Issue №02 · Week of June 15, 2026

Anthropic's most powerful model lasted three days before a Commerce Department order took it dark globally. Meanwhile, international courts began assigning liability for what AI systems say.

Inside: The Trump Administration's push for AI physicians, a Nature Medicine benchmark study that OpenEvidence wants retracted, 42 state AGs serving OpenAI, and AI may be inflating the bills it was supposed to shrink.

Anthropic's Most Powerful AI Lasted Three Days

Emerging Policy

Anthropic released Fable 5 in early June, its newest frontier model, which the company said outperformed anything it had made generally available before. It was reportedly good at finding software vulnerabilities — enough that Anthropic had withheld the underlying Mythos model from general release on cyber-risk grounds, restricting it to vetted cyber defenders and infrastructure operators.

Then it disappeared. On Friday, June 12, a Commerce Department order used export controls to target foreign-national access to Fable 5 and Mythos 5; Anthropic responded by disabling both models globally. The reported trigger was Amazon, an Anthropic investor — telling officials it had jailbroken the model. Anthropic disputes the basis, saying the same technique works on other public models that face no such controls.

Days earlier, Anthropic had argued the government should be able to block dangerous AI deployments. While ironic, Anthropic's fight dates back further: in February, Defense Secretary Pete Hegseth had designated the company a "supply chain risk" — a label normally reserved for foreign adversaries — after it refused to let Claude be used for autonomous weapons and mass domestic surveillance.

The models are still dark. Anthropic is reportedly negotiating with Commerce over terms that could partially restore them.

Why it matters: Hospitals are adopting AI faster than any governance can keep up — and an overnight shutdown is a reminder that you don't control the tools you've come to depend on.

Can You Trust The Answer?

Research, Evidence & Benchmarks

Surgeons want AI in the OR — just not the kind being built: A 53-surgeon survey across five continents found 86.8% believe intraoperative AI can improve performance, yet 79.2% have never used such a tool; anatomy recognition and risk detection scored highest, while decision guidance ranked last.

OpenEvidence is everywhere, but answer fidelity is under scrutiny: the app fielded 30 million clinical questions last month — while a pharmacy journal review documented medication-answer inaccuracies and source-summarization errors, with the authors noting adoption by pharmacists and clinicians is likely to keep growing regardless.

A Nature Medicine study found frontier models outperformed it: Researchers testing GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6 against OpenEvidence and UpToDate Expert AI found general-purpose LLMs won all three benchmarks, including 100 real physician queries reviewed blind by 12 clinicians; OpenEvidence has since asked Nature Medicine to retract the paper.

KPMG's AI report became a case study in what it was supposed to warn against: After research firm GPTZero alleged that only 5 of 45 citations in KPMG's agentic AI report matched their sources, with roughly half of factual claims described as false or unsupported, KPMG pulled the report and said it is reviewing how it was published.

Who Owns AI Liability?

Policy & Regulation

UK doctors and the NHS face unresolved liability exposure as AI embeds in clinical care: The Medical Protection Society is warning that clinicians can be held negligent for AI-caused patient harm under current law and is calling for AI tools to be reclassified as products under the Consumer Protection Act 1987, shifting accountability toward developers.

Most health AI tools operate outside the regulatory oversight providers assume they have: A Bipartisan Policy Center report finds that ambient scribes, prior authorization systems, and other common health AI applications largely fall outside FDA review, with oversight fragmented across five federal agencies and variable state rules — and a proposed rollback of EHR transparency requirements could increase provider liability further.

42 state attorneys general have subpoenaed OpenAI over ads, health data, minors, and model safety: The multistate investigation, spearheaded by New York AG Letitia James and served June 12, came five days after OpenAI reportedly filed IPO paperwork with the SEC; Florida separately sued the company June 1 over alleged marketing of ChatGPT to children and concealment of safety risks.

A Munich court held Google liable for false claims its AI Overviews generated: The court rejected Google's defense that users understand not to trust AI results, ruling that fabricated output surfaced prominently in search is attributable to Google, not third-party sources — a distinction that matters for any company deploying AI-generated content in user-facing products.

Hallucinations and liability aren't limited to healthcare: A federal judge sanctioned all four lawyers in a Mississippi case after both sides submitted AI-generated filings with unverified citations, ultimately canceling the trial.

Is the Future Robot Doctors, or Boring Automation?

Hospitals and Care Delivery

• The Trump administration is pushing AI toward physician work: backing an AI prescription-writing experiment in Utah, granting $50M+ in conversational AI for cardiovascular care, and working on a regulatory pathway for independent AI physicians. UCSF's chair of medicine warned that AI will eventually be trusted beyond what it deserves, with potentially fatal consequences.

• Virtual nursing may cut readmissions: A nine-hospital UNC study found virtual nursing at discharge was associated with a 72% relative drop in 30-day ED returns (3.7% vs. 13.3% under standard care), with the largest gap in rural hospitals.

• Discharge-summary tools cut burnout but introduce errors: Stanford's MedAgentBrief discharge summary tool reduced physician burnout scores, but a review of 100 AI-generated drafts found omissions in 25% and inaccuracies in 20%.

• Nurses are bargaining for AI guardrails: Munson Medical Center nurses ratified a three-year contract with formal AI guardrails, approved by 93% of the Traverse City Munson Nurses Association. The provisions define AI contractually and create a channel for nurses to raise concerns before new tools are deployed.

Frontier AI’s Healthcare Foray

Operations & Markets

• Token-based pricing is straining budgets as AI moves from pilots to enterprise scale — while CIOs say ambient documentation is finally delivering EHR ROI at systems like St. Luke's and UMMC.

• AI may be making care less affordable: by enabling richer documentation and higher-complexity billing; PwC projects a 9% increase in employer market medical costs next year.

• Abridge touted partnerships with Nvidia and Eli Lilly (with interest from Cigna and Aetna), expanding from ambient documentation into foundation models for clinical conversations, real-time claims adjudication, and evidence-based care and research.

• Anthropic made five major moves in healthcare this year: a HIPAA-compliant Claude platform, a $200M Gates Foundation commitment, cybersecurity coverage for 150 organizations, new drug-design models, and federal oversight policy frameworks.

• SpaceX's S-1 filing reveals healthcare ambitions ahead of its June 12 IPO at a $1.75T valuation: rural hospital connectivity, telemedicine, orbital drug manufacturing, and AI-assisted diagnosis.

By CareChronicle. Did we miss something? Let us know.