, conduct, setting, and applicable law. The proposed distinct opt-in workflow is therefore a stronger governance recommendation, not a quotation of a categorical existing rule. [S116]

HIPAA permits various uses and disclosures for treatment, payment, and operations, subject to its requirements. California’s Confidentiality of Medical Information Act adds a separate layer, including authorization rules and specified exceptions. Neither should be collapsed into the recording-consent analysis. A use can be permissible under one framework yet require additional analysis under another. [S094, S117]

A meaningful explanation should tell patients that audio will be processed to produce draft documentation, that the clinician reviews the result, and how to request a pause or alternative. It should describe material retention and third-party processing without overwhelming patients with technical detail. Where information is used beyond the immediate service, the disclosure should distinguish the purpose and applicable choice rather than bury it in a generic promise of “improving care.”

The proposed consent model separates recording choice from unnecessary secondary use. It also provides an operational path for patients who decline. A refusal should not be converted into a clinical disadvantage merely because the institution has designed no alternative. The exact legal entitlement varies; this report advocates a clearer protection rather than claiming every jurisdiction already provides it.

Special contexts need additional review: children, substitute decision-makers, interpreters, multiple family members, behavioral-health and substance-use information, and cross-border telemedicine. A single consent checkbox should not be assumed to resolve each situation. Nor should a public-policy article publish identifiable examples simply because they demonstrate the technology’s risk.

32. Cloud processing and model improvement: what the contracts actually need to say

A business-associate agreement is a necessary contractual safeguard in many deployments, not a general license to use clinical conversations for any purpose. HHS guidance explains that business associates may use or disclose protected information within authorized functions and contractual limits; independent research is not automatically transformed into a permitted business-associate activity. Cloud providers and subcontractors require appropriate treatment under the applicable framework. [S054, S106]

Microsoft’s current Dragon Copilot privacy documentation provides an important concrete example. It describes collection of encounter audio and clinician edits, anonymization of selected data within a designated research environment, use of resulting anonymized data to improve Dragon Copilot models, and limited controlled human review. This is the vendor’s description, not an independent audit or proof that identical terms govern the author’s legacy DAX implementation. [S103]

Published Microsoft terms also contain processing permissions concerning model development and de-identification. The privacy white paper and contractual text should be read together with the executed agreement, service-specific terms, version dates, and any negotiated amendments. It would be inaccurate to say simply that all enterprise health data is never used for improvement; it would also be inaccurate to assume identifiable patient conversations are unrestricted public training data. [S105]

HHS distinguishes valid de-identification from merely removing names. The recognized approaches involve the applicable Safe Harbor or Expert Determination requirements. De-identification itself is a use of PHI that must be authorized; after valid de-identification, HIPAA’s PHI restrictions no longer apply, although contracts and other law can still matter. [S107, S108]

The policy questions remain even where processing is lawful. Who benefits from derived data? Can the provider negotiate limitations? What can the patient understand or decline? Are sensitive contextual details necessary for the service? Can data be linked back through other sources? What happens to derived models or datasets after a customer terminates? These questions call for precise answers rather than a vague claim that the system is “HIPAA certified.” HHS does not approve products as HIPAA-compliant in the way that marketing sometimes suggests. [S055]

A responsible procurement review should document the data map, subprocessors, locations, permitted purposes, retention periods, deletion exceptions, model-improvement rights, incident notification, evidence access, and restrictions on advertising or unrelated commercialization. The institution should be able to explain its choices to patients and clinicians in ordinary language.

33. Cybersecurity, availability, and the dependence problem

Clinical AI creates both confidentiality and operational questions. If a recording cannot upload, a service becomes unavailable, an identity system fails, or an integration produces inconsistent versions, the clinic still has patients to treat. The essential safeguard is a tested downtime pathway that preserves clinical authority and necessary documentation.

This report does not claim that cloud deployment is inherently insecure or that local processing is inherently safe. Each architecture has different attack surfaces and operational responsibilities. Risk assessment should address access, authentication, data integrity, service availability, vendor dependencies, and recovery. HIPAA’s technical safeguards provide part of the legal framework; they are not a complete clinical-safety specification. [S098]

Newly connected actions raise another concern. A text-generation error can remain a draft; an agent with access to ordering, messaging, or scheduling can affect additional systems. Least-privilege permissions, separate service identities, clear approval gates, and limits on cross-system actions are proposed design safeguards. The institution should know which actions a tool can perform, not only what the front-end demo shows.

An outage also tests whether an alternative is real. If every clinician has been trained only to accept and correct an AI workflow, the organization may lose resilience when it fails. Maintaining human capability and a usable manual route is therefore a continuity-of-care issue as well as an autonomy issue.

Security controls should not obstruct incident investigation. Authorized exports and targeted retention can preserve evidence without allowing unrestricted screen capture or uncontrolled copying. The objective is secure accountability, not a false choice between privacy and proof.

Part V — The law: device regulation, professional responsibility, and distributed liability

34. Six legal questions that must not be collapsed into one

A tool can influence care without necessarily being an FDA-regulated medical device. A regulated device can be lawfully used without holding a professional license. A non-device system can still cause negligent harm or violate privacy law. A physician can remain responsible while a vendor or institution also bears responsibility. These distinctions follow from the separate federal device, state professional, tort, privacy, and institutional frameworks discussed below. [S024–S029, S045–S046, S094]

The first question is what the software is intended to do. The second is whether that function falls within device regulation, a statutory exclusion, an exemption, or an enforcement policy. The third is whether the activity constitutes a professional function that state law reserves to a licensed person. The fourth is whether the deployment preserves lawful professional judgment and patient safety. The fifth is whether particular conduct caused legally compensable harm. The sixth is whether data collection and processing were lawful.

A slogan such as “the AI has no license” does not answer these questions. Software does not ordinarily obtain a physician’s license. The relevant issue is whether the people and organizations using it are lawfully providing services and whether independent performance of a reserved clinical function is permitted. Conversely, a disclaimer calling software administrative cannot settle the issue if its intended functionality is diagnostic or therapeutic.

The report’s argument is accordingly function-specific. The expanded DAX–Epic workflow warrants a documented classification and safety analysis because it interprets conversations into clinically consequential structured outputs. The present record does not establish an unlawful unapproved device, autonomous prescribing, or an unlicensed-practice offense.

This precision is important strategically. An allegation broader than the evidence allows a vendor to answer the wrong question: it can demonstrate that an order still requires a signature and treat the entire complaint as resolved. A narrower account forces examination of the actual defect, the difficulty of correction, and the basis for the recommendation.

35. The FDA boundary and the limits of a universal certification proposal

The FDCA device definition in 21 U.S.C. §321(h) and the software provisions in §360j(o) supply the starting point. The latter exclude specified functions, including certain administrative, record, and decision-support activities. For the particular non-device CDS exclusion, all applicable statutory conditions must be satisfied. [S024, S025]

FDA’s January 29, 2026 final guidance explains its current interpretation, including the nature of inputs, recommendations to professionals, and the clinician’s independent ability to review the basis. It also describes enforcement discretion for certain single-option recommendations meeting other conditions. A selectable recommendation is therefore not automatically a regulated device; nor does a signature requirement automatically establish non-device status. The guidance is nonbinding and should not be described as a statute. [S026]

A key distinction is between processing spoken clinical information and using characteristics of a voice as a physiological or diagnostic signal. Ordinary ambient speech capture does not automatically settle the medical-signal criterion. A tool intended to infer a disease from vocal characteristics would present a different intended-use question. The applicable classification must reflect the actual function rather than the shared presence of a microphone. [S024–S027]

The proposed institutional requirement is a function inventory with a written regulatory rationale. Each materially consequential feature should identify its intended use, authorization or exclusion relied upon, population and setting, limitations, approval controls, and changes since the earlier deployment. If the organization cannot obtain that information, it should not treat an old documentation-only assessment as sufficient for new order or diagnosis functions.

A universal demand for “FDA certification” is legally imprecise. Device pathways can include clearance, De Novo authorization, approval, exemptions, and other policies; non-device functions do not simply obtain the same certificate. State legislation can require lawful deployment, documented classification, appropriate validation, and professional safeguards, but it cannot command FDA to authorize every excluded function. [S027]

The stronger legislative objective is a minimum accountability floor across both device and non-device clinical AI: transparent intended use, safety validation proportionate to risk, meaningful override, accurate attribution, incident response, and access to evidence. That proposal should be reviewed for federal preemption and state authority rather than advertised as an effortless solution.

FIGURE 10 · A function-specific regulatory inquiry

36. EHR certification, discrimination law, and public-program obligations

Health-IT certification and FDA authorization answer different questions. The predictive decision-support provisions under 45 C.F.R. §170.315(b)(11) address specified source attributes and risk-management practices for interventions supplied as part of certified modules. They do not establish that every third-party model integrated into an EHR has been independently validated or cleared as a device. [S029]

Federal nondiscrimination requirements add another layer. The text of 45 C.F.R. §92.210 prohibits covered discrimination through patient-care decision-support tools and requires reasonable efforts to identify specified uses and mitigate risk. Coverage, implementation, and the effect of litigation or later regulatory changes must be checked in a concrete enforcement matter; the existence of a regulation should not be used to ignore litigation affecting related Section 1557 provisions. [S052]

Operational testing should examine language, accent, disability, demographic representation, specialty, and clinical setting where relevant. A single aggregate accuracy score can conceal poor performance for a subgroup. That is a reason to measure subgroup performance, not permission to infer protected characteristics from patients’ names or appearances or to build discriminatory shortcuts.

Public-program participation can impose additional quality, record, and risk-management obligations. For a deemed health center, HRSA’s Chapter 21 expressly addresses clinical complaints, near misses, and tracking of referrals and diagnostic tests. Those requirements make a formal safety response appropriate even when the immediate problem first appears in an IT ticket. [S053]

The policy lesson is that software classification is not the sole source of accountability. A function outside FDA device regulation may still be subject to contract, privacy, nondiscrimination, billing, professional, and organizational requirements. “Not a medical device” is not equivalent to “unregulated” or “risk-free.”

37. California: present law, forthcoming protections, and the vetoed proposal

As of October 28, 2026, California’s legal position is layered. Current professional and privacy duties already apply. Two recently enacted measures take effect in January 2027. A separate override and retaliation proposal was vetoed. These categories must not be blurred when addressing the Medical Board or an employer.

Authority	Status at the report cutoff	What it contributes—and does not establish
BPC §§2234 and 2266	In force	Professional-conduct and adequate/accurate-record requirements; not automatic discipline for a corrected AI draft.
BPC §2056	In force	Protection for medically appropriate physician advocacy within its terms; not immunity from legitimate review.
AB 3030 / HSC §1339.75	In force	Disclosures for specified AI-generated clinical communications, subject to human-review exception; not universal recording consent.
AB 489	In force in 2026	Restrictions on AI representations implying licensed professional status; not a prohibition on all decision support.
Civil Code §1714.46 / AB 316	In force in 2026	Limits the defense that AI autonomously caused harm; preserves other defenses.
SB 1120 / relevant utilization-review statutes	In force	Individualized review and restrictions on AI substitution in covered utilization management; not a direct ambient-scribe device rule.
AB 1979, Chapter 854 (2026)	Enacted; effective January 1, 2027	Reasonable steps to retain professional judgment and limits on independent AI performance of legally reserved functions.
SB 503, Chapter 857 (2026)	Enacted; effective January 1, 2027	Developer/deployer responsibilities concerning foreseeable biased impacts of covered CDS.
AB 2575	Vetoed September 30, 2026	Proposed override, retaliation, and superseding-cause protections; cannot be invoked as enacted law.

The table is a map, not an exhaustive statement of every element or exception. Statutory text and current implementation control. [S032–S040, S045–S047, S112]

AB 1979 is especially relevant to the professional-judgment boundary. It distinguishes clinical decision support from specified administrative activities and excludes documentation and communication that do not involve professional judgment. Its “reasonable steps” requirement invites attention to practical controls, but the statute does not automatically establish that every difficult interface is unlawful. [S032]

SB 503 focuses on biased impacts and the responsibilities of developers and deployers. That matters for linguistic and demographic performance, but the law should not be mischaracterized as a general guarantee of perfect accuracy or as a medical-malpractice compensation statute. [S033, S034]

The veto of AB 2575 leaves a more specific policy debate. The Governor objected to the anti-retaliation provision’s “solely” formulation and to requiring the Labor Commissioner to adjudicate clinical scope and standard-of-care issues without appropriate expertise. A revised proposal should solve those problems by using an effective causation framework and a competent clinical review mechanism. [S035, S036]

38. Malpractice: neither acceptance nor override should be judged by hindsight alone

In a negligence case, the questions ordinarily include duty, breach, causation, and harm. California’s professional standard focuses on reasonable skill, knowledge, and care under comparable circumstances. Neither an incorrect output nor an adverse outcome automatically supplies every element. [S087]

If a physician accepts an inaccurate suggestion, the investigation should ask what review was reasonable, how visible the error was, what information was available, and whether the physician corrected known problems. If the physician rejects a suggestion that later appears correct, the investigation should ask whether the rejection was reasonable at the time. The correct answer produced by one model on one occasion is not automatically the legal standard of care.

