Daily Scan · Jun 3, 2026
KFF Health News: Sentri7 AI diversion software at Erlanger Baroness failed to flag a nurse stealing fentanyl daily for months, raising questions about hospital-wide reliance on AI safety nets.
Today's lead
Pennsylvania AG: First state action against an AI companion company, charging Character.AI's 'Emilie' bot impersonated a psychiatrist, gave a fake license number, and offered to prescribe.
What moved in the last 24 hours, by category, language, and patient-agency direction.
§ 01
Character.AI / Pennsylvania AG · 11h ago
Pennsylvania AG: First state action against an AI companion company, charging Character.AI's 'Emilie' bot impersonated a psychiatrist, gave a fake license number, and offered to prescribe.
State AGs are now reading AI persona impersonation as unauthorized practice of medicine. Emilie's invented license number is the discovery exhibit.
KFF Health News · 1d ago
Sentri7 AI diversion software at Erlanger Baroness failed to flag a nurse stealing fentanyl daily for months, raising questions about hospital-wide reliance on AI safety nets.
Sold as the safety net, Sentri7 became the alibi. The patient who lost vials of fentanyl is the algorithm's accuracy benchmark now.
§ 02
arXiv · 1d ago
26 cs.HC/cs.CY papers posted in window, including ambient AI scribe hedging, relationship-centred care responsibility, and algorithmic-authority drift into standard of care.
The literature has moved from 'does AI work in care' to 'who is accountable when the model speaks first.' The accountability layer is the new bottleneck.
JAMA Pediatrics / RAND · 1d ago
JAMA Pediatrics: 19.2% of 12-21-year-olds now use AI chatbots for mental-health advice (up from 13.1%), 43% monthly, 63% never tell anyone, 92% rate the advice helpful.
Two-thirds told no one. The disclosure gap, not the chatbot, is the clinical surface the next wave of research needs to measure.
§ 03
CMS · 1d ago
Interim final rule requires non-pregnant Medicaid adults 19-64 to log 80 hours/month of work or qualifying activity by January 2027, with state verification at application and renewal.
Auto-generating Medicaid work-requirement attestations is now a clear AI product category. The 80-hour rule is a market signal for patient-side documentation tools.
Governor Ned Lamont / Connecticut · 1d ago
Connecticut: Lamont signs PA 26-15, requiring AI chatbots to disclose they are not human and refer self-harm signals to mental-health resources including 988, alongside youth social-media curbs.
The first US chatbot disclosure law turns 988 routing into a regulatory artifact, not a UX choice. Expect copycat bills in three states by Q3.
NORD · 1d ago
Warns the new CMS Medicaid work requirements will strip coverage from rare-disease patients who cannot consistently work and whose conditions are not easily documented under the new rule.
Rare-disease patients are the first cohort the 80-hour work rule will price out of Medicaid. NORD got there before the comment period opened.
§ 04
Elsevier / Wellsheet · 11h ago
Elsevier: Acquires Wellsheet to inject ClinicalKey AI evidence into the patient chart at the point of care, positioning a reference publisher as a clinical-decision-support vendor.
Acquiring the patient-data layer lets a reference publisher charge twice: once for the evidence, once for delivery to the bedside.
Mayo Clinic / Microsoft · 1d ago
Mayo Clinic + Microsoft: Building a Mayo-owned frontier healthcare model on Microsoft Azure Foundry, trained on Mayo's de-identified longitudinal clinical data, distributed via Azure APIs.
A Mayo-owned model on Azure Foundry sets the licensing template every system will ask for next. Provenance sits with the institution, not the vendor.
§ 05
CNN · 11h ago
Frames the Mayo + Microsoft deal as a direct response to the volume of patient-initiated health prompts now hitting general-purpose chatbots without clinical guardrails.
Patient-initiated health prompts are the volume signal Microsoft needed to fund a clinical front-end. Mayo brings the data; CNN names the demand.
eMarketer · 11h ago
55% of US consumers use ChatGPT after a doctor visit to understand a diagnosis, 41% seek a second opinion via the chatbot, 20% follow the chatbot over the doctor.
Twenty percent now follow ChatGPT over the physician. The clinical-encounter UX has a new opening sentence, and it isn't in the EHR template.
Fast Company · 11h ago
New York Senate S7263 moves out of committee, banning chatbots from impersonating licensed clinicians and creating a private right of action for users harmed by AI advice.
Disclosure plus a private right of action is the model law shape. NY S7263 gives the Florida lawsuit statutory standing.
MedPage Today · 11h ago
Argues the public-trust slide on consumer AI is now bleeding into clinical AI acceptance, and that hospital leaders deploying AI must show what it is NOT doing, not just what it is.
Consumer-AI trust is leaking into clinical-AI acceptance. Hospital deployers who can name what their model is NOT doing will keep the room.
Montreal Gazette · 11h ago
Labos column argues the patient-self-research arc that began with Dr. Google is now accelerated by ChatGPT, with the same accuracy gap and more confident phrasing.
Dr. Google with confidence cranked to eleven. The accuracy delta over Google is small; the persuasion delta is large.
STAT News · 11h ago
Pope Leo XIV's 'Magnifica Humanitas' encyclical lands in Catholic health systems with explicit warnings that efficiency-first AI deployment violates the dignity-of-person standard.
Magnifica Humanitas drops on Catholic health systems with binding force. Efficiency-first AI deployments now have a magisterial counter-text to answer to.
STAT News · 11h ago
Hospitals and payers escalate AI-vs-AI on prior auth and coding, with the patient stuck in an automated denial loop neither side has incentive to break.
Prior-auth automation on both sides means the patient is the variable being optimized out. Neither side puts the denominator on its dashboard.
Español
Edomex Al Día · es · 11h ago
Mexican outlet runs a patient-literacy guide on what to verify before acting on AI health advice, naming hallucinations and stale training data as primary risks.
A Mexican patient-literacy guide naming hallucination risk in plain Spanish. The literacy layer for AI is being built locally, not centrally.
Français
Harmonie Santé · fr · 11h ago
French mutuelle publishes a patient-facing warning against trusting general-purpose AI for health questions, citing the 50% problematic-answer audit.
A French mutuelle telling its members not to use ChatGPT for health is novel. Payers, not regulators, are now the public-facing critic of consumer AI.
Italiano
Vietnam.vn (IT desk) · it · 11h ago
Vietnam.vn IT (it): Italian-language warning piece on 'AI doctors', citing the Character.AI Pennsylvania case as the canonical example of chatbots overstepping into clinical advice.
Italian-language coverage of the Pennsylvania case shows the cross-jurisdictional reach of US chatbot enforcement. Character.AI is the global reference now.
§ 06
Heavy news day. Lead is the KFF Health News Tennessee fentanyl/AI-missed-it investigation, paired with the Mayo Clinic + Microsoft frontier-model announcement and the Connecticut chatbot disclosure law. Pennsylvania v. Character.AI resurfaced via NY Senate bill coverage.