Who Gets the Last Word? AI, Work, and Education After the Election

California has started assigning responsibility for automated decisions. The next test is whether workers, students, and public institutions have enough information, time, and authority to exercise it.

A worker receives a poor performance rating from a system whose measurements she cannot inspect. A student receives instant feedback from a tutoring application but cannot tell whether it has misunderstood the assignment. A college signs a systemwide software contract before faculty have agreed on what students should be allowed to delegate. A small employer sees an opportunity to reduce paperwork but discovers that the software also ranks applicants.

These situations share a governance problem. Who gets to decide whether the machine's answer is good enough, and what happens when a person disagrees?

The election changes the officials who will supervise parts of this transition. It does not reset the legal calendar. California enacted several relevant measures in September 2026. Some obligations have later operative dates or require institutions to develop standards. The incoming governor and legislature will inherit those tasks. A Democratic Congress can investigate the use of automated systems, negotiate federal legislation and appropriations, and press agencies for evidence. It cannot make every desired employment or education rule effective through a hearing alone.

This article's central forecast is that the next two years will be shaped by implementation, procurement, labor relations, and disputes over particular decisions. A comprehensive national settlement on AI remains a scenario. California already has a more immediate question: can it turn a right to human judgment into a working process without overwhelming the people expected to exercise that judgment?

The distinction matters in Fresno and the Central Valley. A technology that helps a small organization translate a notice or complete a form may be valuable. The same broad category of technology can create a serious problem when it determines whose application is rejected or whose work is considered inadequate. Treating every use as identical would make both public oversight and practical adoption worse.

What the adoption numbers actually measure

The Census Bureau's Business Trends and Outlook Survey offers a useful starting point. In its May 2026 account, the Bureau reported that approximately one fifth of businesses were using AI. But the survey's wording changed on November 17, 2025, from use in producing goods or services to use in any business function. The newer estimates should not be joined to the older series as if the question had remained constant.¹

These are dated national survey estimates, not a November census of every employer. They do not measure student use. They also do not support a claim that one fifth of American jobs has been automated. Adoption is a description of use; productivity, quality, displacement, and wages are separate outcomes.

That separation should govern the policy debate. An organization may use AI extensively while retaining its workforce. Another may use one consequential automated process that substantially affects a small group. Counting subscriptions or logins will miss the difference. A useful oversight program must ask what the system does, whose position it changes, and whether the affected person can obtain a meaningful correction.

A century of recurring questions, with a new technical form

American institutions have spent much of the last century negotiating the relationship between technology, authority, and opportunity. Mechanized production, the expansion of office work, mass communication, computing, and the internet each changed the tasks organizations could perform. None supplied its own answer to how gains should be distributed or how errors should be challenged.

The recurring questions are familiar. Does an employer use increased output to improve compensation, expand production, lower prices, or reduce staffing? Does a school use a new tool to give teachers more time with students or to substitute a cheaper interaction for a difficult educational relationship? Does a public agency simplify access or transfer more administrative work to the applicant?

Those are choices about institutional design. Technology affects the menu of possibilities; contracts, budgets, law, competition, and professional practice help select the outcome.

The last forty years provide a particularly useful comparison. As organizations moved more work into software, decisions increasingly depended on classifications, databases, and standardized workflows. A person could be harmed by an incorrect record long before a generative model produced a persuasive paragraph. AI can intensify that problem by making an uncertain inference appear fluent, precise, and administratively convenient.

The important historical lesson is therefore more specific than a prediction that all new technology eventually creates better jobs. A new system can create aggregate benefits while particular workers, communities, or students experience lasting losses. Adjustment costs do not disappear because a national total eventually improves. Equally, protecting every existing task indefinitely can deny people useful tools and leave institutions less capable of serving them.

A serious policy response needs to recognize both propositions. It should make legitimate experimentation possible while requiring stronger evidence and clearer remedies when a system controls a consequential decision. The standard for drafting an internal meeting summary need not be the standard for terminating an employee, assigning a student to a restrictive program, or screening a candidate out of public employment.

History also cautions against announcing the result before the organizational changes occur. A demonstration can show that software completes a task. It cannot, by itself, show how employers will redesign jobs, what teachers will stop doing, which complementary skills will become scarce, or how a county agency will handle appeals. Those are empirical questions for the next administration and Congress.

