Comparative intelligence · California · WHO/Europe · European Union · United Kingdom · Australia
The next phase of health AI is institutional governance, not model novelty
Within a single week four jurisdictions moved, and none of them moved on model capability. California legislated duties. WHO/Europe published the evidence that institutions are not ready to discharge them. The European Commission designated a general AI service under platform law. The United Kingdom proposed to make device regulation easier to update. The common subject is governance, and the common gap is readiness.
California’s 2026 health-AI bills against WHO/Europe’s governance principles
| Dimension | California instrument | Second instrument | WHO/Europe finding | What the comparison shows |
|---|---|---|---|---|
| Governance structure | AB 1979 (health care AI framework) — passage complete, enrolling | SB 503 — enrolled, developer/deployer duties as described in floor material | Governance gaps named as a primary barrier | The bills allocate duties; the report asks whether any institution is ready to discharge them. Allocation without readiness produces paper compliance. |
| Transparency | AB 2575 — tool inventory and disclosure provisions; Senate-amended text not yet read | — | Transparency to clinicians and patients treated as a precondition, not an output | California is legislating inventories. Inventories tell an institution what it runs; they do not tell a patient what touched their care. |
| Professional oversight | SB 903 — would bar AI systems from providing or advertising therapy | AB 2575 — health-care-worker use and override, per the amended text | AI literacy deficits identified as a persistent barrier | Both jurisdictions have arrived at the clinician as the control point. Neither has resolved what competence to operate a specific system means. |
| Bias and equity | SB 503 — identification and mitigation of known or reasonably foreseeable biased outputs | — | Biased and fragmented datasets named first among barriers | The strongest convergence on the table. California would impose duties; WHO/Europe supplies the reason they are necessary. |
| Accountability | SB 503 — developer and deployer duties | AB 1979 — framework duties | Unclear accountability identified as a barrier | Developer/deployer allocation is the most consequential design choice in either instrument, because whatever is unallocated lands on the clinician. |
| Patient participation | — | — | Patient and clinician co-design recommended | A gap in the California record. No tracked 2026 bill reaches patient participation in AI governance. |
Six threads worth following
Every instrument surveyed ends at the clinician, because the clinician is the only actor already licensed, insured and disciplinable. That is administratively convenient and analytically wrong: responsibility should follow control, and a clinician using a system she cannot inspect does not control it. This is the thread the record-integrity proposal attacks directly.
No medical board in the surveyed jurisdictions has adopted risk-tiered AI guidance. The KPSGILL forecast is that contested discipline cases will arrive before guidance does, and the falsification condition is stated on the ledger.
SB 503, as floor material describes it, splits duties between developer and deployer. That split is the most consequential drafting choice in the 2026 California set, because whatever falls between the two lands on the person holding the stethoscope.
Bias moves from an ethics topic to a statutory duty when a bill requires identification and mitigation of known or reasonably foreseeable biased outputs. The unresolved question is the standard of care: reasonably foreseeable by whom, assessed against which population?
The EU’s designation of a general AI service under the Digital Services Act is not medical-device regulation, but it applies systemic-risk duties touching physical and mental well-being to a service millions use for health questions. Two regulatory regimes are converging on the same conduct from different statutes.
If a clinician is disciplined over an AI-assisted decision, the evidentiary record must show what the system produced and what she did with it. Where that record does not exist, the investigative-phase proposal is the operative safeguard, not AI policy.
What follows for Reform
This analysis does not itself recommend anything. What it establishes is that the KPSGILL framework has independent international support for its central claim — that readiness, not capability, is the binding constraint — and that the California record now supplies concrete duty language to test the framework against. Two consequences for the agenda:
First, the AI Governance Readiness Framework is published as a proposal with its counterargument attached, and its risk-tiering remains unbuilt and is stated as unbuilt.
Second, patient participation is an identified gap in the California record. WHO/Europe recommends patient and clinician co-design; no tracked 2026 California bill reaches it. KPSGILL opens that as an evidence gap rather than a proposal, on the same basis as coverage architecture: a position without a published record behind it is an opinion.
Records: WHO/Europe report · UK MHRA amendments · dossiers: AB 1979, AB 2575, SB 503, SB 903.