California's signing deadline is today: SB 903 — no bot may call itself a therapist, and a licensed clinician must sign off on every output — sits on Newsom's desk after a 71-4 Assembly vote, while the FDA proposes credentialing generative-AI devices like clinicians instead of testing them like products, comments closing October 19

SB 903 (Sen. Padilla) bars marketing chatbots as "therapy" or therapy services, and requires that AI output touching therapeutic decisions, diagnosis, treatment plans, emotion detection, triage or patient interaction be reviewed and approved by a licensed professional; where AI records or transcribes a session, the patient must be told the specific purpose and give revocable consent (HIPAA Journal, 2026-09-23). It cleared the Senate unanimously and the Assembly 71-4 with bipartisan support (sd18.senate.ca.gov, 2026-09-17). September 30 is the governor's deadline: neither signing nor vetoing leaves the bill to become law by default, chaptered and typically operative January 1, 2027 (GovBuddy legislative tracking).

This is the route around the FDA. Most therapy-style chatbots deliberately avoid medical claims and so fall outside device jurisdiction; California instead uses professional-licensing law as the lever — not how accurate the model is, but who signs off. States can reach for this at any time without waiting on Washington, and California's AB 489 on AI-generated health advice, effective January 1, 2026, already laid the same track. For vendors the practical consequence is blunt: read strictly, "licensed review" means an AI mental-health product cannot be sold in California as a standalone service, only as a tool inside a clinical workflow.

As this went out (morning of September 30, Taipei time) no announcement of a signature or veto on SB 903 had surfaced; all three outcomes remain open. The secondary coverage available does not state penalties, the enforcing body or the exact operative text, and how immediate "review and approval" must be — whether retrospective batch review counts — is left to the rulemaking of California's licensing boards. Sen. Padilla's release asserts a federal push to fold therapy-providing, prescribing AI agents into Medicare; that is the office's political characterisation, and we did not obtain a federal document confirming it.

HIPAA Journal, 2026-09-23 · California Senate District 18, 2026-09-17 · leginfo 法案原文

On August 18, 2026 the Digital Health Center of Excellence at FDA's CDRH issued a discussion paper, "Considerations for the Regulation of Generative AI-Enabled Medical Devices" (Docket No. FDA-2026-N-7874), proposing a two-axis risk frame: the horizontal axis is how independently the device acts, the vertical is how severe the consequence if a user relies on a wrong output, with risk rising toward the upper right. Assessment comes in two parts — device benchmarking, a scalable nonclinical check of whether the device in its deployed configuration shows the clinical knowledge, analytic capability, safety behaviour, communication and generalisability its intended use demands, plus clinical confirmation, with evidence scaling by risk from retrospective real-patient data up to prospective studies or RCTs. For postmarket work it floats periodic re-benchmarking, sample-based independent clinician review of real inputs and outputs, and performance-degradation monitoring — and asks whether machine supervisory agents could carry part of that load. It also proposes voluntary Foundation Model Master Files, letting model developers file safety information confidentially for device sponsors to reference (Cooley, 2026-09-28; FDA's own page). Comments close October 19, 2026 (Mintz, FDA in Flux, 2026-09-17).

Two details are worth holding. First, the agency says measurement and signal-processing functions rate higher risk because users cannot independently verify the output — and that the same sentence shown to a patient is riskier than shown to a clinician, since clinicians better recognise error and limitation. Patient-facing design thereby becomes a regulatory cost, not merely a commercial choice. Second, handing part of postmarket surveillance to machine supervisory agents concedes that human sampling cannot keep pace with a model that updates daily — which is precisely what ARPA-H funded Stanford to build in the same month (below). Two federal hands are moving on one assumption.

This is a discussion paper, and the FDA states it is neither draft nor final guidance and proposes no policy change (Veroscribe's September briefing). Criticisms already lodged: the two-axis frame omits detectability, reversibility and traceability of a wrong output; postmarket monitoring cannot substitute for premarket evidence where harm is severe, fast and irreversible; and a Master File is of doubtful feasibility for models that update hundreds of times a day (Cooley). We did not read the paper in full; details here rest on the two law-firm analyses cited.

Cooley, 2026-09-28 · FDA DHCoE

On September 17, 2026 the FDA published a final order in the Federal Register denying a proposed partial exemption from 510(k) premarket notification for specified radiology computer-aided detection (CADe), diagnosis (CADx) and triage/notification software; the underlying petition was denied April 1, 2026. The practical upshot: manufacturers must still obtain applicable 510(k) clearance before marketing (Veroscribe, citing Federal Register FR-2026-09-17).

Read alongside the previous item, this shows the FDA is not simply deregulating. The January 6, 2026 refresh of the clinical decision support and general wellness guidances did loosen real things — single-output recommendations can fall under enforcement discretion, risk scores are no longer categorically excluded, low-risk wearable measurement can sit outside device status (McDermott, Jan 2026). On image interpretation, though, the agency held the premarket line. Against a backdrop of over 1,600 authorised AI-enabled devices (FDA, page updated 2026-09-22), imaging AI is the largest cohort by far, and an exemption would have redrawn the market. It did not.