Workflow constraints belong in that inquiry, but they do not lower the patient’s entitlement to reasonable care. A clinician cannot simply say that the employer demanded speed and therefore any error was excusable. Equally, an institution should not pretend that imposed conditions are irrelevant to foreseeability and causal responsibility. The appropriate analysis considers both the professional decision and the system in which it was made.

Reasoned disagreement should be documented when clinically material, but demanding an extensive defense of every rejected AI suggestion would create its own burden and encourage defensive medicine. A proportionate system records the recommendation, relevant context, and a concise explanation where necessary. It should not require the physician to prove a negative each time the software proposes something inappropriate.

There is also a danger of overreliance on the availability of AI. If a model is not validated for the patient or setting, failure to follow it should not be treated as disregard of an established warning merely because it was displayed. Conversely, a well-supported, relevant alert may deserve attention. The duty is to exercise reasonable judgment, not to always agree or always disagree.

The proposed legal clarification is therefore modest but important: acceptance, rejection, and signature should be treated as evidence within a complete analysis, not as conclusive shortcuts that determine fault while excluding the model, interface, and institutional conditions.

39. California’s cost-containment cases: a useful analogy with limits

Wickline v. State of California, 192 Cal.App.3d 1630, addressed alleged harm associated with a payer’s cost-containment decision. The court reversed the judgment against the state on the facts. Its discussion of professional responsibility and potentially defective utilization review should not be converted into a holding that physicians always bear exclusive liability—or that payers are always liable. [S057]

Wilson v. Blue Cross of Southern California, 222 Cal.App.3d 660, reversed summary judgment because the evidence presented a triable issue as to whether the insurer’s conduct was a substantial factor in the patient’s death. The decision did not adjudicate an AI system, but it rejected an overbroad reading that a treating physician’s role necessarily removes a payer from the causation chain. [S058]

The analogy to mandatory AI is not that cost-containment software and ambient generation are identical. It is that external systems can constrain clinical choices while the physician retains duties. Causation should be evaluated rather than assumed away by pointing to the last licensed professional in the sequence.

This is also a warning to physicians. A report of a dangerous constraint is useful, but a clinician’s reasonable protective actions remain important. The most defensible record combines specific notice, attempts to obtain a workable alternative, careful correction, and appropriate escalation—not passive acceptance of known inaccuracies followed by a claim that the employer was responsible for everything.

For institutions, the analogous warning is that a “physician must review” disclaimer may be important but not dispositive. If implementation makes meaningful review foreseeably difficult and the organization ignores documented defects, the complete factual record may matter to liability. Whether it establishes a viable claim depends on the particular duties, harms, causation, and defenses.

40. Product liability, warnings, learned intermediaries, and federal preemption

A clinical-AI claim against a vendor may encounter threshold questions about whether the disputed offering is treated as a product, a service, or a combination; what duty is alleged; what warnings were given; and how the defect caused harm. It is premature to assume that every software error supports strict products liability in every jurisdiction.

Himes v. Somatics, a 2024 California Supreme Court decision involving a medical device, addresses failure-to-warn causation in the learned-intermediary setting. It does not establish a DAX-specific liability rule. Its broader relevance is that the clinician’s intermediary role and the patient’s informed choice require a careful causation analysis rather than an assumption that warning the doctor automatically ends every manufacturer responsibility. [S059]

Federal device regulation creates additional complexity. Riegel v. Medtronic, 552 U.S. 312, held that federal law preempted certain state claims involving a premarket-approved device. Buckman v. Plaintiffs’ Legal Committee, 531 U.S. 341, addressed implied preemption of fraud-on-the-FDA claims. Neither case means that every claim involving every FDA-cleared software product is barred. Both caution against assuming FDA authorization automatically expands civil remedies. [S060, S061]

A proposal to strengthen FDA oversight should therefore be accompanied by analysis of patient remedies and state authority. Safety evaluation, disclosure, postmarket monitoring, and compensation are separate policy tools. A device may be more closely reviewed while some particular tort theories become more difficult. A sound reform package anticipates that possibility rather than promising that a certificate will transfer all liability to the manufacturer.

Warnings also must be usable. A lengthy disclaimer in a contract may differ from a clear on-screen notice at the point of a clinically consequential action. But the existence of a bad outcome does not prove the warning was inadequate. Investigators need the actual labeling, training, display, and user behavior, not a generalized claim that physicians cannot read every manual.

41. Insurance disputes, ERISA, and algorithmic due process

A coverage dispute can have a different remedial structure from a malpractice action. In Aetna Health v. Davila, 542 U.S. 200, the Supreme Court addressed complete ERISA preemption where claims fell within the benefits-enforcement framework and no independent legal duty supported them. That principle can affect the forum and remedies for some employer-plan disputes. It does not mean all insurance-related AI claims are preempted. [S062]

The applicable plan type matters: fully insured coverage, self-funded employer plans, Medicare Advantage, Medicaid, and Original Medicare do not share identical rules. State restrictions on utilization review must be applied within their jurisdiction. A persuasive article should not promise state-law damages without checking the governing plan and federal preemption.

Algorithmic notice is another issue. In K.W. ex rel. D.W. v. Armstrong, 789 F.3d 962, the Ninth Circuit addressed procedural protections in Medicaid budget reductions. It is not an ambient-AI case. Its relevance is the need for sufficiently intelligible information to permit a meaningful challenge to an adverse decision, rather than leaving a beneficiary unable to understand how the result was reached. [S069]

For contemporary AI, a meaningful explanation should identify the actual coverage criterion and patient-specific basis. “The algorithm determined this” should not substitute for the governing rule. Conversely, a payer need not disclose every trade secret merely to provide an adequate reason. The policy task is to ensure reviewability, contestability, and access to the information necessary to correct errors.

This also reinforces the provenance argument. A patient appealing a denial should not be forced to challenge an inaccurate diagnosis without knowing that it originated from an uncorrected draft. Institutions should design correction and appeal processes that can trace the relevant fact without exposing unrelated sensitive information.

42. Medical Board, FTCA, and employment exposure are separate tracks

California’s Medical Board can examine inadequate or inaccurate records and conduct meeting the statutory grounds for discipline. An error caught before acceptance is materially different from an inaccurate signed record that remains uncorrected. Patient harm is relevant but not required for every possible recordkeeping violation. The analysis must use the actual statutory ground rather than label all inaccuracies gross negligence. [S045, S046]

FTCA deeming can provide substantial protection for qualifying health centers and eligible individuals performing covered services. The terms and scope under 42 U.S.C. §233 matter, and substitution of the United States is not automatic for every activity or dispute. That framework does not generally eliminate state licensure accountability, employment consequences, or every privacy and contractual issue. [S053, S096]

Employment protections likewise need a separate analysis. Labor Code §1102.5 may protect qualifying disclosures based on a reasonable belief of a legal violation. Lawson v. PPG Architectural Finishes identifies §1102.6 as the governing burden framework for covered whistleblower claims. Not every complaint about inconvenience is protected, and not every adverse action after a complaint proves retaliation. [S047, S067]

BPC §2056 specifically protects medically appropriate advocacy within its terms and preserves legitimate review and disciplinary authority. Health and Safety Code §1278.5 may be relevant to covered health facilities, but its facility definition must be satisfied; it should not be assumed to cover every freestanding clinic merely because it provides healthcare. [S112, S123]

The strategic objective is to make a truthful clinical record: what happened, how it affected care, what was requested, what safeguards were available, and how the organization responded. It is not to insert legal accusations into every chart or to treat safety reporting as immunity from performance standards.

Part VI — Other states and countries: different routes to the same accountability problem

43. State experimentation: Colorado, Utah, and Illinois

Colorado illustrates how quickly a legal summary can become stale. Its 2026 legislation repeals and reenacts the earlier AI framework with requirements for covered automated decision-making technology materially influencing consequential decisions. The official summary describes documentation, notice, human-review instructions, material-update information, and record retention, with requirements beginning January 1, 2027. A report that simply repeats the original 2024 law and its old effective dates would be unreliable. [S070]

Utah provides a contrasting experiment. The state’s Doctronic pilot concerns bounded prescription renewals, not unrestricted autonomous diagnosis or new prescribing. When the official status page (last updated in June 2026) was reviewed on October 29, 2026, it stated that the pilot remained in Phase 1 (it gives a one-year term, October 2025 to October 2026, renewable for a further year; no renewal had been announced on the state’s pages as of that review), requiring authorization by a licensed practitioner for all prescription requests, and that robust evidence of benefit is not yet available. It also describes excluded activities and phased evaluation. This is more precise than headlines claiming that an AI has generally been licensed to practice medicine. [S071]

Illinois’s Wellness and Oversight for Psychological Resources Act establishes restrictions concerning AI in therapy and psychotherapy. Its scope and exceptions must be read as enacted rather than summarized as a ban on all mental-health technology. It reflects a state choice to preserve professional responsibility in a particularly sensitive domain, not a national rule applicable to every documentation tool. [S072]

These approaches are not interchangeable. One emphasizes consequential decisions and developer/deployer obligations; another permits a constrained regulatory pilot; another limits particular professional uses. The comparative lesson is to define the function, the population, the decision authority, and the mechanism for review before claiming success or failure.

California can learn from such experiments without copying them wholesale. A safe harbor for a bounded refill task would not answer the problems of a complex oncology plan. A ban on autonomous psychotherapy would not resolve coding provenance. Legislation should be specific enough to address risk while flexible enough to accommodate different clinical contexts.

44. Europe: meaningful human involvement, product rules, and changed timelines

The European Union treats some AI systems as high-risk under its AI framework, but implementation dates have changed. The Council’s June 29, 2026 announcement describes delayed application dates of December 2, 2027 for stand-alone high-risk systems and August 2, 2028 for high-risk systems embedded in products. These dates should not be confused with every other obligation under the AI Act. [S073, S074]

European data-protection law also raises questions about consequential automated decisions. In SCHUFA Holding, C-634/21, the Court of Justice examined automated scoring whose result played a determining role in a third party’s decision. It was not a healthcare case. Its relevance is that a nominal human step does not necessarily settle the substantive role of automation. The analogy should not be represented as a U.S. liability rule. [S076]

The EU’s revised Product Liability Directive addresses software within its framework and contains evidentiary provisions relevant to complex products. National transposition and temporal application matter. It should not be presented as an already available universal remedy for every October 2026 healthcare-AI dispute. [S075]

The policy principle is useful: a person who nominally approves a result may not exercise meaningful control if the decision is practically determined upstream. Yet legal regimes differ in definitions, exceptions, enforcement, and remedies. A comparative article should identify those differences rather than suggest that Europe has one settled solution that California can enact verbatim.

45. Britain, Australia, and Canada: deployment evidence matters

NHS guidance treats ambient scribing as a clinical and information-governance deployment problem, with attention to safety assessment, organizational responsibility, explanation, review, and objections. The UK’s clinical safety practices offer a useful model for a documented deployment safety case rather than reliance on vendor demonstrations alone. They are guidance in a different legal system, not California statutes. [S077, S078]

Australia’s TGA reported on October 16, 2026 that eight of twelve reviewed scribe products contained functions likely to meet its medical-device definition, two required more information, and two appeared unlikely to qualify. The review examined regulatory characterization; it did not test safety, performance, privacy, or every product. The products reviewed reportedly included Microsoft Dragon Medical One; the TGA did not say which products fell into which category, and Dragon Medical One should not be casually treated as identical to the author’s DAX–Epic configuration. [S079]

That finding supports a narrow but important proposition: marketing a system as a scribe does not conclusively describe all its functions. It does not prove that DAX or Abridge violates U.S. law, and Australia’s conditional CDSS exemptions are not identical to the U.S. statutory non-device exclusion.

Health Canada’s machine-learning-enabled device guidance addresses lifecycle evidence and planned changes. Its relevance is the recognition that performance and risk can change after deployment. A one-time approval or favorable pilot should not excuse a failure to evaluate significant updates. Canadian device classification and privacy requirements remain jurisdiction-specific. [S080]

Across these systems, the recurring governance questions are recognizable: intended use, evidence, transparency, human control, changes, and ongoing monitoring. The strongest international comparison is not a contest to identify the strictest country. It is a search for enforceable mechanisms that preserve benefits while making responsibility traceable.

Part VII — A practical institutional and legislative program

46. What a defensible deployment should demonstrate before it becomes mandatory

The proposed minimum is a documented safety case for the exact integrated workflow. It should identify the clinical function, intended users, population, setting, model and EHR versions, known limitations, approval gates, failure recovery, and responsible owners. A vendor’s general accuracy statement cannot answer whether a local order-matching configuration is safe.

