CareChronicle

Issue №03 · Week of June 22, 2026

Hospitals are rushing AI into everything from revenue cycle and inbox triage to clinical detection and diagnosis — just as public sentiment toward AI is moving the other direction. The operational argument for automation keeps getting stronger, but hallucinations and medical errors remain unsolved, and “human in the loop” is starting to seem less like a safety model than a liability rubber stamp.

This week: Anthropic’s Fable 5 remains under government scrutiny, hospitals create AI literacy roles while agents confidently fabricate incorrect answers, and Midjourney decides whole-body ultrasound screening should be a spa business.

CONTINUING STORY

Anthropic's Most Powerful AI Lasted Three Days

Last week we reported on the chronology of the U.S. government's takedown of Anthropic's latest Mythos-class models — Mythos 5 and its sibling, Fable 5. Additional details have emerged.

The Economist's initial reporting kicked off a viral claim that Mythos broke into "almost all" classified systems belonging to the NSA. The original author has since walked it back: the alleged breach happened under approved internal red-team testing. Anthropic contends that the models only flagged minor, known bugs.

Trump met Anthropic's CEO, Dario Amodei, at the G7 Summit. He told Axios he had viewed Anthropic as a national security threat "a week ago, maybe", but said talks are ongoing and he doesn't want to shut the company down — while keeping Defense Production Act powers on the table. Wired reports that co-founder Tom Brown has replaced Amodei — whom a White House official reportedly called "a weirdo" — in discussions with the administration.

OpenAI has evaded similar criticism, despite shipping a model that may outperform Mythos. GPT-5.5-Cyber, announced this week, hit 85.6% on CyberGym — edging out Mythos 5's 83.8%; GPT-5.5-Cyber was gated to vetted defenders and cleared with government reviewers before release.

Previously, OpenAI president Greg Brockman gave $25 million to pro-Trump super PAC MAGA Inc. in September, reportedly the cycle's largest single gift, and the company has spent more than a year in talks with the White House discussing a potential equity stake.

More than 100 security experts have already signed a letter arguing the Anthropic ban hurts defenders more than attackers.

OPINION

Healthcare has an AI problem

Justin Flechsig

Hospitals are adopting AI at breakneck speed and deploying it faster than anyone can prove it works.

The appeal is obvious: AI promises a silver bullet for everything crushing health systems — negative margins, relentless documentation, too few clinicians, too many messages, too many workflows held together by human effort. More than 80% of physicians now use AI professionally, and writing in The Atlantic, Ben Mazer calls this medicine's "Uber moment": in April, a Harvard–Stanford study pitted ChatGPT against hundreds of doctors on real cases, and the bot won. But the study's own lead author flinched at the result, calling it an academic exercise — not a green light to start replacing clinicians. Mazer, a pathologist, says he's lost count of the AI tools his own hospital keeps sending to him.

He's right to worry. ECRI, the patient-safety nonprofit, ranked the misuse of AI chatbots the #1 health-technology hazard of 2026 — above every device and drug-delivery risk on its list. Mass General Brigham's BRIDGE benchmark found the strongest AI model scored 92 on standardized exams but just 44.8% on real-world clinical text. And OpenEvidence — now in thousands of hospitals — handed out flatly wrong dosage guidance, as we covered last week.

Hospitals have never lacked decision support. Clinicians already drown in alerts, flags, and reminders they've learned to tune out — a Northwestern cardiologist notes a warning can be fully available and still change nothing. AI may just be the newest alert no one acts on. Nor does "a human in the loop" necessarily save you: the same Atlantic piece describes a randomized trial in which a single wrong AI suggestion pulled doctors off course. If containing hallucinations is hard for the frontier labs that build these models, expecting a community hospital's IT team to manage it is fantasy — and only 8% of physicians say their organization's AI rules are even clear. Adoption is lapping governance.

Meanwhile, public sentiment on AI has soured — only 16% of Americans expect AI to benefit society — while fewer than half of Americans can even afford healthcare.

That should lead every responsible healthcare executive to ask themselves: when the tool gets a dose wrong, who answers for it? And what's the plan for the day patients start wondering whether they're talking to a person at all? AI is one of the most powerful tools for patient access hospitals have ever had; deploying it deserves consideration commensurate with its power for good and harm.

Policy & Regulation

AI is at the center of US policy discussions

• Global rules: Trump met with AI CEOs at the G7 to discuss U.S.-led global AI standards and a proposed safety forum.

• White House shuffle: The Trump administration's AI policy team is reorganizing as David Sacks and Sriram Krishnan step back, with Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and Chief of Staff Susie Wiles absorbing influence.

• Clinical trials: HHS launched Operation TrailBlazer, directing six agencies — including the FDA, NIH, ONC, NCI, and OIG — to cut early clinical trial timelines by six to twelve months and explore EHR-based trial notifications.

