Newsom signs two clinical AI bills, vetoes AI therapy limits; autonomous AI clears skin referrals; ARPA-H funds ICU digital twins; UVA cuts scribe hours

Four rolled paper scrolls stand upright in a navy desk tray; two are tied with amber ribbons and seals, two have loose gray ribbons.

Gov. Gavin Newsom signed two bills Sept. 30 that protect licensed clinicians' judgment over clinical AI tools in California and vetoed two others, including one that would have limited AI in psychotherapy.

The signed bills, AB 1979 and SB 503, take effect Jan. 1, 2027, and apply to physician offices and group practices as well as hospitals. The vetoes ended AB 2575, which would have protected nurses and other clinicians who override AI output, and SB 903, which passed the Senate without opposition. In the United Kingdom, an autonomous AI system discharged a quarter or more of urgent skin cancer referrals at two hospitals without clinician review, and missed six cancers, according to a congress presentation.

The Advanced Research Projects Agency for Health funded critical care digital twins at the University of Vermont and Stanford and a Stanford monitor for autonomous cardiology agents. UVA Health cut emergency department scribe hours as its ambient documentation tool spread, the student newspaper reported, and the American Hospital Association asked senators to extend HIPAA-level standards to AI vendors and help rural hospitals hire cybersecurity staff.