Validation should include realistic conversation rather than only clean dictated commands. Test cases should distinguish old referrals from new referrals, unilateral from bilateral findings, differential considerations from established diagnoses, negated from affirmed symptoms, family history from patient history, and conditional plans from current decisions. Specialty, language, and setting should be represented where the system will be used.

Error evaluation needs denominators. A count of defects without the number and type of encounters cannot establish a rate. A high overall agreement score can conceal rare but serious failures. The organization should assess both frequency and clinical consequence, identify what reached the signed record, and avoid equating reviewer disagreement with patient injury.

The review process should itself be tested. Can a clinician reject an inaccurate diagnosis without losing valid narrative? Can a patient decline recording without the encounter breaking? Can the user tell whether an order is merely suggested or transmitted? Does a material regeneration invalidate an earlier review? How quickly can a defect be corrected across connected systems?

A staged rollout should have predetermined stop or restriction criteria. These may include reproducible high-consequence errors, inability to preserve documentation during correction, unclear order activation, or lack of a workable fallback. The thresholds should be set by qualified governance personnel for the actual context; this report does not invent a universal acceptable error rate.

An incident should result in a documented response: acknowledge, investigate, mitigate, communicate, verify the fix, and monitor recurrence. Training may be part of the answer, but it should not be the default explanation before the defect is reproduced. Equally, a correctable training gap should not be mislabeled as proof that the entire product is inherently unsafe.

47. A procurement and contracting agenda that distributes accountability

The contracting process should make responsibilities explicit before an error occurs. The vendor should identify its intended functions and regulatory rationale; the integrator should identify matching and transformation rules; the institution should identify configuration ownership; and clinical leadership should approve the conditions of use. No contract should leave every participant assuming that someone else evaluated the same hazard.

The agreement should address material updates, notice, validation support, incident response, audit access, retention and legal holds, service levels, downtime, exit assistance, and the ability to suspend particular functions. Liability caps, indemnity, defense obligations, and exclusions should be reviewed against the realistic risks rather than accepted as routine boilerplate. These are negotiation recommendations, not assertions about the author’s employer’s contracts.

A physician’s review duty should not be used as a universal contractual disclaimer for negligent design or implementation. Conversely, a vendor cannot reasonably insure every unrelated clinical error simply because its software was used. The objective is to align responsibility with actual control, breach, and causation, while preserving patients’ legally available remedies.

Institutions should require usable evidence access. A promise to cooperate is weak if relevant model versions, source links, or logs are unavailable by the time an incident is reported. Evidence access should survive contract termination when legally required, with secure procedures and reasonable limits.

Financial incentives should be disclosed internally. Does the vendor’s fee depend on volume, coding uplift, recovered revenue, avoided spending, or another outcome? None is necessarily improper, but each can influence design and evaluation. A clinical committee should understand the metric being optimized before judging a system’s success.

The governance body should include frontline professionals, patient representation, privacy and security expertise, quality personnel, and financial leadership. That is not a demand that every procurement decision become a referendum. It is a way to prevent a narrow financial case from being mistaken for a complete clinical assessment.

48. A functional-autonomy and patient-protection legislative framework

The proposed legislation should define clinically consequential AI functions by what they do, not by product branding. It should cover material generation or modification of diagnoses, treatment recommendations, medication directions, orders, triage, and coverage determinations, with proportionate treatment of clerical tools. Definitions should avoid sweeping ordinary calculators and harmless formatting into the same category as high-risk decisions.

The core protection should be a practical ability to exercise lawful clinical judgment. Employers should provide a usable override and fallback for documented safety defects and appropriate patient objections. Adverse action should not be based on good-faith safety reporting or professionally justified disagreement merely because it reduces an adoption or acceptance metric. Legitimate peer review and accountability must remain available.

The law should require provenance for clinically material outputs and a reviewable action history. It should define appropriate access for clinicians, patients, regulators, and litigants while protecting unrelated information and trade secrets. The aim is enough evidence to determine responsibility, not universal public release of recordings or source code.

Developer and deployer duties should be complementary: intended-use and limitation information, appropriate validation, update disclosure, monitoring, incident escalation, and correction. A mandatory deployment should carry a higher obligation to demonstrate that review is feasible than an optional low-risk drafting feature.

A liability provision should reject a categorical defense that a clinician’s failure to catch an AI error necessarily severs every other actor’s responsibility. It should not create automatic vendor liability or immunize the clinician. The drafting should preserve ordinary causation, comparative-fault, and patient-protection principles and account for federal preemption.

Consent provisions should distinguish recording from secondary use, provide a workable explanation and alternative, and address sensitive contexts. Workforce provisions should require an impact statement for substantial deployments, including expected task changes, staffing, retraining, and measurable distribution of time savings. These are proposals requiring legislative analysis, not descriptions of current comprehensive law.

The AB 2575 veto offers a drafting lesson: do not create a protection that requires proving the forbidden conduct was the sole reason for an adverse action, and do not assign clinical adjudication to an agency without a workable expertise mechanism. A revised bill should specify coordination and standards rather than leaving enforcement aspirational. [S035]

49. Turning the Medical Board presentation into an actionable administrative request

Public comment is useful for creating a record and requesting an agenda item. It is not itself a rulemaking petition, an investigation, a binding advisory opinion, or a civil judgment. The Board’s meeting notice, as reviewed for this report, lists meetings on November 18–20, 2026, with Panel A and Panel B listed for November 18. The agenda, which must be posted at least ten days before the meeting, should determine the correct forum for general policy testimony. [S088]

A more concrete route is a petition under Government Code §11340.6. It should identify the substance of the requested regulation, the reasons, and the Board’s legal authority. Section 11340.7 establishes the acknowledgment and 30-day denial-or-hearing-scheduling framework, subject to its terms; it does not require final adoption of a rule within thirty days. BPC §2018 supplies rulemaking authority concerning the practice of medicine within delegated limits. [S109–S111, S122]

The request should have three tracks. First, ask for a public agenda item and guidance explaining professional expectations. Second, petition for appropriate recordkeeping and review standards within the Board’s authority. Third, request legislative recommendations for matters that exceed that authority, such as comprehensive vendor duties or employment remedies. The Board cannot make a policy statement bind civil courts on negligence, rewrite FDA law, or regulate every healthcare profession.

The proposed authorship clarification should avoid inventing an existing irrebuttable presumption. A better request is that Board evaluations not treat the signature alone as conclusive proof of the origin of every statement or of the absence of system causation. Investigators should obtain relevant versions, audit information, workflow evidence, and the physician’s actions. This would not excuse inadequate review; it would improve factual reconstruction.

Similarly, ask the Board to evaluate reasonable care in the actual workflow environment without announcing that imposed constraints automatically lower the standard. The purpose is contextual assessment and identification of all relevant conduct, not an institutional-pressure defense to avoid patient protection.

A requested confidential AI-safety intake channel could categorize interface defects, unsafe configurations, and reported retaliation, route matters to the appropriate agency, and publish de-identified trends. Existing Board complaint procedures already have confidentiality features, but practical anonymity cannot be guaranteed: facts may reveal the source, and some allegations must be investigated with the affected parties. The Board cannot promise immunity from all employment consequences or substitute for every labor remedy. [S113]

If a purported advisory statement would create a new generally applicable binding rule, the APA may require formal rulemaking rather than informal guidance. Government Code §11340.5 limits use of unadopted regulations. A legally durable request should respect that boundary instead of asking the Board to create a sweeping immunity through an advisory letter. [S114]

50. Proposed consolidated petition language

Subject: Petition for rulemaking and request for public consideration of AI-generated clinical documentation, professional control, and record provenance

I request that the Medical Board of California consider rulemaking within its statutory authority and issue appropriate nonbinding guidance addressing the use of artificial intelligence that generates or materially modifies clinical diagnoses, assessments, treatment plans, and proposed orders. This request is made under Government Code sections 11340.6 and 11340.7, with reference to Business and Professions Code section 2018 and the Board’s responsibilities concerning professional conduct and accurate medical records.

The requested rulemaking should establish proportionate expectations for identifying clinically material AI contributions, preserving relevant versions and action histories, reviewing content before it becomes actionable, and correcting errors without destroying accurate narrative. It should clarify that a signature is an important attestation but is not, standing alone, a complete account of the origin or transformation of every statement. Nothing requested would immunize inadequate professional review or diminish patient protection.

The Board should examine whether clinical judgment remains practically available where an institution technically permits an alternative but imposes scheduling, chart-closure, compensation, or performance conditions that make the alternative unusable. The requested inquiry should distinguish legitimate efficiency expectations from constraints that foreseeably prevent reasonable review and correction after specific hazards are reported.

The Board should direct an appropriately scoped stakeholder process involving clinicians, patients, EHR and AI developers, health centers, hospitals, privacy experts, and other relevant regulators. The process should examine source attribution, laterality, diagnostic uncertainty, referral intent, material post-review changes, and coupled deletion of narrative when a proposed diagnosis is rejected.

I further request a documented intake and referral process for reports of unresolved AI-related clinical hazards and associated retaliation concerns, with clear explanations of confidentiality limits and jurisdiction. Matters beyond the Board’s authority should inform legislative recommendations and referral to the appropriate agency, rather than be treated as outside the patient-safety concern altogether.

The factual basis includes my reported October 27–28, 2026 experiences and an interface warning that rejecting a proposed diagnosis will delete associated ambient assessment-and-plan content. The interface display does not establish autonomous signing or transmission of orders, and I do not ask the Board to assume that it does. I request a function-specific, evidence-based review of how the workflow preserves clinical meaning and professional control.

Please acknowledge receipt and identify the proposed disposition under the applicable petition procedures. I am prepared to provide supporting material through an approved confidential process that protects patient information. This submission seeks appropriate rulemaking and guidance; it does not ask the Board to adjudicate civil liability or declare a specific product unlawful without investigation.

Kanwar Gill, MD

Drafting note: before filing, confirm the current submission instructions, add the date and contact information, attach a concise factual declaration and proposed regulatory text if appropriate, and have counsel assess scope and any concurrent employment or patient-safety proceedings. This text is not represented as already submitted.

51. Proposed public statement for the November meeting

Good morning. My name is Dr. Kanwar Gill. I am a California family physician.

I ask the Board to address a change in the role of clinical AI: tools introduced as scribes are now generating diagnoses, treatment-plan content, and proposed orders inside the electronic record. The issue is not whether AI can be helpful. It is whether professional control remains real when the software’s interpretation is wrong.

In my recently expanded Epic–DAX workflow, I encountered a unilateral concern represented as bilateral, differential discussions presented as encounter diagnoses, and a proposed new referral when I was reviewing an existing referral. The interface also warns that rejecting a suggested diagnosis will delete its associated assessment-and-plan note. I am not claiming that these suggestions were autonomously signed or that patient injury has been established.

The concern extends beyond a button. A manual pathway is not a meaningful alternative if scheduling, chart deadlines, or performance penalties make it unusable. A physician’s signature is important, but it should not erase the need to examine how software generated, changed, or presented the clinical content.

I request a formal agenda item and consideration of a rulemaking petition addressing practical override, accurate provenance, appropriate audit preservation, and investigation of reported defects. I also request a clear channel for confidential safety reports and interagency referral, with truthful limits on confidentiality and authority.

The Board should evaluate reasonable clinical judgment in the actual workflow environment without granting immunity for careless review. Patients deserve accountability from every participant that controls a material risk: the clinician, institution, integrator, and developer.

Human communication should not be forced into a format designed primarily for software extraction. AI should support the encounter, not convert the physician into a compulsory validator of decisions made elsewhere.

52. Strategic footprint: the strongest case, the strongest defense, and the next evidence

The strongest case is concrete: the workflow can change clinical meaning; rejecting a proposal can remove valid narrative; pressure can make nominal alternatives unusable; and the signature alone does not reconstruct the causal chain. The interface evidence and contemporaneous reports justify investigation. They do not establish all elements of malpractice, unlawful practice, or device noncompliance.

The strongest defense is also concrete: the output is a draft, clinicians can review and reject it, the software may be performing extraction rather than independent diagnosis, training may resolve some errors, and the institution may have a lawful and usable alternative. That defense should be tested with an observed workflow demonstration, not dismissed because it is inconvenient to the argument.

The preemptive response is to request evidence: show the exact feature configuration; reproduce each reported error securely; identify when a suggestion becomes active; demonstrate rejection without narrative loss; explain whether post-review regeneration occurs; disclose the actual fallback; and measure total work rather than initial drafting speed. These requests are harder to dismiss than an assertion that every AI recommendation constitutes illegal practice.

For a reporting physician, the immediate priorities are to correct affected records, verify whether any erroneous orders were released, use approved safety reporting, request a workable interim pathway, and preserve relevant evidence through authorized channels. No public article should substitute for necessary patient-care follow-up. Nor should the physician export identifiable patient material to strengthen a public narrative.

For the institution, the appropriate response is not automatic capitulation or automatic defense of the rollout. It is a prompt, documented evaluation with interim controls, qualified clinical review, vendor coordination, and feedback to the reporter. If the issue is a training gap, demonstrate the safe method. If it is a configuration or design defect, correct it and verify the correction.