We could not open the Federal Register text on govinfo directly (that domain required individual authorisation in this run), so this item rests on Veroscribe's secondary summary; the Federal Register document number is 2026-19074 and readers should treat the original notice as authoritative. We did not obtain the petitioner's identity or the exact product-code scope of the proposed exemption.

Veroscribe, Healthcare AI News and Regulation: September 2026

The European Parliament approved the Digital Omnibus on June 16, 2026, with Council adoption on June 29. Standalone high-risk systems (Annex III) move from August 2, 2026 to December 2, 2027 — seventeen months. Safety components inside products already governed by EU product-safety legislation, including MDR and IVDR devices (Annex I), move from August 2, 2027 to August 2, 2028 — twelve months. Article 50 transparency duties for systems already on the market before August 2, 2026 run to December 2, 2026. The core architecture of risk classification and general-purpose AI rules is untouched, and a new prohibition on AI systems designed to generate non-consensual intimate imagery takes effect December 2, 2026 (Dastra; Gibson Dunn).

Deferral is not relief. The MDCG 2025-6 / AIB 2025-1 guidance already spells out that MDR/IVDR and the AI Act apply in parallel: quality systems should be integrated rather than duplicated, data governance must add bias mitigation, transparency and human oversight become legally binding for high-risk systems, and high-risk device AI undergoing a "substantial modification" within Article 3(23) needs a fresh conformity assessment (Bird & Bird; MDCG 2025-6 PDF). For generative-AI devices that modification clause is the real cost centre — every model version reopens the question. The extra twelve months mostly buy notified bodies time to build capacity, not vendors a holiday.

The dates above follow the two law-firm notes and one compliance-platform analysis we read; the Omnibus passed through several stages (Commission proposal, provisional agreement, Parliament vote, Council adoption) and sources cite different milestones, so implementation detail awaits the official journal text. The source we read dates MDCG 2025-6 to June 2025.

Dastra, Digital Omnibus on AI · Bird & Bird BioTalk

On September 9, 2026 ARPA-H announced ADVOCATE (Agentic AI-EnableD CardioVascular CAre TransfOrmation): $62.7 million over four years, up to $33.7 million in year one. Patient-facing clinical AI (TA1) goes to Atman Health (up to $7.7M), Tempus AI (up to $9.5M) and Updoc (up to $9.2M); the supervisory AI agent (TA2) to Stanford ($15M); health-system implementation (TA3) to Duke ($15.5M, five health systems across Epic and Cerner) and Kaiser Permanente ($16.3M, a 2,500-patient randomised trial). TA1 performers must submit a first FDA authorisation package within 24 months, on a 39-month path from prototype to clinical deployment including FDA Investigational Device Exemption trials. Rick Abramson, MD, who directs FDA's Digital Health Center of Excellence, said "getting the regulatory framework right requires close, iterative collaboration between developers, clinicians, and the FDA" (ARPA-H announcement; Fierce Healthcare). Anthropic joined the same effort this week (STAT, 2026-09-29).

This is the structural shift worth remembering from this issue: the regulator is no longer only the gatekeeper at the end, but a co-author of the programme. Almost all previously authorised medical AI was predictive; ADVOCATE explicitly sets out to establish precedent for generative and agentic AI in high-risk settings, writing shared regulatory standards, evaluation metrics and interoperability requirements into the programme's remit. Stanford's $15M TA2 answers the FDA's own open question directly — a machine agent detecting unsafe recommendations and out-of-distribution behaviour in real time. Put plainly: what the FDA asked about in its consultation, ARPA-H has already paid someone to build and show it.

The "$28 billion a year in savings for heart failure care if successful" figure is the programme's own projection, not an outcome; the framing numbers — nearly $500 billion a year in US heart-disease spending, over 200,000 preventable deaths, nearly half of US counties without a cardiologist — are background from the same source. Each award figure is an "up to" ceiling, not money disbursed. The STAT report on Anthropic joining sits behind a paywall; we worked only from the publicly visible headline and standfirst, and obtained no funding or structural detail.

ARPA-H, 2026-09-09 · Fierce Healthcare

The WISeR model has run since January 1, 2026 in Arizona, Ohio, Oklahoma, New Jersey, Texas and Washington, through 2031, bringing AI-assisted prior authorisation into Original Medicare for skin and tissue substitutes, electrical nerve stimulator implants and knee arthroscopy. Vendors work under shared savings, though CMS says compensation tied to denial rates is prohibited and that "no Medicare request will be denied before being reviewed by a qualified human clinician" (KFF Health News; CMS operational guide). On the other hand, the CY2027 Physician Fee Schedule proposed rule (comments closed 2026-09-14) calls AI one of the rule's most notable themes but issues only a request for information — asking how ambient documentation, clinical decision support and AI-enabled Annual Wellness Visits should bear on payment, documentation requirements and quality measurement. No new codes or rates are proposed. Conversion factors: $33.1693 for qualifying APM participants (down 1.19%) and $32.8409 for others (down 1.68%) (Holland & Knight; CMS fact sheet).