• AI dividend: Sen. Bernie Sanders introduced legislation that would pay Americans $1,000 annually on the premise that AI is a public resource built on collective human knowledge.

• Cyber warning: The Five Eyes alliance warned that emergent AI models pose an urgent global cyber risk by enabling faster and more sophisticated cyberattacks.

Hospitals & Care Delivery

Where are hospitals actually deploying AI?

We’re seeing less sci-fi replacement than workflow pressure relief: hospitals are reaching first for scribing, payment, documentation, and simple automations. With that said, the latest Y Combinator cohort shows evidence of investment into creating the “AI physician”.

Revenue cycle / denial management: Baptist Health says Epic’s AI coding assistant cut chart review time 20% and saved an estimated $924,000 a year. A Becker’s panel on hidden denials put the larger pressure in context: hospitals spent an estimated $43 billion in 2025 chasing payments insurers already owed, while DRG downgrades and short pays often never trigger a formal denial alert.

Ambient / documentation / inbox pressure: An Epic Cosmos study found portal messages rose 153% from 2020 to 2025, while phone calls fell only 6% and in-person visits rebounded. In Q1 2025 alone, roughly 42 million Epic patients messaged a clinician. Hospitals are reaching for tools that lighten the load, from ambient to inbox triage.

Operations and automation: Boston Children’s says more than 50 automations saved 30,000 hours in the first half of 2026, with staff using GPT tools saving an estimated two to three hours per week. A nine-hospital virtual nursing study found 30-day ED returns were 3.7% after virtual-nurse discharge, versus 13.3% after standard discharge.

Clinical detection and care gaps: Forrest Health says an Epic-embedded tool flagged 173 incidental lung nodules in six weeks, with a human navigator reviewing each case before routing. Boston Children’s and OpenAI reanalyzed 376 previously unresolved rare-disease cases and confirmed 18 diagnoses, a 4.8% additional diagnostic yield after prior specialist review.

AI and Healthcare Trends

AI governance moves to the org chart

Weekly roll-up: Lumeris added symptom checking to its AI platform, DeepIntent pushed further into pharma targeting, and WHOOP partnered with HealthEx on wearable data consent. AI is becoming background plumbing across payers, pharma, devices, and care navigation.

Cyber release: OpenAI launched Daybreak, including a GPT-5.5-Cyber model it says beat GPT-5.5 on CyberGym, plus Codex Security tooling that has scanned more than 30 million commits. Cyber reportedly outperformed Claude’s Fable 5, which remains under government scrutiny.

AI literacy: Mass General Brigham named Chenzhe Cao director of AI literacy, adding to its AI governance and product bench. Hospitals are creating bespoke roles for rollout — another sign healthcare AI adoption is escalating from pilot work to institutional infrastructure.

Agents drift: Aultman Health’s CIO described an internal AI agent that gave confident fabricated answers instead of admitting uncertainty. UCHealth and Jefferson are responding with retrieval, tuning, and governance. Health systems are figuring this out in real time.

Ultrasound spa: Midjourney plans to spend $74 million with Butterfly Network to launch a whole-body ultrasound screening business, starting in the Bay Area. Radiologists say the marketing is ahead of the evidence: no clinical validation data have been published, and performance claims remain contested.

Governance split: Amazon’s Eric Brandwine argues human-in-the-loop governance is the wrong model because reviewers normalize alerts and stop catching failures. That lands differently in a healthcare week full of hallucinations, medical errors, and agents going off script.

Research, Evidence & Benchmarks

• Clinical drafts still carry harm: UCSF’s hepatology model produced appropriate case-specific recommendations most of the time, but reviewers found misleading content in 10% of drafts and severe-harm risk in 3.4%.

• Consumer AI needs rules: A JMIR viewpoint argues direct-to-consumer health LLMs create governance gaps around privacy, safety, accuracy, and equity that generic AI frameworks do not cover. Its proposed risk tiers are anticipatory, but the core warning is already live: patients are using medical chatbots before medicine has rules for them.

• Mental health is fragile terrain: A review found GenAI chatbots can reduce anxiety and depressive symptoms short-term, but flagged dependency, sycophancy, maladaptive belief reinforcement, and unclear liability. In mental health, the clinical risk is not only being wrong; it is being agreeable at the wrong moment.

• Equity cuts both ways: A JMIR commentary argues ChatGPT Health could worsen disparities by legitimizing self-rationing, misfiring on emergency triage, or standing in for care patients cannot access.

• Trust is conditional: In a preregistered experiment with 1,502 U.S. adults, human nurse advice was rated more credible than AI advice. But message framing sometimes mattered as much as source identity, and people skeptical of traditional medicine rated the AI nurse as more competent.

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