For legislators, the essential task is to close accountability gaps without freezing useful innovation. A law that simply says “a human must approve” may leave the same problem intact. A law that grants blanket immunity for overrides may sacrifice patients. The more durable framework preserves practical authority, traceable evidence, proportionate responsibility, and meaningful remedies.

Conclusion — The person at the center must not become the person left holding every risk

Healthcare AI can make good care easier. It can also turn the human encounter into a stream of inputs for diagnosis generation, coding, payment optimization, staffing analysis, and automated review. The same record can serve all of those purposes, but their objectives are not identical.

The central question is who owns the decision in practice. If a clinician can understand, correct, reject, and replace an output while preserving accurate records and adequate time, AI may genuinely support professional judgment. If the clinician is constrained by defaults, deletion coupling, opaque transformations, and productivity penalties, a final signature may conceal rather than demonstrate meaningful control.

The economic question is equally direct: who receives the saved minute and the additional dollar? Patients may receive access and attention. Clinicians may receive relief. Institutions may receive capacity. Vendors may receive revenue. Workers may receive new opportunities—or displacement and heavier exception work. Those outcomes should be measured and openly negotiated, not inferred from a marketing slogan.

The legal response should refuse two extremes. It should not make physicians powerless and then hold them solely responsible. It should not use software as an excuse for negligent practice. Accountability should follow actual conduct, control, information, and causation, with evidence sufficient to distinguish them.

A patient’s uncertainty is not a defect in the conversation. A physician’s independent judgment is not an obstacle to efficient software. A worker’s expertise is not merely a cost to remove. The purpose of healthcare is not to maximize the amount of human behavior that can be made machine-readable.

The technology must remain accountable to the patient, the clinical encounter, and the public—not the other way around.

Preservation Appendix A — The preceding article

Archival status. The preceding article is retained below for the record, including its proposed public statement; a closing strategy note addressed to the author is withheld from this public edition. Its substantive wording is retained; citations have been normalized to the expanded report’s source register, and presentation has been standardized. This appendix is a record of the earlier draft, not a second independent validation of its claims. The expanded report supplies the more detailed qualifications, updated source treatment, distinctions between current and future law, and administrative-remedy analysis. Read the main report where a formulation in the earlier draft is broader or less precise.

When the AI Scribe Starts Choosing the Diagnosis

Protecting patients, professional judgment, and accountability as artificial intelligence moves from documentation into clinical care

By Kanwar Gill, MD
October 28, 2026

A patient encounter is not a sequence of commands addressed to a computer.

It is a conversation between human beings. It includes uncertainty, reassurance, questions, disagreement, family history, grief, humor, and explanations of conditions that the clinician may ultimately reject. A physician may discuss a medication without prescribing it, mention a diagnosis while explaining why it is unlikely, or ask about an existing referral without intending to create another one.

The patient should not have to understand the limitations of a language model to receive care. The physician should not have to reshape every expression of empathy, every differential diagnosis, and every discussion of uncertainty into language optimized for software extraction.

Artificial intelligence can assist this relationship. It should not quietly redefine it.

The central question is no longer simply whether AI can produce a readable clinical note. Commercial systems now describe capabilities that generate diagnosis-linked documentation, extract proposed orders, and prepare clinical actions inside the electronic health record. The regulatory and professional question is what safeguards must accompany those functions—and who remains accountable when they fail. [S001]

1. What changed in my clinical workflow

During the introduction of expanded functionality in my Epic–DAX workflow on October 27–28, 2026, I encountered several discrepancies between what I intended clinically and what the software presented.

A conversation reviewing the status of an existing referral generated a proposed new referral. A unilateral clinical concern was represented as bilateral. Conditions discussed as possibilities appeared as encounter diagnoses rather than remaining clearly identified as differential considerations.

An especially concerning problem arose when I tried to reject an incorrect diagnosis. The interface explicitly warned:

“Associated ambient A&P note will be deleted.”

That warning matters. The assessment and plan are not disposable text. They may contain the physician’s reasoning, counseling, follow-up instructions, and explanation of why a particular diagnosis or treatment was—or was not—chosen.

A workflow that connects rejection of an incorrect diagnosis to removal of its associated narrative creates a practical obstacle to correction. The physician must first preserve the useful documentation, reject the incorrect clinical label, and then reconstruct the record.

These observations have limits. The interface displayed suggested diagnoses and order choices. It does not establish that the software independently signed or transmitted an order. Nor does the warning prove that the original recording or transcript was permanently destroyed.

Nevertheless, the interface itself substantiates a narrower concern: rejecting a structured clinical suggestion is tied to removal of associated assessment-and-plan content. That is a legitimate design and patient-safety issue even before harm occurs.

My understanding of the instructions at my workplace is that recording is mandatory and that a usable alternative documentation pathway has not been offered. That understanding should be checked against the written policy and actual configuration. The distinction matters: a feature that is optional and easily bypassed presents a different governance problem from one that every clinician is required to use.

The objective is not to condemn an entire product based on two days of experience. It is to investigate specific failures before they become embedded in routine care.

2. The industry is moving beyond transcription—but not every generated order is an autonomous medical decision

There is a continuum between clerical assistance and clinical decision-making.

At one end, software converts speech to text. It may then summarize that text, organize it under headings, or extract a decision the physician has explicitly made. Further along the continuum, it may infer diagnoses, recommend treatment, select order candidates, populate clinical fields, or initiate downstream actions.

These functions should not be treated as interchangeable.

For example, converting “I am ordering a complete blood count today” into a proposed laboratory order can be clerical assistance. Inferring that a laboratory test is needed from a general conversation is a different function. Generating a draft is different again from releasing an order.

The relevant question is not whether AI touched the record. It is what the system decided, what it merely recorded, and what happened before meaningful clinician review.

What the vendors actually describe

System	Publicly described functionality	Important qualification
Microsoft DAX/Dragon Copilot with Epic	Generates assessment-and-plan diagnoses from ambient recordings and compares them with Epic diagnoses. DAX can also identify potential medication, referral, procedure, imaging, laboratory, and follow-up orders.	Microsoft describes order suggestions as aids, states that new orders cannot be created in the DAX Orders tab itself, and explains that EHR integration and local configuration determine the subsequent workflow.
Epic AI Charting/Art	Captures clinical conversations, drafts notes, and extracts diagnoses and orders.	The presence of Epic’s interface does not identify which vendor generated every component or how an individual organization configured it.
Abridge	Its October 2026 “Now in Practice” release describes outpatient orders appearing in Epic’s Heard in Visit Recording workflow and being staged for review and signature.	Abridge also describes review controls and safeguards. Those claims should be evaluated, not ignored. They do not establish that Abridge powers my DAX implementation.
Oracle Health Clinical AI Agent	Describes creation of draft orders from conversations, including medication-related, laboratory, imaging, and follow-up workflows.	Oracle states that clinicians review and approve the proposed orders.

These are vendor-described capabilities, not findings that the products unlawfully diagnose or prescribe. They demonstrate why evaluation must extend beyond the accuracy of a typed note to the accuracy of clinical interpretation and EHR integration. [S001–S007]

The industry’s strongest response is therefore substantial: these systems are intended to assist, and clinicians retain final approval.

But that is the beginning of a safety analysis, not its conclusion.

3. Human review must be possible in practice—not merely present as a button

A physician may technically have authority to reject an AI suggestion while facing a workflow that makes rejection difficult, destructive, or time-consuming.

Consider three hypothetical statements:

“We discussed cancer as one possibility, but we have not established that diagnosis.”

“Let us see whether the cardiology referral from last month has been scheduled.”

“We could consider changing the medication later, but we are not changing it today.”

A safe system must preserve uncertainty, chronology, negation, and intent. It must also distinguish the patient’s words from the clinician’s conclusions and a family member’s history from the patient’s own condition.

A fluent paragraph can still get those distinctions wrong.

The practical test of oversight should be whether a clinician can identify the supporting information, understand the proposed clinical meaning, reject or revise it, and preserve accurate documentation without excessive effort. FDA’s decision-support framework similarly focuses on the professional’s ability to independently review the basis of a recommendation—not simply the existence of an acceptance control. [S026]

That principle suggests a basic design requirement:

Rejecting an incorrect diagnosis should not require sacrificing an otherwise accurate clinical narrative.

Likewise, disagreement with a suggested plan should not prevent completion of a clinically appropriate encounter.

These are proposed safeguards, not a claim that every existing interface is legally required to offer a particular button or layout.

4. A documentation error can become a care error

The danger is not confined to the original note. Consider the potential sequence: a provisional possibility becomes a structured diagnosis; that diagnosis is copied into future records; a subsequent clinician treats it as established; and later testing, counseling, or treatment follows.

That sequence is a risk scenario, not a claim that it occurred in my reported encounters. It explains why errors affecting clinical meaning deserve more attention than spelling or formatting defects.

The same distinction matters for coding. The current outpatient ICD-10-CM guidelines instruct that uncertain diagnoses—such as suspected or rule-out conditions—should not be coded as established. Instead, the encounter is coded to the highest supported degree of certainty. Different rules apply to specified inpatient settings. An AI system therefore cannot safely treat every spoken diagnostic possibility as equivalent across settings. [S086]

The record should also distinguish between an AI proposal, a clinician’s accepted assessment, a pending order, a signed order, and an action actually transmitted or performed. Without those distinctions, investigators may be unable to reconstruct where an error entered care.

5. The concern extends across specialties and professions

The following are illustrative failure modes that should be tested, not allegations that these events occurred in every specialty or that all listed functions are deployed everywhere.

Clinical setting	What AI-assisted documentation or decision support must preserve
Family medicine, internal medicine, and geriatrics	Multimorbidity, medication reconciliation, historical versus active conditions, differential diagnoses, and the difference between reviewing care and initiating it.
Emergency medicine and critical care	The distinction between a time-critical possibility, an exclusion, a confirmed finding, and an action already taken. A delayed correction may be particularly consequential.
Cardiology, pulmonology, and neurology	Whether a risk estimate, suspected syndrome, or diagnostic possibility is being discussed rather than established; the basis for a proposed test or intervention.
Oncology and hematology	Suspected versus confirmed malignancy, treatment intent, staging uncertainty, contraindications, and the distinction between discussing options and selecting a regimen.
Surgery, orthopedics, urology, and otolaryngology	Correct side, site, procedure, operative versus nonoperative options, and consent. A discussed procedure must not become a scheduled or ordered procedure without authorization.
Obstetrics, gynecology, and pediatrics	Which person is being discussed, gestational or developmental context, weight-dependent decisions, confidentiality, and family statements versus clinical findings.
Psychiatry, addiction medicine, and palliative care	Sensitive disclosures, hypothetical statements, risk assessment, patient preferences, decisional capacity, and discussions that are not diagnoses or treatment commitments.
Radiology and pathology	Detection versus interpretation, measurement versus diagnosis, uncertainty, comparison studies, and whether AI generated findings or merely organized a specialist’s findings.
Anesthesiology and procedural medicine	Actual medication administration versus proposed dosing, changing physiology, procedural context, and decisions requiring immediate verification.
Dentistry and dental specialties	Correct tooth, surface, quadrant, clinical findings, radiographic limitations, restorative options, periodontal assessment, and treatment necessity versus insurance coverage.
Podiatry	Laterality, wound location and severity, vascular and neurologic context, and the difference between preventive care, imaging, referral, and operative decisions.
Chiropractic, rehabilitation, physical and occupational therapy, and speech-language care	Scope-of-practice boundaries, contraindications, functional findings, treatment goals, referral triggers, and patient-specific rather than template-generated plans.
Nursing, advanced practice, and pharmacy	Correct ordering authority, medication intent, administration status, handoff accuracy, and whether the system proposes an action outside the user’s lawful role.

The common requirement is not that every profession use identical software. It is that the system preserve clinical meaning and respect the authority of the professional responsible for the decision.

6. Radiology: AI assistance is real; “physicians only rubber-stamp reports” is not an established general fact

Reports that computers now read everything and radiologists merely sign the results should not be repeated as an industry-wide fact without evidence.

Several distinct functions are often grouped together: prioritizing a worklist, detecting a suspected abnormality, measuring anatomy, generating a draft report, and independently making a narrowly defined diagnostic determination.

Those distinctions matter. FDA’s current guidance differentiates software that drafts a report from a radiologist’s findings from software that analyzes the medical image itself. The latter may fall within device oversight even where a related text-based function receives different treatment. [S026]

The American College of Radiology’s ARCH-AI program emphasizes institutional governance, evaluation, integration, and ongoing oversight. Its Assess-AI initiative addresses real-world monitoring of algorithm performance. Those programs support scrutiny of implementation—not an assumption that every deployment is safe or that radiologists have generally ceased interpreting studies. [S084]

Autonomous AI is not categorically illegitimate. FDA granted De Novo authorization to IDx-DR for a narrowly defined diabetic-retinopathy screening use, with specified patients and equipment. That is not authorization to diagnose every eye disease or to replace all ophthalmic judgment. [S085]

The appropriate question is therefore:

Is this specific autonomous or assistive function validated and authorized for this specific use, and are its limitations respected?