The California rules now in the pipeline

The governor signed an expanded group of AI measures on September 30, 2026. The employment and higher education provisions discussed here are enacted legislation, rather than campaign proposals. Their scope and timing differ. Describing all of them as a single immediate ban on automated decision-making would be inaccurate.²

SB 947 is particularly important because it addresses how a consequential employment decision is reached. Its chaptered text prohibits specified uses connected to labor, safety, and civil-rights violations and provides a route for obtaining certain information about a primarily automated disciplinary or termination decision. Enforcement is assigned to public authorities, with statutory exceptions and limits. An employer should not assume that adding an employee's signature to a machine-generated recommendation necessarily answers every legal question.³

SB 951 has a different purpose. It improves visibility into certain displacement events. Cal-WARN generally concerns covered establishments with at least 75 employees and includes a mass-layoff threshold of at least 50 employees in a 30-day period, subject to the statute's definitions and exceptions. The bill adds relevant information where AI accounts for displacement wholly or in substantial part. It also calls for state reporting on employment effects. This is a notice and evidence mechanism; it is not a guarantee that a displaced worker receives an equivalent job.⁴

AB 2392 addresses institutional preparation. Its framework includes participation by academic, labor, privacy, student, and other relevant representatives. The interim contract report is due within 60 days after execution, not as a universal condition that must be satisfied before a contract can be signed. That distinction matters when evaluating whether an institution complied and whether its process was adequate.⁵

These measures create three different policy instruments: constraints on decisions, visibility into displacement, and standards for institutional use. Their effectiveness should be evaluated separately.

Human review has to change the decision process

Human review can become a ceremonial step. A manager may technically approve a recommendation while lacking the time, evidence, or authority to reject it. A teacher may be assigned to supervise automated feedback across so many students that supervision becomes a search for obvious failures. A public employee may be held responsible for a decision generated by a system whose rules are controlled elsewhere.

The practical test is whether the reviewer can do something consequential. Can the reviewer inspect the relevant inputs? Can the affected person correct a factual error? Can the organization explain why the recommendation matters? Can a reviewer stop or reverse the decision without being penalized for departing from a productivity target?

An institution should distinguish three functions. Someone must check the underlying facts. Someone must assess whether the tool is suitable for the particular use. Someone must exercise the judgment that the institution is legally or professionally required to retain. A single checkbox does not establish that all three occurred.

Consider an illustrative scheduling system. It assigns fewer shifts to a worker after interpreting recent unavailability as a prediction of future unreliability. The factual issue may be whether the availability record is correct. The suitability issue is whether the prediction has enough relevance to justify its use. The judgment issue is whether the employer should make the proposed change in the actual circumstances. Different errors require different remedies.

This example is an analytical scenario, not a report about a particular company. Its value is that it exposes the gap between a technically reviewable output and a decision process that people can realistically challenge.

For policymakers, a strong next step would be to ask regulated organizations to describe the point at which human judgment enters and the evidence available there. Requiring an explanation of that workflow can be more revealing than asking whether the organization has an AI policy. Written policies matter, but a policy that cannot be followed during ordinary operations offers weak protection.

Who captures the gains at work?

The labor question extends beyond job counts. AI could make some work easier while making the remaining work more demanding. A person may spend less time drafting routine material but more time checking exceptions, resolving ambiguous cases, and explaining decisions to frustrated customers. If management counts only the tasks that were accelerated, it may underestimate the labor needed to maintain quality.

A realistic productivity assessment therefore needs a full cost account. It should include purchasing, integration, training, correction, security, appeals, and the cost of errors. It should also count benefits that are difficult to capture in a short financial report, such as faster service or improved accessibility. Neither a vendor's demonstration nor a single employee's frustrating encounter supplies the complete picture.

Distribution depends on bargaining and organizational choices. Workers may seek paid training, limits on surveillance, access to the information used to assess them, and a voice in redesigning roles. Employers may seek flexibility to improve output and retain competitiveness. Customers may value lower prices and quicker service. The resulting settlement can produce shared gains, concentrated gains, or gains offset by deterioration in quality.

The new congressional majority could make these distributional questions more visible through hearings and requests for data. Federal appropriations could support research, workforce training, and public-sector capacity. Changes to private employment obligations would require the applicable legal path, including passage through both chambers and presidential action for new federal legislation. Congressional interest does not itself create an enforceable workplace right.

California can act through its own labor law, public employment practices, procurement, and education systems, subject to governing legal limits. Its influence is substantial, but officials should distinguish the requirements they can enforce from the outcomes they hope to encourage. A training appropriation can purchase instruction. It cannot guarantee that employers will offer enough suitable positions or that workers can afford the time required to participate.

Schools need a theory of learning, not only an acceptable-use policy

In education, the central question is what students should learn to do independently and what they should learn to do with assistance. Those are educational choices. A prohibition or permission rule cannot substitute for them.