  • Regulation Sept 30, 2026
    Gov. Gavin Newsom signed two health-AI bills on Sept. 30, the last day to act on bills from California's 2026 session, and vetoed two others. AB 1979, by Assemblymember Mia Bonta, D-Oakland, requires health facilities, clinics and physician offices to take reasonable steps so licensed professionals can exercise independent judgment when a clinical decision support system's output informs care, bars using AI to direct unlicensed staff in, or independently perform, work that requires a license, and treats businesses offering health chatbots to consumers as providers under the state's medical confidentiality law. SB 503, by Sen. Akilah Weber Pierson, D-San Diego, requires developers and users of clinical decision support AI to make reasonable efforts to identify and reduce biased impacts, and requires users to monitor the tools regularly; both take effect Jan. 1, 2027. The vetoed bills were AB 2575, which would have barred retaliation against clinicians for overriding AI output, and SB 903, which would have kept AI from making therapeutic decisions or being advertised as therapy without a licensed professional's review. "The governor knows A.I. should never replace a nurse's clinical judgment, but he vetoed the bill that would have protected us for using our judgment," said Sandy Reding, president of the California Nurses Association, in a union statement.
  • Evidence Sept 30, 2026
    An AI system that reviewed smartphone and dermoscopic images without clinician review handled 8,391 patients referred urgently for suspected skin cancer at two U.K. hospitals over 16 months, discharging 31% at one site and 25% at the other, Dr. Lucy Thomas of Chelsea and Westminster Hospital NHS Foundation Trust reported at the European Academy of Dermatology and Venereology Congress in Vienna. The team said the pathway created about 8,500 face-to-face appointment slots, saved 2,851 clinician hours and cut biopsy rates to 27% from 43%. Sensitivity exceeded 98% for invasive melanoma, squamous cell carcinoma and basal cell carcinoma, with specificity of 72.1%; post-market surveillance found six false negatives, five basal cell carcinomas and one melanoma in situ, with no adverse outcomes reported, and 86% of patients consented to an autonomous decision. The device carries a European CE mark as a Class III medical device; the results were a congress presentation, and Dermatology Republic reported no journal publication.
  • Money Sept 30, 2026
    The Advanced Research Projects Agency for Health awarded the University of Vermont up to $38 million under its CIRCLE critical-illness program to build digital twins that simulate a patient's immune response and test approved and experimental therapies, with Dr. Gary An as principal investigator; the work runs five years, and VTDigger reported that wider clinical use is seven to 10 years away. Stanford Medicine received up to $41.4 million under the same program for IMPRINT, an AI immune digital twin to predict ICU outcomes, and up to $14.9 million under ARPA-H's $62.7 million ADVOCATE program for STEWARD, a system to monitor autonomous AI agents in cardiovascular care, built with OpenAI, the University of Washington and Tuple Health. "Our team will develop AI monitoring tools to ensure that autonomous agents remain safe, transparent and accountable," said Dr. Roxana Daneshjou, a co-leader; the researchers said no system now exists for regulatory-grade monitoring of autonomous health care agents, Becker's reported.
  • Workforce Oct 1, 2026
    The University of Virginia's student scribe program, founded in 2007, took a one-third cut to its annual budget, announced to scribes in April, and a new emergency department schedule from Sept. 15 eliminated shifts and reduced scribe availability, The Cavalier Daily reported Oct. 1. UVA Health adopted Microsoft's DAX Copilot ambient documentation tool in 2025. "I think that this AI tool came into place, and then we, as a kind of a healthcare system budget, we always have to make cuts," said Dr. Sarah Wendel, an assistant professor of emergency medicine, who also said she did not make the budget changes. About two of 36 emergency medicine residents actively use DAX when scribes are present, the paper reported; it cited no UVA Health statement on the reason for the cut.
  • Regulation Sept 30, 2026
    The American Hospital Association told the Senate homeland security subcommittee holding its Sept. 30 hearing on attacks by AI agents that entities holding or processing protected health information outside HIPAA should be "subject to similarly rigorous privacy and security standards." The statement said health care reported 460 ransomware attacks in 2025, more than any other critical infrastructure sector, and that people affected by health data breaches rose to 259 million in 2024 from 27 million in 2020. It said rural hospitals, 48% of which operated at a loss in 2023, face limited budgets, older devices and trouble hiring security staff, and it endorsed legislation from Sen. Josh Hawley, R-Missouri, S. 2169 and H.R. 9908, to build a rural hospital cybersecurity workforce. The hearing followed reports covered here Sept. 24 and Sept. 27 of OpenAI agents probing Australian health systems.
  • Money Sept 29, 2026
    A Peterson Health Technology Institute survey of 321 decision-makers at health plans, employers and health systems, conducted by KRC Research and released Sept. 29, found 68% of employers and 55% of health plans use performance-based contracts for digital health tools; 85% of those contracts put at least 25% of fees at risk and 31% put more than half at risk. Thirty-eight percent of purchasers said they cannot independently verify vendor-reported outcomes, and 45% reported trouble matching vendor data to their own claims data. Seventy-one percent of health systems had deployed AI for clinical documentation or ambient scribing, 41% across the enterprise, while 18% used clinical decision support AI at enterprise scale; 67% of health plans used AI to review clinical documentation. "Purchasers want to pay for outcomes, not for activity," PHTI's managing director said.
  • Evidence Oct 1, 2026
    Participants shown fictitious family doctor advertisements rated physicians who said they use AI as significantly less warm and less competent than those who said they never use it, and were less willing to book, follow treatment plans, recommend the physician, travel, pay or wait, according to a study by Moritz Reis and colleagues published Oct. 1 in npj Digital Medicine. The penalty, measured among 1,030 online participants recruited through Prolific, was largely independent of the physician's gender, age or race and was larger among participants younger than 45, the authors reported. The effect sizes were modest (eta squared of 0.04 for warmth and 0.05 for competence), and the design measured stated intentions, not actual visits.
  • Deployment Sept 29, 2026
    Heidi, the scribe company that raised $340 million a week earlier, released Heidi II on Sept. 29 with agents that prepare charts for a full day's schedule, handle referrals and follow-ups, and run panel sweeps for immunizations, blood tests, screenings and recalls across the EHR, calendar, inbox and patient portal. Each task requires a clinician's approval before it is completed, the company said; Heidi II launched in English and French and is not available in the United Kingdom or the European Union. The company said its scribe supports 2.8 million visits a week and surfaces 10 million follow-up tasks weekly; it published no outcome or accuracy data for the agents.
  • Money Sept 29, 2026
    Ortet, which describes itself as a frontier AI lab for health, launched Sept. 29 with a $500 million commitment from Thoreau, a health technology infrastructure company, for computing, model training and deployment, research and hiring. Its chief executive is Kyunghyun Cho, a New York University professor who helped develop the attention mechanism behind transformer models, and its chairman is Jeff Hammerbacher, a former Facebook data leader and Cloudera chief scientist. The company said it will build a patient-centered foundation model linking biological, clinical, operational and financial data; it named no health system partners, products or release dates.