It is not enough to say either “a computer did it” or “a physician signed it.”

7. Dentistry and insurance: identifying an abnormality is not the same as deciding coverage

A dentist’s account of helping label radiographs for an insurer raises legitimate questions about how the resulting system will be used. It does not, standing alone, prove that the insurer intends unlawful automated denials.

Expert annotation can be a legitimate part of developing and validating diagnostic systems. The American Dental Association has established standards addressing image annotation and data collection for AI, alongside guidance concerning independent validation. FDA has also cleared specific dental image-analysis products, including Overjet Dental Assist. [S081, S082]

However, three questions must remain separate: what an image shows, what treatment the patient clinically needs, and what a particular benefit plan covers.

A radiographic measurement does not necessarily resolve the entire clinical question. Nor does FDA clearance for an image-analysis function amount to blanket approval of every insurance decision made using its output.

Fairness requires acknowledging vendor explanations. Overjet describes its Review Pass workflow as confirming that payer clinical requirements are met or directing a case toward manual review, and expressly says the system does not deny claims. That is the company’s stated model—not independent proof of every implementation, but also not evidence of automated denial. [S083]

The accountability inquiry should examine actual decisions: whether a qualified reviewer assessed the individual case, what information was available, how exceptions were handled, and whether the patient received a meaningful explanation and appeal.

California already regulates part of this territory. Under SB 1120’s amendments, covered utilization-review systems must not substitute AI for the required professional review of medical necessity. The statute expressly includes specialized health plans and requires attention to individual clinical information. Applicability must still be evaluated for the particular plan and regulatory framework. [S040]

Litigation shows why implementation records matter. In Estate of Lokken v. UnitedHealth Group, a March 2026 federal order addressed discovery concerning an allegedly AI-influenced coverage process, including implementation, oversight, incentives, and projected cost savings, while denying broader valuation and profit discovery. It was a discovery ruling—not a finding that every allegation was proven or a universal rule that AI-assisted utilization review is unlawful. [S065]

8. Productivity is a legitimate objective—but it must not be assumed to equal safe capacity

Reducing documentation burden can benefit patients and clinicians. A fair critique should recognize that possibility rather than assume every deployment is solely a revenue strategy.

The evidence, however, is not uniform.

A randomized trial involving 238 outpatient physicians across 14 specialties compared DAX, Nabla, and usual care. Nabla reduced the measured time-in-note outcome; DAX did not produce a statistically significant reduction in that outcome. Some well-being measures suggested benefits from either tool. The study evaluated a particular period and workflow, not every subsequent release or expanded order-generation feature. [S015]

A separate multisite observational study found AI-scribe adoption associated with approximately 13.4 fewer EHR minutes per eight scheduled patient hours and 0.49 additional visits per week. These were modest associations—not a guarantee that each clinician could safely add several patients every day. [S016]

An organization’s evaluation should therefore measure net work, including review, correction, order reconciliation, recovery of missing narrative, patient questions, and downstream follow-up—not merely how quickly the first draft appears.

A proposed efficiency measure might be:

Time saved generating documentation, minus time spent verifying and correcting it, minus additional work attributable to erroneous or unnecessary outputs.

That is an analytical framework, not a validated billing formula.

A business case based on additional visits should also account for the work those visits generate after the encounter. A shorter note-writing interval does not eliminate responsibility for laboratory results, referrals, medication questions, and care coordination.

The concern is not that organizations may seek efficiency. It is that unproven efficiency assumptions could become mandatory workload increases before safety and total workload are measured.

9. The physician faces two liability questions—not one

The first is familiar: what happens when a physician accepts an incorrect AI suggestion?

The second is equally important: what happens when a physician rejects an AI suggestion that later appears correct?

Both scenarios deserve attention. Neither establishes automatic liability.

California’s professional-negligence framework asks whether the clinician exercised the skill, knowledge, and care reasonably expected under comparable circumstances. A bad outcome alone does not prove negligence. A model’s correct answer, viewed retrospectively, does not by itself establish that the physician’s contemporaneous judgment was unreasonable. [S087]

Accepting an incorrect output

A claimant may argue that the physician failed to review the final record or order adequately. The analysis should also examine whether the system introduced a foreseeable error, whether rejection was workable, whether the employer knew of the defect, and whether an alternative was available.

Those are potentially relevant facts—not automatic defenses for the clinician or automatic liability for the vendor.

Overriding an output that later proves correct

A claimant may argue that the physician disregarded an available warning. The physician may have had reasonable grounds to disagree: unreliable input, an inappropriate model population, conflicting examination findings, patient preferences, or a recommendation outside the system’s intended use.

The appropriate inquiry is whether the decision was reasonable with the information available then—not whether hindsight can identify one correct output among many.

AI should neither become an automatic legal standard nor an automatic excuse.

At the same time, a rule providing absolute immunity for every override would be unsound. Deliberately ignoring a reliable, relevant warning can be unreasonable. The proposed protection should preserve professional judgment while retaining ordinary accountability for negligent conduct.

10. The physician’s signature should not erase everyone else’s role

The entity developing a model controls different risks from the company integrating it into an EHR. The employer selecting the configuration controls different risks from the clinician reviewing the final encounter.

A complete investigation should distinguish:

• Model behavior and validation.
• EHR matching, defaults, and release controls.
• Employer deployment, training, workload, and alternatives.
• Clinician review, correction, and clinical judgment.

These categories describe an investigative framework, not a predetermined allocation of fault.

California already rejects one oversimplified defense. Civil Code §1714.46, effective January 1, 2026, prevents a defendant that developed, modified, or used AI from defending a harm claim simply by asserting that the AI acted autonomously. Other defenses and comparative-fault arguments remain available. The statute neither guarantees recovery nor assigns all responsibility to any one participant. [S037]

Accordingly, the assertion that “all liability legally belongs to the physician” is too broad. So is the opposite assertion that software involvement excuses an inaccurate signed record.

Contracts may affect indemnification and defense arrangements, but no conclusion about a particular vendor’s contractual responsibility should be drawn without reviewing the actual agreements.

11. Medical Board accountability remains distinct from malpractice coverage

California Business and Professions Code §2266 addresses adequate and accurate medical records. Section 2234 includes disciplinary grounds such as gross negligence, repeated negligent acts, and incompetence. An isolated corrected software error does not automatically satisfy those standards, but AI involvement does not eliminate them. [S045, S046]

A sound regulatory review should examine both professional conduct and the surrounding system. Did the clinician identify the discrepancy? Was it corrected? Was the defect reported? Could the clinician reasonably bypass it? What happened after leadership received notice?

Reporting a hazard is important, but it is not prospective immunity. Once a clinician recognizes a problem, reasonable interim precautions remain necessary.

Health centers funded under the federal Health Center Program add another distinction. FTCA protection can apply to qualifying deemed health centers and eligible individuals performing covered activities within the relevant scope. It is not automatic for every activity or every person. HRSA expressly requires risk management addressing clinical complaints and near misses, including referral and diagnostic tracking. [S053]

That framework supports treating these reports as clinical safety concerns—not merely help-desk preferences. Civil malpractice protection does not itself answer professional-licensure questions.

12. Is the software a medical device?

The correct answer is function-specific.

Under federal law, some clinical decision-support functions are excluded from the medical-device definition. The statutory criteria address the information processed, the nature of the recommendations, and whether a healthcare professional can independently review their basis without relying primarily on the software. [S025]

The January 29, 2026 FDA guidance also recognizes enforcement discretion for certain functions that provide a single clinically appropriate option while otherwise satisfying the relevant criteria. Thus, the proposition that “any AI-generated diagnosis or treatment suggestion necessarily requires FDA clearance” is not an accurate statement of current policy. [S026]

Conversely, calling a product a “scribe” does not settle its classification. Nor does adding a physician signature step conclusively establish that every function is outside device oversight. FDA evaluates the actual intended function and applicable criteria. [S027]

The right questions for an integrated system are concrete: does it record an expressed decision, infer one, recommend a new one, or execute one? What supports the output? How can that support be reviewed? Is the function time-critical? Does it analyze a medical image or other regulated signal? What authorization, exclusion, or enforcement policy is being relied upon?

A recent international warning

Australia’s Therapeutic Goods Administration reported in October 2026 that, among 12 digital-scribe products reviewed, eight contained functions likely to meet its medical-device definition, two required further information, and two appeared unlikely to qualify. None of the reviewed suppliers had identified its product as a medical device.

That finding supports scrutiny of the gap between product labels and functions. It is not a U.S. classification decision, does not establish that DAX or Abridge violates U.S. law, and was not a comprehensive safety or performance evaluation. [S079]

The U.S. framework is also evolving. FDA’s August 2026 generative-AI discussion paper solicited input on regulatory approaches; it was not itself a final rule or an authorization for any particular product. [S124]

13. Why “California should require FDA certification for everything” needs refinement

The objective—independent scrutiny and manufacturer accountability—is sound. The proposed mechanism needs precision.

FDA does not issue one universal “medical AI certification.” Depending on the function, the relevant pathway may involve clearance, De Novo authorization, approval, an exemption, or non-device treatment. California cannot simply compel FDA to certify functions that federal law excludes from device regulation. [S027]

Federal preemption also matters. In Riegel v. Medtronic, the Supreme Court held that federal law preempted certain state-law claims involving a device that had received premarket approval. That holding does not eliminate every claim involving every device, particularly every software clearance. It does show why additional FDA oversight should not be presented as automatically guaranteeing greater civil recovery. [S060]

A stronger state proposal would require a documented regulatory classification for each clinically material function, compliance with any applicable federal authorization, and independent safety and usability controls even where a function is not federally regulated as a device.

EHR certification is another distinct layer. Federal health-IT requirements address transparency and risk management for certain predictive decision-support interventions supplied within certified modules. They are not a blanket FDA authorization or a guarantee that every third-party integration has been clinically validated. [S029]

Device status, professional scope of practice, institutional safety, and civil liability are related—but separate—questions.

14. California’s legal landscape has already changed

The argument should not suggest that California has done nothing. It should identify what is already protected and what remains unresolved.

Measure	Status as of October 28, 2026	Relevance
AB 3030	Effective	Requires disclosures and a human-contact pathway for specified AI-generated patient communications, with an exception for communications reviewed by a licensed or certified human provider. It is not a universal recording-consent statute.
AB 489	Effective January 1, 2026	Addresses AI-related representations implying that a system is a licensed health professional. It is not a blanket prohibition on clinical decision support.
AB 316 / Civil Code §1714.46	Effective January 1, 2026	Rejects the defense that AI autonomously caused the harm, while preserving other defenses.
SB 1120	Effective	Restricts AI use in covered utilization-review decisions and preserves required individualized professional review.
AB 1979	Enacted; effective January 1, 2027	Requires covered healthcare settings to take reasonable steps to preserve licensed professional judgment when care is informed by clinical decision support.
SB 503	Enacted; effective January 1, 2027	Establishes duties concerning foreseeable biased impacts, developer information, monitoring, and mitigation for covered clinical decision-support systems.
AB 2575	Vetoed September 30, 2026	Would have addressed clinician reliance and overrides, retaliation, and a specified liability-shifting defense. It is not current law.

These distinctions are supported by the Medical Board’s guidance, enacted statutory text, and the Governor’s veto message. [S032–S040]

AB 1979 directly addresses the professional-judgment boundary

The new law also restricts using AI to independently perform clinical functions legally reserved to licensed persons. It excludes documentation and communication that do not involve professional judgment. That distinction is highly relevant to the transition from scribing to clinical interpretation.

However, its January 2027 effective date must be respected. It is not a basis for declaring an October 2026 deployment already in violation. Nor does it expressly guarantee any physician’s preferred charting method. Its practical implementation will require attention to what preserving independent judgment actually demands. [S032]

The vetoed bill identifies an unresolved accountability issue

AB 2575 would have prevented certain defendants from claiming that a clinician’s failure to override a clinical decision-support output was a superseding cause severing their liability. It would also have created specified protections related to clinicians’ use of professional judgment. Other defenses and comparative-fault evidence would have remained available. [S036]

The Governor’s veto objected to the anti-retaliation mechanism’s drafting and enforceability, including its “solely” standard and the Labor Commissioner’s capacity to resolve clinical scope-of-practice and standard-of-care questions. Those objections should inform a revised proposal rather than be portrayed as a rejection of all clinician autonomy. [S035]

15. Mandatory recording creates a separate patient-consent issue

A policy requiring clinicians to use a recording system does not itself establish patient consent.

California Penal Code §632 generally requires all parties’ consent for recording confidential communications, subject to statutory qualifications and exceptions. HIPAA separately permits many treatment, payment, and healthcare-operations uses without a distinct patient authorization, but that does not automatically satisfy California recording law or authorize every secondary use of recordings. [S094, S116]

An organization should be able to explain what is recorded, who receives it, how long it is retained, whether it is used for model development, and what happens when a patient declines or asks to pause.