A teacher might allow a tool to help students generate competing explanations and then require them to test those explanations against evidence. Another assignment may require unaided writing because the purpose is to practice constructing an argument. A technical course may appropriately teach students how to examine generated code. A foundational exercise may instead assess whether students understand the underlying procedure.

These are not inconsistent positions. The legitimate use depends on the learning objective and the stage of instruction. An institution that treats every assignment the same will either restrict useful practice or make its assessments difficult to interpret.

The governance problem becomes harder when students do not have equal access to devices, reliable internet, paid software, quiet space, or assistance from an experienced adult. A school may inadvertently reward access to a better tool while claiming to measure mastery. Conversely, an appropriately designed public service could make support available to students who previously had little. Both outcomes are plausible; neither should be assumed from the technology alone.

Institutions should therefore evaluate a tool through evidence about learning, accessibility, and student experience. A high satisfaction score is helpful but cannot establish that students retain the material. A rise in completed assignments does not necessarily show stronger reasoning. An apparent decline in teacher workload may conceal additional time spent repairing errors or changing assessments.

AB 2392's standards process offers California a place to address these questions in public higher education. Its existence should not be described as proof that all necessary safeguards are already operating. The incoming administration and institutional leaders will have to translate requirements into training, contracts, support, and actual teaching practice. The quality of that translation will determine much of the practical effect.

Procurement is where general promises become specific obligations

A public institution can have careful principles and still sign an inadequate contract. Procurement determines what the purchaser can inspect, what information the vendor must supply, whether material changes require notice, and what happens when performance deteriorates.

A useful procurement record should identify the particular use and the population affected. It should specify the evidence that supports that use, the limits of the evidence, the information necessary to investigate an incident, and the process for ending or changing the arrangement. These are proposed governance practices, not a claim that every item is already required by every California AI statute.

The NIST AI Risk Management Framework provides a voluntary reference for managing risks. It does not certify a product as safe or replace applicable law. Its value for a purchaser is organizational: it offers a way to ask consistently who is responsible, what is being measured, and how risks will be monitored as use changes.⁶

A small district or employer may not have the staff to perform an extensive technical review. That creates an argument for shared expertise, common contract provisions, and cooperative evaluation. It also creates a risk that a standardized approval will be treated as permission for uses that were never tested. A system suitable for summarizing public information may not be suitable for evaluating a person.

Procurement should include a plan for leaving. If an institution cannot recover necessary records, preserve continuity, or move to another provider at a reasonable cost, its future bargaining position weakens. A low introductory price can therefore be a poor guide to the long-run cost of the arrangement. The relevant question is what the institution will need to operate responsibly over the life of the service.

Fresno's test: connect technology policy to local capacity

Fresno County's community health assessment and improvement planning identify education, internet access, economic stability, childcare, and safe opportunities for physical activity among the conditions relevant to community well-being. The county's 2024–2028 improvement plan was revised in May 2026. That local record supports treating technology adoption as part of a wider capacity problem rather than an isolated software purchase.⁷

For a Central Valley worker, a retraining opportunity may be unusable if its schedule conflicts with employment or caregiving. For a student, access to an application may accomplish little if the course assumes background knowledge the student has not yet acquired. For a small organization, a requirement to provide human review may fail if no one has been trained to investigate an error.

Fresno State, community colleges, schools, employers, worker representatives, and public agencies offer different potential points of coordination. Naming those institutions does not establish that a particular partnership or funded program exists. It identifies where a practical regional strategy would need to connect instruction, employer demand, public services, and evaluation.

A credible local initiative could begin with a limited set of use cases and publish its results. It might assess whether a tool reduces the time needed to complete a routine administrative task while preserving accuracy and accessibility. It could then examine who benefits, who needs extra support, and whether the time saved actually improves service. This would generate more useful evidence than an undifferentiated announcement that the region is embracing AI.

The region should also avoid defining success entirely through the recruitment of new technology firms. Improvements within existing organizations can matter directly to residents. A school that gives teachers more effective preparation time, a small employer that handles paperwork more reliably, or a county office that makes an application easier to understand may produce substantial value even without a headline investment announcement.

The strongest local safeguard is a clear connection between a claimed benefit and an observable result. If a system is purchased to improve access, measure access. If it is purchased to reduce error, measure errors and corrections. If it is purchased to support learning, evaluate learning with an appropriate design. The procurement justification should not quietly change when the original benefit proves difficult to demonstrate.

Three plausible paths through 2028

Managed adoption is the base case. California institutions implement the new requirements unevenly, improve some procedures, and continue experimenting. Congress increases scrutiny and negotiates selected measures. Some organizations obtain meaningful gains; others discover that integration and supervision cost more than expected. This is a directional judgment, not a measured probability.

