AI in Healthcare: Vermont just made the therapy bot a crime  June 21st 2026

The healthcare AI story this week is a regulator drawing a hard line: Vermont became the first US state to make it unprofessional conduct to let an AI system diagnose or treat a mental-health patient on its own. Underneath the policy news, the clinical-AI map shifted again — AlphaFold co-creator John Jumper left Google DeepMind for Anthropic, a consumer AI lab pushed a 60-second full-body ultrasound scanner past the regulated lane, and OpenAI claimed its free model now out-scores physician-written answers on its own health benchmark. Four fronts, one theme: the guardrails and the gold rush are now arriving in the same week.

Key Takeaways

• Vermont set the precedent builders feared. Deploying a wellness or therapy-adjacent product in Vermont now carries licensing-board and consumer-protection exposure the moment the model makes a therapeutic decision unsupervised — audit your product surface there before August.
• Anthropic is buying clinical-science credibility, not just compute. Landing AlphaFold's Nobel laureate signals that "Claude for biology" is a roadmap, not a demo — watch which life-science partners follow him out of DeepMind.
• The FDA boundary is the new product strategy. Midjourney's scanner ships as "wellness" specifically to avoid 510(k)/De Novo review — expect more consumer-imaging hardware to launch in the unregulated body-composition lane and layer clearances on later.
• The patient-facing LLM is now a clinical claim. OpenAI says GPT-5.5 Instant beats physician-written responses on its in-house health rubric — a marketing line today, but it moves the liability conversation from "chatbot" to "medical advice" fast.

The Big Story

Vermont Gov. Phil Scott signs H.816, a therapy-bot ban, into law · Transparency Coalition · June 18, 2026
→ Governor Phil Scott signed H.816 on June 17, making Vermont the clearest state-level "AI cannot be the therapist" statute yet: it regulates the use of AI in the provision of mental health services so that licensed providers cannot offload therapeutic decisions to a model. For health-tech builders the clinical significance is the liability surface, not the headline — once a product crosses from administrative support into diagnosis or a treatment plan, it sits in the same enforcement lane as an unlicensed practitioner. The unresolved line is where "wellness chatbot" ends and "therapeutic decision" begins, and that ambiguity is exactly what compliance teams have to resolve before shipping.

Also This Week

Nobel laureate John Jumper is leaving DeepMind for rival Anthropic · TechCrunch · June 20, 2026
→ The AlphaFold co-creator, who shared a recent Nobel Prize in chemistry, is leaving Google DeepMind after nearly nine years for Anthropic — the strongest signal yet that the frontier labs intend to compete head-to-head on AI-for-biology. For drug-discovery and protein-design teams, the practical question is which foundation model their pipelines anchor to over the next 24 months as the talent that built structure prediction migrates to a new house.

Midjourney enters medical imaging with a 60-second full-body scan · PYMNTS · June 18, 2026
→ Midjourney's prototype runs 40 Butterfly ultrasound-on-chip modules under a co-development agreement worth up to $74 million over five years and produces a whole-body image in roughly 60 seconds with no radiation — but it is explicitly not FDA-cleared for diagnostic use, launching first as body-composition "wellness" that sidesteps clearance. That sequencing — ship unregulated, layer clearances on later — is the regulatory tell every imaging builder should study.

OpenAI says GPT-5.5 Instant now beats physician-written health answers · Search Engine Journal · June 18, 2026
→ OpenAI pushed a health update to its free GPT-5.5 Instant model, claiming responses now rate higher than physician-written ones across accuracy, communication and completeness on its own HealthBench rubric, with flagged-factuality issues down 71% over two months. The evaluation leans on OpenAI's internal benchmark and its panel of 260+ physicians rather than independent peer review — so read it as a positioning move that quietly reframes a free consumer chatbot as medical-grade advice, and watch how regulators respond to that framing.

Worth Reading

• AI Legislative Update — June 19, 2026 · Transparency Coalition · June 19, 2026 — A running multi-state ledger of AI bills — Arizona, California, Illinois, New York, Rhode Island and more — useful for tracking which jurisdictions move next on health and therapy AI.
• Improving health intelligence in ChatGPT · OpenAI · June 18, 2026 — OpenAI's own write-up of the GPT-5.5 health update and its physician-evaluation methodology — the primary source behind this week's "is the chatbot now medical advice?" debate.

Vermont didn't ban AI from mental health — it banned AI from being the one in the room without a license. That distinction is the whole fight.

— Alexis