The policy should also address interpreters, accompanying family members, minors, and sensitive disclosures. These are not peripheral workflow details; they are part of preserving a trustworthy encounter.

My proposed standard is straightforward: patients should have a meaningful, understandable choice, and clinicians should have a safe way to continue care when recording is inappropriate or declined.

That is a policy recommendation—not a claim that existing law provides an identical opt-out rule in every circumstance.

16. What an enforceable accountability framework should require

The reform objective should be to preserve innovation while preventing clinical authority and legal responsibility from becoming disconnected.

Function-level transparency. Organizations should disclose which functions merely document, which infer or recommend, and which can initiate actions. Material updates should trigger renewed assessment rather than inherit approval from an earlier, less capable version.

A usable correction and bypass pathway. Clinicians should be able to reject diagnoses, orders, or plans without losing valid narrative. A safe documentation-only or manual alternative should be available when a material defect, patient preference, or clinical circumstance warrants it.

Validation of clinical meaning. Testing should include laterality, uncertainty, negation, historical versus current conditions, family-member attribution, existing versus new referrals, conditional medication decisions, and specialty-specific risks. Testing should also examine relevant languages and populations rather than assume uniform performance.

Traceable authorship and action. Systems should preserve who or what proposed an entry, who changed it, and when it became signed or actionable. Relevant model versions and configuration changes should be retrievable. Preservation must be secure and proportionate; this is not a proposal to keep every patient recording indefinitely.

Accountability proportional to control. A physician’s signature should not automatically extinguish claims concerning negligent design, integration, or deployment. Equally, vendor involvement should not excuse unreasonable clinical conduct. Patient remedies and professional standards should remain meaningful.

Protection for reasoned clinical disagreement. Good-faith safety reports and professionally justified departures from AI recommendations should not trigger retaliation merely because they interfere with adoption targets. Legislation should be drafted so clinical questions are assessed with appropriate expertise, addressing the concerns identified in the AB 2575 veto.

Measured—not assumed—productivity gains. Schedule increases should follow demonstrated net efficiency and safety in the relevant setting. Error correction, downstream workload, patient experience, and clinician burden should be part of the assessment.

Ongoing review after deployment. Near misses, recurring error mechanisms, performance changes, and significant updates should trigger investigation and proportionate corrective action. An effective program must be capable of limiting or withdrawing a feature, not merely providing additional training.

These are proposed requirements. Some overlap with existing or forthcoming law; others would need legislation, regulation, procurement standards, or organizational policy.

17. What the Medical Board should examine

The Medical Board’s role is not to become the FDA or regulate every dental, rehabilitation, or insurance function. Its contribution should be to clarify professional expectations and identify when deployment practices undermine lawful physician judgment.

A useful public examination would ask whether a physician can realistically reject AI-generated content; whether record authorship and order status are distinguishable; how known defects should be reported and mitigated; and how institutional constraints should be considered in a professional-conduct investigation.

The Board should also clarify that AI output is not automatically the standard of care, and that neither accepting nor rejecting it conclusively establishes negligence.

Coordination with other professional boards, health-plan regulators, and federal agencies is appropriate where their jurisdictions intersect. The goal should be consistent patient protection without pretending one agency controls the entire technology chain.

Most importantly, enforcement should not overlook the conditions under which a clinician is required to work.

If an institution mandates an AI-shaped workflow, meaningful physician control must exist before the signature—not merely responsibility after an error.

Conclusion: technology should accommodate care, not require care to perform for technology

The human encounter is not defective because it contains nuance that software struggles to interpret.

An uncertain diagnosis is not unfinished data waiting to be converted into certainty. A conversation about treatment options is not an instruction to choose one. A discussion of an old referral is not permission to create a new order.

Artificial intelligence may become a valuable participant in the work surrounding care. But when it shapes diagnoses, treatment plans, and clinical actions, it must be evaluated as more than a convenient writing tool.

The appropriate demand is neither a blanket prohibition nor unquestioning adoption. It is a system in which patients understand the technology’s role, clinicians retain practical decision-making authority, developers and employers remain accountable for what they control, and professional review does not become a mechanism for transferring every system failure to the last person who clicks “sign.”

The patient must remain the center of the encounter. The technology must remain accountable to that purpose.

Source register

Each bracketed citation in the text links here. A source supports the specific statement it is cited for; an official page that could not be retrieved in the 29 October verification pass is noted in the publication record.

1. S001 — Microsoft: Manage orders in DAX Copilot for Epic ↩ back to text

   Type: Vendor technical documentation

   Use and limitation: Suggested orders and local integration; not proof of autonomous signing.

   https://support.microsoft.com/en-us/dax-copilot/epic/manage-orders-in-dax-copilot-for-epic

2. S002 — Microsoft: DAX Copilot for Epic FAQs ↩ back to text

   Type: Vendor technical documentation

   Use and limitation: Summarization, recordings, editing, workflow and retention; local terms must be checked.

   https://support.microsoft.com/en-us/dax-copilot/epic/dax-copilot-for-epic-faqs

3. S003 — Microsoft: Intelligent diagnoses retrieval ↩ back to text

   Type: Vendor technical documentation

   Use and limitation: AI diagnoses compared with EHR diagnoses; configuration-specific.

   https://learn.microsoft.com/en-us/industry/healthcare/dragon-copilot/specialty-models/intelligent-diagnoses-retrieval

4. S004 — Microsoft: Dragon Copilot 3.6 release information ↩ back to text

   Type: Vendor technical documentation

   Use and limitation: Diagnosis-aware notes and Epic configuration; does not establish local deployment version.

   https://learn.microsoft.com/en-us/industry/healthcare/dragon-copilot/whats-new/3-6

5. S005 — Epic: Art for Clinicians ↩ back to text

   Type: Vendor product description

   Use and limitation: Notes, diagnoses, orders, messages and clinical summaries; availability may vary.

   https://www.epic.com/software/art/

6. S006 — Abridge: Now in Practice, October 2026 ↩ back to text

   Type: Vendor release, October 2026

   Use and limitation: Order staging, review and revenue-integrity features; not evidence that Abridge powers the reported DAX workflow.

   https://www.abridge.com/blog/now-in-practice-october-2026

7. S007 — Oracle: Order creation capabilities for Clinical AI Agent ↩ back to text

   Type: Vendor announcement, February 2, 2026

   Use and limitation: Draft clinical orders subject to review; savings are vendor-reported.

   https://www.oracle.com/news/announcement/oracle-health-adds-order-creation-capabilities-to-clinical-ai-agent-2026-02-02/

8. S008 — Oracle: Health AI Agents ↩ back to text

   Type: Vendor product/roadmap page

   Use and limitation: Separates available capabilities from features expressly described as planned.

   https://www.oracle.com/health/clinical-suite/ai-agents/

9. S009 — Ambience: Business ↩ back to text

   Type: Vendor commercial description

   Use and limitation: Revenue capture, coding and ROI claims; not independent financial validation.

   https://www.ambiencehealthcare.com/business

10. S010 — Ambience: Platform roadmap ↩ back to text

    Type: Vendor roadmap

    Use and limitation: Revenue integrity and workflow strategy; roadmap is not proof of deployment.

    https://www.ambiencehealthcare.com/blog/ambience-healthcare-unveils-platform-roadmap-to-rebuild-healthcare-with-ai

11. S011 — Hippocratic AI: New products ↩ back to text

    Type: Vendor commercial description

    Use and limitation: Non-diagnostic voice-agent pricing and scope; not equivalence to an entire licensed nursing role.

    https://hippocraticai.com/new-products/

12. S012 — Hippocratic AI: Polaris ↩ back to text

    Type: Vendor safety architecture

    Use and limitation: Vendor-reported benchmarks, escalation and boundaries; independent local validation still needed.

    https://hippocraticai.com/polaris/

13. S013 — Fathom: Primary care coding automation ↩ back to text

    Type: Vendor commercial description

    Use and limitation: Automation, cost and coding claims are vendor claims, not workforce-outcome findings.

    https://www.fathomhealth.com/primary-care-coding-automation

14. S014 — AKASA: Autonomous mid-cycle ↩ back to text

    Type: Vendor announcement, October 2026

    Use and limitation: Autonomous coding/CDI ambitions and claimed coverage; not proof that all coding staff have been replaced.

    https://akasa.com/press/autonomous-mid-cycle

15. S015 — Randomized trial of ambient AI scribes in outpatient practice ↩ back to text

    Type: Peer-reviewed randomized research

    Use and limitation: 238 physicians; product-specific findings; study period does not test every later software update.

    https://doi.org/10.1056/AIoa2501000

16. S016 — JAMA: Ambient AI scribe adoption, efficiency and productivity ↩ back to text

    Type: Peer-reviewed observational research

    Use and limitation: 13.4 fewer EHR minutes per eight scheduled hours and 0.49 additional weekly visits; association, not a guaranteed capacity gain.

    https://jamanetwork.com/journals/jama/fullarticle/2847319

17. S017 — Clinical note quality pilot, PubMed record 41996389

    Type: Peer-reviewed pilot research

    Use and limitation: Selected audited notes and potential-risk ratings must not be represented as population injury rates.

    https://pubmed.ncbi.nlm.nih.gov/41996389/

18. S018 — Multiperspective study of ambient scribes in general practice

    Type: Peer-reviewed research

    Use and limitation: Time and experience findings in a defined setting; generalization requires caution.

    https://www.nature.com/articles/s41746-026-02454-3

19. S019 — Specialty assessment of ambient AI documentation

    Type: Peer-reviewed research

    Use and limitation: Simulation and specialty performance are not identical to clinical outcomes.

    https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2843515

20. S020 — Wong et al.: External validation of a proprietary sepsis prediction model

    Type: Peer-reviewed research, 2021

    Use and limitation: Historical Epic sepsis model at a particular site; not an evaluation of DAX or every current Epic algorithm.

    https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307

21. S021 — BLS: Medical transcriptionists, Occupational Outlook Handbook ↩ back to text

    Type: Official labor projection

    Use and limitation: Current 2025–2035 projection: decline of 4%; projection is not an observed AI layoff count.

    https://www.bls.gov/ooh/healthcare/medical-transcriptionists.htm

22. S022 — ILO: Generative AI and jobs, 2025 update ↩ back to text

    Type: International labor research

    Use and limitation: Task exposure is not an estimate that every exposed job disappears.

    https://www.ilo.org/publications/generative-ai-and-jobs-2025-update

23. S023 — BLS: Industry and occupational employment projections overview ↩ back to text

    Type: Official labor analysis

    Use and limitation: Projection vintages differ; technology and demand can move employment in opposite directions.

    https://www.bls.gov/opub/mlr/2026/article/industry-and-occupational-employment-projections-overview.htm

24. S024 — FDCA medical-device definition: 21 U.S.C. §321(h) ↩ back to text

    Type: Federal statute

    Use and limitation: Device definition depends on intended use and statutory terms.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title21-section321&num=0&edition=prelim

25. S025 — Software exclusions: 21 U.S.C. §360j(o) ↩ back to text

    Type: Federal statute

    Use and limitation: Non-device software and CDS criteria; not a universal clinical-AI exemption.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title21-section360j&num=0&edition=prelim

26. S026 — FDA: Clinical Decision Support Software, final guidance January 29, 2026 ↩ back to text

    Type: Final nonbinding FDA guidance

    Use and limitation: Function-specific criteria and independent review; guidance is not itself a statute.

    https://www.fda.gov/media/109618/download

27. S027 — FDA: CDS frequently asked questions ↩ back to text

    Type: FDA explanatory material

    Use and limitation: Failing one CDS exclusion does not by itself resolve every other exclusion or enforcement policy.

    https://www.fda.gov/medical-devices/software-medical-device-samd/clinical-decision-support-software-frequently-asked-questions-faqs

28. S028 — FDA: Digital-health guidance inventory

    Type: FDA guidance/status inventory

    Use and limitation: Distinguishes final guidance, draft guidance and discussion documents.

    https://www.fda.gov/medical-devices/digital-health-center-excellence/guidances-digital-health-content

29. S029 — ASTP/ONC: Decision support interventions certification criterion ↩ back to text

    Type: Federal health-IT certification requirements

    Use and limitation: 45 C.F.R. §170.315(b)(11); predictive DSI supplied within certified modules; not blanket FDA clearance.

    https://healthit.gov/test-method/decision-support-interventions/

30. S030 — FDA: AI to support regulatory decisions for drugs and biologics ↩ back to text

    Type: Draft guidance, January 2025

    Use and limitation: Risk-based context-of-use credibility framework; draft, not a final binding rule.

    https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

31. S031 — FDA/EMA: Good AI practice in drug development ↩ back to text

    Type: Joint principles, January 2026

    Use and limitation: Human-centric, risk-based lifecycle principles; not approval of a drug or model.