A stronger public-capacity path is possible. Institutions pool evaluation resources, make training usable, and publish meaningful results. Workers and educators participate early enough to shape the use rather than merely respond to it. Public purchasers obtain clearer rights to inspect and remedy failures. The result could be wider access to useful tools with better protection against consequential mistakes.

A compliance-without-capacity path is also possible. Organizations produce notices, policies, and training records while leaving frontline staff unable to exercise judgment. Vendors and purchasers treat documentation as evidence of effectiveness. Displacement becomes visible only after it happens, and educational adoption proceeds faster than evaluation. The formal requirements exist, but their practical value is limited.

The strongest challenge to this article's cautious forecast is that rapid improvement in tools could make some current oversight practices unnecessary or too costly. That possibility deserves attention. A process that remains burdensome after a particular risk has demonstrably fallen should be revised. But improvement should be shown for the actual use and population. General technical progress does not establish that a particular employment or educational decision has become reliable.

The opposite challenge also matters: a major failure or a sharp labor-market disruption could produce much stronger restrictions than the base case anticipates. That would change the legislative incentives and the urgency of public intervention. Neither optimistic nor restrictive scenarios should be treated as inevitable.

What would change the forecast?

The most useful indicators are operational. Watch the implementation of SB 947 before its July 2027 operative date; the quality of explanations provided for consequential employment decisions; the technology-related information appearing in covered displacement notices; and the standards and training recommendations presented to system leaders under AB 2392 by January 1, 2028.

Also watch results that legislation cannot establish on its own: corrected decisions, time spent on review, learning outcomes, access across different groups, worker retention, service quality, and total operating cost. Track these with a clear reference period and a meaningful comparison. A change after adoption is not automatically a change caused by adoption.

For Fresno and the Central Valley, the decisive question is whether institutions build the capacity to use and challenge these tools. A person cannot exercise judgment merely because a policy says that judgment remains human. The person needs information, time, training, and the authority to act on a disagreement. The post-election policy test is whether those conditions become part of ordinary work and education.

Related analysis

Read Who Is Responsible When the Algorithm Enters the Exam Room? California’s New Test for Clinical AI for clinical AI and physician judgment; Opportunity Needs an Operating Budget: Schools, College, Research and Child Care After the Election for education, research, college costs, and family opportunity; After the Election, What Will Change at Work—and Who Will Pay? and After the Election, Follow the Household Dollar: Taxes, Benefits, Wages, and the Fiscal Bargain for employment and household economic conditions; and A New Congressional Majority, a Divided Government: What the 2026 Election Can Actually Change for the governing balance and the limits of congressional power.

Sources and scope

This article distinguishes enacted requirements from proposed practices and the author's forecasts. Adoption estimates are dated survey results, not measurements of job displacement or educational achievement. The scenarios are reasoned assessments rather than calibrated statistical probabilities.

Source notes

Numbered notes support the sentence that carries the number. A source supports the claim it is cited for, with its own date and scope; it is not an endorsement by the issuing institution.

1. U.S. Census Bureau, “AI Use by Businesses”, May 26, 2026. Business Trends and Outlook Survey results and the November 17, 2025 questionnaire change. The table preserves the May reference period and relevant business denominators. ↩

2. Office of the Governor of California, “California's nation-leading AI framework just got stronger”, September 30, 2026. Signing announcement; statutory text controls individual obligations and operative dates. ↩

3. California SB 947, Chapter 859, Statutes of 2026, chaptered bill text. Operative July 1, 2027; employment automated decision systems, notice and information rights, enforcement, and exceptions. The linked page reproduces the chaptered statutory text; the official enacted text controls. ↩

4. California SB 951, Chapter 860, Statutes of 2026, chaptered text and version history, 2026 session; see also the bill record. California WARN thresholds, AI-related notice provisions, and reporting. Apply the chaptered version and statutory exceptions when assessing an actual event. ↩

5. California AB 2392, Chapter 855, Statutes of 2026, chaptered text and bill record; legislative history. Public higher education generative-AI standards, training, and interim systemwide contract reporting. UC participation is requested, not described here as identical to the statutory direction for CSU and community colleges. ↩

6. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 26, 2023. Voluntary risk-management framework; not a product certification or a substitute for governing law. ↩

7. Fresno County Department of Public Health, Community Health Assessment and Community Health Improvement Plan, 2023 assessment and 2024–2028 plan, revised May 2026. Establishes locally identified priorities, not the effectiveness of any proposed AI program. ↩