AI is failing differential diagnosis at rates above 80 percent on incomplete clinical data, and the tools available to solo practitioners carry none of the governance scaffolding that institutional settings build around them. The clinical question is identical: is this tool safe for the person in front of me? The infrastructure to answer it is not.
Signal 1
AI Fails Differential Diagnosis 80%+ of the Time — New JAMA Study [WARN]
A study published April 13, 2026, in *JAMA Network Open* by researchers at Mass General Brigham evaluated 21 large language models on clinical reasoning tasks using standardized patient vignettes. All 21 models failed to generate an appropriate differential diagnosis more than 80% of the time when working from incomplete clinical information — failure rates reached 90–100% in several models. Final diagnosis accuracy was above 90% only when all pertinent data was supplied upfront. Lead researcher Dr. Marc Succi was direct: these models "are not ready for unsupervised clinical-grade deployment." The study introduced a new benchmark metric — PrIME-LLM — that scores models across the full reasoning arc rather than just final-answer accuracy. Scores ranged from 64% to 78% across all models tested.
What this means for you
AI tools that arrive at a correct final answer do not necessarily reason correctly to get there. If you are using AI to support clinical pattern recognition or assessment logic, the tool may confirm what you already believe while missing what you do not. The liability is yours, not the tool's.
Signal 2
Maine Bans Unlicensed AI Therapy; Four Other States Advance Health AI Bills [WARN]
Maine's legislature passed LD 2082 in the week of April 6-13, 2026, prohibiting any person from offering therapy or psychotherapy services via AI unless provided by a licensed professional. The bill directly addresses AI systems marketed as mental health tools to the public. Separately, Louisiana's HB 475 advanced from committee, requiring health care providers to verbally notify patients before using AI to transcribe clinical visits. California advanced three health AI bills in the same week: SB 1146 (AI in healthcare advertising), AB 2575 (AI use in health care broadly), and AB 1979 (health care services AI). Minnesota's HF 3893 and Missouri's SB 1444 are both moving through committees targeting AI in mental health settings.
What this means for you
The state-level action is no longer theoretical. Maine has law on the books. Six other states have bills in active motion this month. If you are using AI tools in any client-facing capacity — documentation, intake, communication, or recommendations — patient disclosure is becoming a legal expectation, not a best practice.
Signal 3
81% of Physicians Now Use AI in Practice — And Their Top Fear Is Practitioner Skill Loss [WITNESS]
The AMA released survey data on March 12, 2026, showing that 81% of physicians now use AI in their practices — more than double the rate from 2023 (38%). Average use cases per physician rose from 1.1 to 2.3 in the same period. The top concern cited by 88% of respondents: potential skill degradation among newer practitioners who rely on AI before developing foundational clinical reasoning. Data privacy followed at 86%.
What this means for you
The mainstream clinical world has normalized AI use while simultaneously raising a governance alarm about what that use is doing to the practitioners themselves. The question is no longer whether to use AI — it is whether you have a clear understanding of where your judgment ends and the tool's output begins.
Signal 4
Washington and Utah Pass Laws Restricting AI in Insurance Prior Authorization [INSTRUCT]
Washington's SB 5395, signed into law in April 2026, increases restrictions on health insurers using AI in prior authorization decisions. Utah's SB 319 amends state health insurance preauthorization law with new AI disclosure requirements, mandating that insurers inform practitioners and patients when AI is involved in coverage determinations. These follow Tennessee's SB 1580, which bars AI systems from presenting themselves as mental health professionals. Nineteen AI bills passed into law across states in April 2026 according to tracking by Plural Policy.
What this means for you
If your clients navigate insurance coverage for services, AI is now making — or informing — those coverage decisions in ways that insurers may not be disclosing. You are on the downstream end of automated gatekeeping. Knowing whether an authorization denial was AI-generated is now a question you are legally entitled to ask in Washington and Utah, and likely more states by Q3.
Signal 5
HHS HTI-5 Proposed Rule Would Let AI Systems Access Your Patients' Health Records [INSTRUCT]
HHS and the Office of the National Coordinator for Health Information Technology (ONC) proposed the HTI-5 rule, now in a public comment window closing in late February 2026 with final rulemaking expected this year. The rule explicitly expands definitions of health data "access and use" to include autonomous AI systems, advancing a framework where AI can access electronic health information (EHI) automatically — not just when a human initiates a query. The rule also proposes removing over 50% of existing Health IT Certification criteria.
What this means for you
This is infrastructure-level change. If finalized, it creates a legal and technical pathway for AI systems — including those embedded in EHRs like Epic's ambient documentation tool and diagnostic copilots — to access patient records autonomously. Practitioners who are not asking their EHR vendor what is currently active on their account are making governance decisions without information.
The Pattern
Every signal this week points in the same direction: AI has crossed the threshold from pilot technology to ambient infrastructure, and the governance frameworks are lagging behind adoption. States are legislating because the federal framework has not arrived. Researchers are benchmarking failure modes because practitioners are already deploying tools that were never clinical-grade. Insurers are using AI in authorization decisions that practitioners did not design and often cannot audit. This is not a technology story. It is a documentation and accountability story. The practitioners who will be protected — legally and professionally — are the ones who can describe exactly where AI touched their work, what it was told, and what decision a human made independently. That is the audit trail. It does not exist yet in most practices.
One Thing You Can Do This Week
Open the last five client-facing documents your practice produced — notes, intake summaries, recommendations, or communications. For each one, answer this question: was AI involved in any part of drafting, summarizing, or informing the content? If yes, is that documented anywhere? That exercise takes 15 minutes. It is also the exercise a state licensing board would conduct if a complaint were filed. Do it before they do. - Signal 1: Mass General Brigham press release (April 13, 2026) + JAMA Network Open "Large Language Model Performance and Clinical Reasoning Tasks" — https://www.massgeneralbrigham.org/en/about/newsroom/press-releases/ai-chatbot-lacks-clinical-reasoning - Signal 2: Troutman Privacy "Proposed State AI Law Update: April 13, 2026" — https://www.troutmanprivacy.com/2026/04/proposed-state-ai-law-update-april-13-2026/ - Signal 3: AMA press release, March 12, 2026 — https://www.ama-assn.org/press-center/ama-press-releases/ama-ai-usage-among-doctors-doubles-confidence-technology-grows - Signal 4: Plural Policy "AI Governance Watch: Nineteen New AI Bills Passed Into Law" — https://pluralpolicy.com/blog/the-ai-governance-watch-april-2026-nineteen-new-ai-bills-passed-into-law/ - Signal 5: HHS.gov HTI-5 proposed rule — https://www.hhs.gov/press-room/hhs-proposes-hti-5-rule.html + Nelson Mullins analysis — https://www.nelsonmullins.com/insights/blogs/ai-task-force/ai/hhs-proposes-rule-to-deregulate-health-it-and-advance-ai-interoperability-hti-5