    https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development

32. S032 — California AB 1979, chaptered text (Chapter 854, Statutes of 2026) ↩ back to text

    Type: Enacted statute, effective January 1, 2027

    Use and limitation: Professional judgment and reserved clinical functions; not yet effective at the report cutoff.

    https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB1979

33. S033 — California SB 503, chaptered text (Chapter 857, Statutes of 2026; approved by the Governor September 30, 2026) ↩ back to text

    Type: Enacted statute (Business and Professions Code §22758), effective January 1, 2027

    Use and limitation: Clinical decision-support bias obligations; effective January 1, 2027; the chaptered text sets no separate phased operative dates.

    https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB503

34. S034 — Governor: September 30, 2026 AI legislation signing announcement ↩ back to text

    Type: Official enactment announcement

    Use and limitation: Verifies signed bills; statutory text controls the details.

    https://www.gov.ca.gov/2026/09/30/californias-nation-leading-ai-framework-just-got-stronger-governor-newsom-signs-more-first-in-the-nation-worker-protections-and-more/

35. S035 — California AB 2575 veto message (AB 2575 was vetoed September 30, 2026) ↩ back to text

    Type: Official veto, September 30, 2026

    Use and limitation: Not law; objections concern drafting and clinical expertise in enforcement.

    https://www.gov.ca.gov/wp-content/uploads/2026/09/VETO-msg-AB-2575.pdf

36. S036 — California AB 2575 legislative text ↩ back to text

    Type: Vetoed proposal

    Use and limitation: Override, retaliation and superseding-cause proposals; not enforceable as enacted law.

    https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB2575

37. S037 — California Civil Code §1714.46 (AB 316, Chapter 672, Statutes of 2025) ↩ back to text

    Type: State statute, effective January 1, 2026

    Use and limitation: Autonomous-AI defense limitation; retains other defenses and does not create automatic strict liability.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV§ionNum=1714.46

38. S038 — Medical Board: 2026 legislative update

    Type: Official board legislative explanation

    Use and limitation: Includes AB 489 professional-title restrictions and relevant healthcare AI measures.

    https://www.mbc.ca.gov/News/Newsletter/2026-Vol173.aspx?embed=true

39. S039 — Medical Board: Generative-AI patient communications notification

    Type: Official board explanation of AB 3030

    Use and limitation: Specified patient communications and human-review exception; not universal recording consent.

    https://www.mbc.ca.gov/Resources/Medical-Resources/GenAI-Notification.aspx

40. S040 — California HSC §1367.01 ↩ back to text

    Type: State statute

    Use and limitation: Covered utilization-management decisions, including AI constraints; plan jurisdiction matters.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=HSC§ionNum=1367.01

41. S041 — California Attorney General: AI legal advisories

    Type: Official enforcement guidance, January 13, 2025

    Use and limitation: Existing law remains relevant to AI; advisory is not a finding against a named employer.

    https://oag.ca.gov/news/press-releases/attorney-general-bonta-issues-legal-advisories-application-california-law-ai

42. S042 — Medical Board: Corporate practice and practice information ↩ back to text

    Type: Official board guidance

    Use and limitation: Clinical control and operational decisions affecting practice; statutory exceptions require separate analysis.

    https://www.mbc.ca.gov/Licensing/Physicians-and-Surgeons/Practice-Information/

43. S043 — California BPC §2401 ↩ back to text

    Type: State statute

    Use and limitation: Specific employment exceptions; do not assume every FQHC falls within any one provision.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=BPC§ionNum=2401

44. S044 — California SB 351 text

    Type: State legislative text; enacted 2025

    Use and limitation: Private-equity/hedge-fund interference in physician and dental practices; not evidence of employer ownership.

    https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB351

45. S045 — California BPC §2234 ↩ back to text

    Type: State statute

    Use and limitation: Disciplinary grounds; a corrected isolated suggestion is not automatically misconduct.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=BPC§ionNum=2234

46. S046 — California BPC §2266 ↩ back to text

    Type: State statute

    Use and limitation: Adequate and accurate medical records; raw audio status requires separate retention analysis.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=BPC§ionNum=2266

47. S047 — California Labor Code §1102.5 ↩ back to text

    Type: State statute

    Use and limitation: Reasonable-belief protected disclosures; not every software disagreement is protected whistleblowing.

    https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=LAB§ionNum=1102.5

48. S048 — California CRD: Employment AI regulations approval ↩ back to text

    Type: Official regulation announcement

    Use and limitation: FEHA automated employment decisions; effective October 1, 2025.

    https://calcivilrights.ca.gov/2025/06/30/civil-rights-council-secures-approval-for-regulations-to-protect-against-employment-discrimination-related-to-artificial-intelligence/

49. S049 — CMS: Medicare Advantage coverage and utilization-management FAQ ↩ back to text

    Type: Official CMS interpretation, February 6, 2024

    Use and limitation: Individual circumstances required; algorithmic predictions do not replace coverage standards.

    https://www.cms.gov/files/document/hpms-memo-faq-coverage-criteria-and-utilization-management-cms-4201-f-02-6-2024-pdf.pdf

50. S050 — CMS: WISeR model ↩ back to text

    Type: Official model description, 2026–2031

    Use and limitation: Original Medicare in six states; compensation and human-review safeguards both matter.

    https://www.cms.gov/priorities/innovation/innovation-models/wiser

51. S051 — CMS: WISeR frequently asked questions ↩ back to text

    Type: Official model FAQ

    Use and limitation: Clinical review, appeal and payment protections; not a model of unrestricted AI denial.

    https://www.cms.gov/priorities/innovation/files/document/wiser-model-frequently-asked-questions

52. S052 — 45 C.F.R. §92.210: Patient-care decision support discrimination ↩ back to text

    Type: Federal regulation

    Use and limitation: Covered entities and reasonable identification/mitigation duties; litigation and coverage must be checked in an actual matter.

    https://www.ecfr.gov/current/title-45/section-92.210

53. S053 — HRSA Health Center Compliance Manual, Chapter 21 ↩ back to text

    Type: Official FTCA risk-management guidance

    Use and limitation: Deeming and scope are conditional; near-miss, referral and diagnostic follow-up controls.

    https://bphc.hrsa.gov/compliance/compliance-manual/chapter21

54. S054 — HHS: HIPAA cloud-computing guidance ↩ back to text

    Type: Official privacy/security guidance

    Use and limitation: Business-associate agreements, permitted uses and risk analysis; not a general data license.

    https://www.hhs.gov/hipaa/for-professionals/special-topics/health-information-technology/cloud-computing/index.html

55. S055 — HHS: Misleading HIPAA marketing claims ↩ back to text

    Type: Official privacy guidance

    Use and limitation: HHS does not approve or certify healthcare products as HIPAA-compliant.

    https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/be-aware-misleading-marketing-claims/index.html

56. S056 — FTC: BetterHelp final-order proceeding

    Type: FTC enforcement settlement

    Use and limitation: Health-data advertising restrictions; not a clinical-AI causation judgment.

    https://www.ftc.gov/legal-library/browse/cases-proceedings/betterhelp-inc-matter-timeline-item-2023-07-14

57. S057 — Wickline v. State of California, 192 Cal.App.3d 1630 (1986) ↩ back to text

    Type: Published appellate decision

    Use and limitation: State prevailed on facts; cost-containment and professional judgment discussion is an analogy, not an AI holding.

    https://law.justia.com/cases/california/court-of-appeal/3d/192/1630.html

58. S058 — Wilson v. Blue Cross, 222 Cal.App.3d 660 (1990) ↩ back to text

    Type: Published appellate decision

    Use and limitation: Reversed summary judgments where causation was triable; not a final damages judgment.

    https://law.justia.com/cases/california/court-of-appeal/3d/222/660.html

59. S059 — Himes v. Somatics, S273887 (Cal. June 20, 2024) ↩ back to text

    Type: California Supreme Court decision

    Use and limitation: Failure-to-warn causation and learned intermediary; not an AI product ruling.

    https://law.justia.com/cases/california/supreme-court/2024/s273887.html

60. S060 — Riegel v. Medtronic, 552 U.S. 312 (2008) ↩ back to text

    Type: U.S. Supreme Court decision

    Use and limitation: Certain state claims concerning a premarket-approved device preempted; not every clearance or claim.

    https://www.supremecourt.gov/opinions/07pdf/06-179.pdf

61. S061 — Buckman Co. v. Plaintiffs’ Legal Committee, 531 U.S. 341 (2001) ↩ back to text

    Type: U.S. Supreme Court decision

    Use and limitation: Fraud-on-FDA theory preemption; does not eliminate every traditional state tort claim.

    https://www.law.cornell.edu/supct/html/98-1768.ZO.html

62. S062 — Aetna Health Inc. v. Davila, 542 U.S. 200 (2004) ↩ back to text

    Type: U.S. Supreme Court decision

    Use and limitation: ERISA complete preemption in qualifying benefits disputes; not all payers or all independent duties.

    https://www.supremecourt.gov/opinions/03pdf/02-1845.pdf

63. S063 — United States ex rel. Schutte v. SuperValu, 598 U.S. 739 (2023) ↩ back to text

    Type: U.S. Supreme Court decision

    Use and limitation: False Claims Act scienter focuses on relevant subjective knowledge; not all errors are fraud.

    https://www.supremecourt.gov/search.aspx?filename=/docket/docketfiles/html/public/21-1326.html

64. S064 — DOJ: Practice Fusion criminal and civil resolution ↩ back to text

    Type: Federal enforcement resolution, January 2020

    Use and limitation: Historical EHR kickback/CDS case; not a generative-AI case or accusation against current vendors.

    https://www.justice.gov/archives/opa/pr/electronic-health-records-vendor-pay-145-million-resolve-criminal-and-civil-investigations-0

65. S065 — Estate of Lokken v. UnitedHealth, discovery order, March 9, 2026 ↩ back to text

    Type: Federal district court procedural order

    Use and limitation: Allegations and discovery, not a final finding of liability.

    https://law.justia.com/cases/federal/district-courts/minnesota/mndce/0%3A2023cv03514/211721/162/

66. S066 — Kisting-Leung v. Cigna, October 2026 class-certification briefing order ↩ back to text

    Type: Federal district court procedural order

    Use and limitation: Class-certification briefing scheduling/limits; not class certification or a merits judgment.

    https://docs.justia.com/cases/federal/district-courts/california/caedce/2%3A2023cv01477/431351/106

67. S067 — Lawson v. PPG Architectural Finishes, S266001 (Cal. 2022) ↩ back to text

    Type: California Supreme Court decision

    Use and limitation: Labor Code §1102.6 burden framework; application depends on protected activity and facts.

    https://law.justia.com/cases/california/supreme-court/2022/s266001.html

68. S068 — Gruber v. Yelp, A155063 (Cal. Ct. App. 2020) ↩ back to text

    Type: Published appellate decision

    Use and limitation: California recording-consent analysis; not an ambient-healthcare-specific ruling.

    https://law.justia.com/cases/california/court-of-appeal/2020/a155063m.html

69. S069 — K.W. ex rel. D.W. v. Armstrong, 789 F.3d 962 (9th Cir. 2015) ↩ back to text

    Type: Published appellate decision

    Use and limitation: Public-benefit notices and meaningful review; not a modern generative-AI medical judgment.

    https://openjurist.org/789/f3d/962/k-w-ex-rel-d-w-v-armstrong

70. S070 — Colorado SB26-189 ↩ back to text

    Type: Enacted state legislation

    Use and limitation: Replaces earlier AI framework; principal requirements effective January 1, 2027.

    https://www.leg.colorado.gov/bills/SB26-189

71. S071 — Utah: Doctronic authorized pilot ↩ back to text

    Type: Official regulatory pilot description

    Use and limitation: Narrow refill phases; live phase and renewal beyond initial term not independently established here.

    https://commerce.utah.gov/ai/regulatory-relief/authorized-ai-pilots/doctronic/

72. S072 — Illinois HB1806 / Public Act 104-0054 ↩ back to text

    Type: State legislative record

    Use and limitation: Therapy AI limits; final act, not introduced synopsis, controls.

    https://www.ilga.gov/Legislation/BillStatus?DocNum=1806&DocTypeID=HB&GA=104&GAID=18&SessionID=114

73. S073 — European Commission: AI Act implementation ↩ back to text

    Type: Official current implementation guidance

    Use and limitation: Updated high-risk dates: December 2, 2027 Annex III; August 2, 2028 regulated-product systems.

    https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

74. S074 — Council of the EU: Digital Omnibus on AI, June 29, 2026 ↩ back to text

    Type: Official adoption announcement

    Use and limitation: 2026 amendments change original implementation timetable.

    https://www.consilium.europa.eu/en/press/press-releases/2026/06/29/artificial-intelligence-council-gives-final-green-light-to-simplify-and-streamline-rules/

75. S075 — EU Product Liability Directive 2024/2853 ↩ back to text

    Type: EU directive

    Use and limitation: Software coverage and national transposition; not already uniform national liability rules at cutoff.