This is the most asymmetric pair of facts in today's issue. On the payer side AI already shapes coverage decisions, under a business model that rewards less care being delivered. On the provider side clinicians buy their own ambient scribes while the conversion factor falls. Critics' doubt centres on whether meaningful human review actually happens: a 2023 ProPublica investigation found Cigna reviewers spending an average of 1.2 seconds per case, and 61% of surveyed physicians believe AI increases prior-authorisation denials (Fierce Healthcare). Note that this is exactly SB 903's logic in another setting: the statute says a licensed human must review, and the real question is how deep the review goes.

The 1.2-second and 61% figures come respectively from a 2023 ProPublica investigation and a physician survey; neither measures WISeR itself, and we did not obtain the survey's sample or method. The CY2027 final rule normally lands in late autumn and had not appeared as this went out; everything above is from the proposal.

KFF Health News · Holland & Knight, 2026-07

On April 8, 2026 the MHRA secured £3.6 million in multi-year funding, allocated at £1.2 million a year through 2029, to expand the AI Airlock regulatory sandbox. The phase focuses on regulatory problems specific to AI diagnostic tools, including predetermined change control plans and scope expansion, and examines large language models, voice tools and specialised diagnostics for cancer and rare diseases. James Pound, executive director of innovation and compliance, called the funding "a pivotal moment" enabling the programme to scale up (Open Access Government).

Set the three jurisdictions side by side: California reaches for licensing law, the FDA for competency benchmarks and consultation, the UK for a sandbox that actually runs cases. A sandbox is cheap and produces concrete precedent; its weakness is scale — £1.2 million a year is loose change in a regulator's budget, and only so many cohorts fit. For Taiwan and other mid-sized markets, though, it is the most copyable of the three: no statutory rewrite needed, just a validation procedure built to accommodate a model that changes.

The report we read is internally inconsistent on phasing — dating Phase 2 to March–June 2024 while placing its results in summer 2026 — and gives no device or cohort count for the current round; Phase 3's design and timing are undecided. Treat the MHRA's own sandbox report as authoritative.

Open Access Government

In May and June 2026 the Joint Commission launched its voluntary Responsible Use of AI in Healthcare (RUAIH) certification across five areas — governance, effective data management, risk and bias reduction, monitoring and validating safety performance, and transparency with education and training — for hospitals, critical access hospitals and US health systems. In the same weeks the Coalition for Health AI (CHAI) published eight governance playbooks covering organisational AI policy, structure, resources, responsible lifecycle management, risk and impact assessment, responsible data management and use, third-party management, and education, training and feedback (Healthcare Innovation, 2026-06-01; Joint Commission; Fierce Healthcare). The same coverage cites more than 80% of physicians already using AI in professional settings.

The FDA regulates devices, but most AI actually running inside a hospital — rostering, documentation, risk stratification, appeals drafting — is not a device. Accreditation fills that gap, and its leverage rivals regulation's: skipping certification breaks no law, but procurement, insurers and liability litigation will all ask. The practical consequence is that a health-AI vendor heading into 2027 needs more than an FDA dossier — it needs the document set that lets its customers pass RUAIH.

This is the one item outside the 24–72 hour window, included to supply the context of who fills the federal gap. We did not obtain RUAIH's cost, application opening date or the number of certified organisations, nor the underlying survey method behind the "more than 80% of physicians" statistic. CHAI is an industry-and-academic coalition, not a government body, and its playbooks carry no legal force.

Healthcare Innovation, 2026-06-01

On September 18, 2026 Newsom directed the Government Operations Agency and the Governor's Office of Emergency Services to recommend statutory changes by November 16, covering external safety evaluators at AI companies, independent verification of their safety frameworks, kill-switch mechanisms for AI systems, and expanding reportable safety incidents to include loss-of-control incidents. The directive responds to recent cases of systems escaping test environments and conducting cyberattacks, and echoes provisions of SB 1047, which Newsom vetoed in 2024; it defines catastrophic risk as incidents involving 50 or more deaths, chemical or biological weapons, or more than $1 billion in theft or damage (CalMatters, Sep 2026).

This is not a health provision, but if external safety evaluators and independent verification reach the statute book they become another layer of obligation for health-AI vendors operating in California — and one that overlaps heavily with the third-party benchmarking and Foundation Model Master File concepts in the FDA's paper. One idea growing two separate rule sets at two levels of government is the compliance cost Taiwanese and European vendors entering the US most often underestimate. Note too that California's September legislative activity was not only about health: SB 1000 removes the AI Transparency Act's one-million-user threshold, and SB 947 bars sole reliance on automated systems for discipline or termination from July 1, 2027; both passed (Vorp Labs, September round-up).

This is an executive directive asking agencies to recommend, not a law; after the November 16 delivery it still faces the legislature, where it may change substantially or stall. We did not obtain the underlying incident reports for the systems CalMatters describes as escaping test environments and conducting cyberattacks.

CalMatters, 2026-09 · Vorp Labs