    https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024L2853

76. S076 — SCHUFA Holding, C-634/21 (CJEU, December 7, 2023) ↩ back to text

    Type: CJEU judgment

    Use and limitation: Credit-scoring/GDPR analogy; actual determining role can matter more than formal downstream approval.

    https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A62021CJ0634

77. S077 — NHS England: Ambient scribing guidance ↩ back to text

    Type: Official guidance, updated July 2026

    Use and limitation: Clinical safety case, deployment governance, assurance and monitoring.

    https://www.england.nhs.uk/long-read/guidance-on-the-use-of-ai-enabled-ambient-scribing-products-in-health-and-care-settings/

78. S078 — NHS: Information governance for ambient scribing ↩ back to text

    Type: Official information-governance guidance

    Use and limitation: Explain use, respect objections and review outputs; jurisdiction-specific.

    https://digital.nhs.uk/data-and-information/information-governance/guidance/using-ai-enabled-ambient-scribing-products-in-health-and-care-settings

79. S079 — Australian TGA: Digital scribe review outcomes ↩ back to text

    Type: Official review, October 2026

    Use and limitation: 8 of 12 likely devices; 2 uncertain; 2 unlikely; not a safety/performance study or a U.S. classification.

    https://www.tga.gov.au/resources/publication/corporate-reports/digital-scribe-review-outcomes

80. S080 — Health Canada: Machine-learning-enabled medical devices guidance ↩ back to text

    Type: Official guidance, April 2026

    Use and limitation: Lifecycle validation and planned changes; Canadian framework differs from U.S. law.

    https://www.canada.ca/en/health-canada/services/drugs-health-products/medical-devices/application-information/guidance-documents/pre-market-guidance-machine-learning-enabled-medical-devices.html

81. S081 — ADA: Artificial intelligence in dentistry ↩ back to text

    Type: Professional standards resource

    Use and limitation: Image annotation and independent validation; standards are not blanket clearance for insurance denials.

    https://www.ada.org/resources/practice/dental-standards/artificial-intelligence-in-dentistry

82. S082 — FDA: Overjet Dental Assist K210187 ↩ back to text

    Type: FDA device database

    Use and limitation: 510(k) substantial equivalence May 19, 2021; device-specific, not blanket authorization of payer decisions.

    https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K210187

83. S083 — Overjet: Insurance verification / Review Pass explanation ↩ back to text

    Type: Vendor description

    Use and limitation: Company says workflow does not deny claims; distinguish from allegations about other deployments.

    https://www.overjet.com/resources/webinars/lunch-learn-ai-insurance-verification

84. S084 — ACR: ARCH-AI ↩ back to text

    Type: Professional governance program

    Use and limitation: Institutional AI governance; not proof that radiologists only rubber-stamp reports.

    https://www.acr.org/Data-Science-and-Informatics/AI-in-Your-Practice/ARCH-AI

85. S085 — FDA: IDx-DR De Novo authorization DEN180001 ↩ back to text

    Type: FDA authorization database

    Use and limitation: Narrow autonomous retinal-screening function; not general authority to practice medicine.

    https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfPMN/denovo.cfm?ID=DEN180001

86. S086 — CDC: FY2027 ICD-10-CM official guidelines ↩ back to text

    Type: Official coding guidelines, effective October 1, 2026

    Use and limitation: Outpatient uncertain-diagnosis coding differs from specified inpatient settings.

    https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Publications/ICD10CM/2027/ICD-10-CM-October-1-2026-FY27-Guidelines.pdf

87. S087 — California CACI 501: Standard of Care for Health Care Professionals ↩ back to text

    Type: Judicial Council jury-instruction text reproduced by Justia

    Use and limitation: Reasonable professional care in comparable circumstances; not outcome-only liability.

    https://www.justia.com/trials-litigation/docs/caci/500/501/

88. S088 — Medical Board meeting: November 18–20, 2026 ↩ back to text

    Type: Official meeting notice

    Use and limitation: Correct forum and public-comment slot must be checked against the final agenda.

    https://www.mbc.ca.gov/about/meetings/details.aspx?id=4098

89. S089 — KPSGILL: Ambient clinical documentation and device boundary

    Type: Author’s prior public commentary

    Use and limitation: Background and continuity, not independent proof of allegations or legal authority.

    https://kpsgill.com/policy/ambient-clinical-documentation-benefits-and-risks/

90. S090 — KPSGILL: AI-assisted diagnosis and physician responsibility

    Type: Author’s prior public commentary

    Use and limitation: Background and continuity, not a substitute for primary authorities.

    https://kpsgill.com/policy/ai-assisted-diagnosis-and-physician-responsibility/

91. S091 — False Claims Act, 31 U.S.C. §3729 ↩ back to text

    Type: Federal statute

    Use and limitation: Knowing false claims and records, materiality and scienter; not every coding error is fraud.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title31-section3729&num=0&edition=prelim

92. S092 — Anti-Kickback Statute, 42 U.S.C. §1320a-7b(b) ↩ back to text

    Type: Federal statute

    Use and limitation: Remuneration and intent requirements; exceptions and safe harbors require separate analysis.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title42-section1320a-7b&num=0&edition=prelim

93. S093 — Physician self-referral law, 42 U.S.C. §1395nn ↩ back to text

    Type: Federal statute

    Use and limitation: Designated health services and financial relationships; not a universal ban on referral software.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title42-section1395nn&num=0&edition=prelim

94. S094 — HIPAA general rules, 45 C.F.R. §164.502 ↩ back to text

    Type: Federal regulation

    Use and limitation: Covered entities and business associates; treatment/payment/operations exceptions and contractual limits.

    https://www.ecfr.gov/current/title-45/section-164.502

95. S095 — HIPAA permitted disclosures, 45 C.F.R. §164.512

    Type: Federal regulation

    Use and limitation: Specified oversight and legal disclosures; public posting of PHI is not automatically authorized.

    https://www.ecfr.gov/current/title-45/section-164.512

96. S096 — Public Health Service / FTCA, 42 U.S.C. §233 ↩ back to text

    Type: Federal statute

    Use and limitation: Conditional federal substitution/protection; does not generally immunize state licensure conduct.

    https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title42-section233&num=0&edition=prelim

97. S097 — Federal Rule of Civil Procedure 37(e) ↩ back to text

    Type: Federal procedural rule

    Use and limitation: Electronically stored information preservation and sanctions; not every routine deletion is spoliation.

    https://www.uscourts.gov/rules-policies/current-rules-practice-procedure/federal-rules-civil-procedure

98. S098 — HIPAA technical safeguards, 45 CFR §164.312 ↩ back to text

    Type: Federal regulation

    Use and limitation: Audit and integrity controls; does not expressly require token-level AI provenance.

    https://www.ecfr.gov/current/title-45/section-164.312

99. S099 — ONC certification criteria, 45 CFR §170.315 ↩ back to text

    Type: Federal regulation

    Use and limitation: Audit logs and reports; certification is not proof of granular text lineage.

    https://www.ecfr.gov/current/title-45/section-170.315

100. S100 — HL7 FHIR R5 Provenance ↩ back to text

     Type: Published technical standard

     Use and limitation: Provenance describes production and transformation; implementation must be verified.

     https://hl7.org/fhir/provenance.html

101. S101 — HL7 FHIR R5 AuditEvent

     Type: Published technical standard

     Use and limitation: Event logging is distinct from complete authorship lineage.

     https://hl7.org/fhir/auditevent.html

102. S102 — HL7 AI Transparency on FHIR: AI Provenance ↩ back to text

     Type: Continuous-build implementation guide

     Use and limitation: Explicitly not an authorized publication; a technical direction, not binding law.

     https://build.fhir.org/ig/HL7/aitransparency-ig/StructureDefinition-AI-Provenance.html

103. S103 — Microsoft Dragon Copilot privacy white paper ↩ back to text

     Type: Vendor privacy documentation

     Use and limitation: Anonymized model improvement and limited human review; do not assume identical terms for legacy DAX or every customer.

     https://learn.microsoft.com/en-us/industry/healthcare/dragon-copilot/whitepapers/privacy

104. S104 — Microsoft Dragon Copilot data-retention FAQs ↩ back to text

     Type: Vendor technical documentation

     Use and limitation: Up to 90 days for specified data; distinguish user-visible DAX availability and actual contract.

     https://learn.microsoft.com/en-us/industry/healthcare/dragon-copilot/about/faqs

105. S105 — Microsoft Product Terms: Dragon Copilot data processing ↩ back to text

     Type: Vendor contract text

     Use and limitation: Broad processing permissions; executed customer terms and effective version control.

     https://www.microsoft.com/licensing/terms/product/PrivacyandSecurityTerms/all

106. S106 — HHS: Business Associates FAQs ↩ back to text

     Type: Official privacy guidance

     Use and limitation: BAAs constrain uses; they are not universal authorization for independent research.

     https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/index.html

107. S107 — HHS: De-identification guidance ↩ back to text

     Type: Official privacy guidance

     Use and limitation: Safe Harbor or Expert Determination; removing names alone is not sufficient.

     https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html

108. S108 — HHS: Business-associate de-identification and own use ↩ back to text

     Type: Official privacy guidance

     Use and limitation: De-identification itself must be authorized; other laws may still restrict subsequent use.

     https://www.hhs.gov/hipaa/for-professionals/faq/may-a-health-information-organization-de-identify-information/index.html

109. S109 — California Government Code §11340.6 ↩ back to text

     Type: State administrative statute

     Use and limitation: Petition must specify action, reasons, and agency authority.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=GOV§ionNum=11340.6

110. S110 — California Government Code §11340.7

     Type: State administrative statute

     Use and limitation: Receipt and 30-day response/hearing scheduling process; not final adoption within 30 days.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=GOV§ionNum=11340.7

111. S111 — California BPC §2018 ↩ back to text

     Type: State administrative statute

     Use and limitation: Medical Board rulemaking within delegated authority.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=BPC§ionNum=2018

112. S112 — California BPC §2056 ↩ back to text

     Type: State physician-advocacy statute

     Use and limitation: Protected medically appropriate advocacy; reasonable review and licensure accountability remain.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=BPC§ionNum=2056

113. S113 — Medical Board: Complaint-review FAQ ↩ back to text

     Type: Official Board procedure

     Use and limitation: Confidentiality is not guaranteed practical anonymity or immunity.

     https://www.mbc.ca.gov/FAQs/?cat=Consumer&topic=Complaint%3A+Review+Process

114. S114 — California Government Code §11340.5 ↩ back to text

     Type: State administrative statute

     Use and limitation: Generally applicable regulations need APA process unless exempt.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=GOV§ionNum=11340.5

115. S115 — California Civil Code §1633.7 ↩ back to text

     Type: State electronic-transactions statute

     Use and limitation: Electronic form alone does not defeat legal effect; not a sentence-by-sentence authorship rule.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV§ionNum=1633.7

116. S116 — California Penal Code §632 ↩ back to text

     Type: State recording statute

     Use and limitation: Confidential-communication consent; fact-specific exceptions and consent form.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN§ionNum=632

117. S117 — California Civil Code §56.10 ↩ back to text

     Type: State medical-confidentiality statute

     Use and limitation: Authorization and specific exceptions; distinct from recording consent.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV§ionNum=56.10

118. S118 — Medicaid FQHC payment, 42 USC §1396a(bb) ↩ back to text

     Type: Federal payment statute

     Use and limitation: Prospective and alternative payment arrangements; more codes do not automatically mean more payment.

     https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title42-section1396a&num=0&edition=prelim

119. S119 — Sherman Act §1, 15 USC §1 ↩ back to text

     Type: Federal antitrust statute

     Use and limitation: Agreement/restraint analysis; shared use of AI alone does not prove conspiracy.

     https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title15-section1&num=0&edition=prelim

120. S120 — Sherman Act §2, 15 USC §2 ↩ back to text

     Type: Federal antitrust statute

     Use and limitation: Monopolization elements; size or efficiency alone does not establish violation.

     https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title15-section2&num=0&edition=prelim

121. S121 — FLSA physician exemption, 29 CFR §541.304 ↩ back to text

     Type: Federal employment regulation

     Use and limitation: After-hours burden is not automatically an overtime violation for an exempt physician.

     https://www.ecfr.gov/current/title-29/section-541.304

122. S122 — OAL: Rulemaking frequently asked questions ↩ back to text

     Type: Official administrative guidance

     Use and limitation: Explains petitions and 30-calendar-day denial notice.

     https://oal.ca.gov/faq/

123. S123 — California HSC §1278.5 ↩ back to text

     Type: State health-facility whistleblower statute

     Use and limitation: Facility definition and statutory conditions must be satisfied; not automatically every freestanding clinic.

     https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=HSC§ionNum=1278.5

124. S124 — FDA: Considerations for the Regulation of Generative AI-Enabled Medical Devices—discussion paper ↩ back to text

     Type: Agency discussion paper—not draft or final guidance

     Use and limitation: Does not establish product classification or new regulatory requirements; docket FDA-2026-N-7874; listed comment deadline October 19, 2026 has passed.

     https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request