Kennedy Says AI Can Beat Doctors On Medical Choices

Have you ever walked out of a clinic with more questions than answers, holding a stack of papers nobody had time to read with you? That feeling is not rare. Appointments get shorter. Charts get longer. And now a senior health official is telling crowds that artificial intelligence might do a better job of reading those charts than the person in the white coat. That claim is bold. It is also messy. I sat with it for a while because it touches something people already feel in their bones: the system is rushed, and rushed systems miss things.

Why This Argument Hit A Nerve

On a late September stage in Washington, Health Secretary Robert F. Kennedy Jr. told an audience that he had spoken with a leading AI executive. The executive, he said, argued that it would already be malpractice for a clinician to diagnose or prescribe without at least checking a model. Kennedy then pushed the point further. Americans should get their records. Those records can run a thousand pages. A visit can last six minutes. A person cannot scan that file in six minutes. A model can. It can distill. It can, in his telling, offer a second opinion that is better informed than any physician in the country.

That last line is the one that travels. It is also the line that makes physicians restless. I’ve found that whenever someone pits a tool against a profession in public, the debate stops being about software and starts being about trust. Who do you believe when the stakes are your body? A person who can look you in the eye, or a system that never gets tired and never admits it is guessing?

The setting matters. The summit was backed in part by AI firms. An executive from one of those firms spoke on a session about building a health stack. Meanwhile, the White House has been urging faster American adoption so rivals do not pull ahead. Voluntary pledges with company leaders landed the same week. If you squint, the policy story and the product story start to look like cousins.

You may have a medical record that’s a thousand pages long. You have six minutes with a doctor today. He’s not going to be able to review it, but the AI can.

That quote is sticky because it is partly true. Records are bloated. Portals are ugly. Copy-paste notes pile up. Prior authorizations eat hours. Nobody I know thinks the average visit is long enough. The leap is the next sentence: that a model’s summary is not just faster, but wiser than a trained clinician. Speed is not the same as judgment. Distillation is not the same as care.

The Promise People Actually Want

Strip away the politics for a minute. What do patients want? They want someone, or something, to notice the lab that never got followed. They want a plain-language map of a confusing history. They want a second set of eyes when a symptom does not match the first story. In my experience, that is the human demand hiding under the slogan.

AI can help with that demand if it is treated as a reader, not a ruler. It can cluster notes. It can flag interactions. It can draft questions you should ask before you leave the room. Used that way, it is a flashlight. Used the other way, it becomes a substitute parent, and substitute parents make sloppy decisions when the data is thin.

• Longitudinal records that a person cannot finish in one sitting
• Medication lists that conflict across clinics
• Imaging reports written in dense shorthand
• Family history that never made it into the current note
• Second-opinion drafts that a clinician can accept or reject

Those are useful jobs. They are also bounded jobs. The moment a model starts telling you not to seek care, or to ignore a red-flag symptom, the flashlight turns into a hazard. Lawsuits already allege that chat tools steered people away from emergency rooms. One pending case involves a pastor who said a chatbot told him not to get help when severe dizziness hit. Courts will sort the facts. The pattern is enough to slow the applause.

What “Better Informed” Really Means

Kennedy’s phrase was better informed than any doctor in the country. Informed by what? By the same literature the profession already uses? By web text that includes both careful reviews and junk blogs? By training data that reflects yesterday’s consensus, including the parts later walked back?

Critics jumped on that point. One physician argued that models often replay the dominant narrative from agencies and media, then hallucinate when they lose the thread. Another ran the cabinet remark through a chatbot and got a cautious reply: if the claim is that AI is more reliable than physicians for medical advice, the evidence does not establish that. Some real-world signals point the other way.

I keep coming back to a simple test. If a model summarizes a thousand-page chart, who checks the summary? If nobody checks, you have automation theater. If a clinician checks, you have a workflow change, not a replacement. Replacement talk sells tickets. Workflow talk keeps people alive.

Six Minutes Versus A Thousand Pages

The six-minute visit is not a myth. It is a design failure. Insurers price time. Health systems schedule density. Electronic records reward clicking. Patients arrive with printouts. The room cannot hold all of that. So yes, a model that reads overnight and greets you with a one-page brief sounds like mercy.

But mercy has conditions. A brief can hide uncertainty. A brief can flatten a rare presentation into a common one. A brief can sound confident when the source notes were hedged. Confidence is a style. Medicine is a probability. Those two things fight.

Perhaps the most interesting aspect is not the software. It is access. Kennedy also repeated a long-running goal: people should control their own records. That part is older than this summit. Portals exist. Downloads exist. They are still clumsy. If AI becomes the layer that makes a personal archive usable, that is a public-interest win even if you dislike the rest of the speech.

Look at that table and the hype shrinks. The useful work is clerical and comparative. The hard work is still human: values, fear, family context, the way a person describes pain when they are trying not to cry. Models do not sit in that room. They sit on a server.

Industry Rooms And Public Trust

It is fair to notice who paid for the microphones. When companies that sell models sponsor a health summit, the audience hears a sales pitch even if the speaker is a cabinet official. That does not make every sentence false. It does mean the incentives are not hidden. Firms want adoption. Officials want national lead. Patients want not to be harmed. Those aims overlap. They are not identical.

I’ve watched enough product launches to know the pattern. First comes the demo that feels like magic. Then come the edge cases. Then come the terms of service. Then come the lawyers. Health is already deep into that sequence. Pending suits over catastrophic advice are not a footnote. They are a preview of the next decade.

Kennedy did mention downsides and asked the vice president to expand. The reply focused on models trained to attack computers, and on a policy instinct: do not shield firms from product liability with clever rules. That instinct will please people who think regulation often becomes a hall pass. It will worry people who think liability without standards just produces silence and fine print.

Most of the downsides are all related to models that were literally trained to hack into your computer.

– Vice President JD Vance, as relayed at the summit

Security risk is real. So is clinical risk. They are not the same animal. A model that phishes your laptop is a cyber problem. A model that delays a stroke workup is a body problem. Policy that only names one animal will miss the other.

The Malpractice Line And Why It Stings

Malpractice is a legal word with teeth. When a public official repeats an executive’s claim that skipping an AI check is already malpractice, the profession hears a threat. Insurers hear a new standard of care being drafted on a stage instead of in a court. Residents hear that their training is being marked incomplete by a chatbot.

Standards of care usually move slowly. They move through case law, boards, and societies. They do not usually arrive as a punchline at a sponsored summit. That is why the sentence landed like a brick. Even people who like decision support do not want the support tool declared mandatory by branding.

There is a quieter version of the same idea that I can live with. Decision support should be available. It should be logged. It should be easy to override with a reason. It should not pretend to be a license. That version respects both the six-minute clock and the license on the wall.

Tyranny Talk And Expert Fatigue

Kennedy framed AI as a way to escape what he called medical tyranny. He has long attacked public-health guidance from the pandemic years on masks, shots, and related rules. In that frame, models are not just summarizers. They are dissidents that can contradict official lines.

Here is the catch. Models are trained on the corpus that exists. That corpus includes official guidance, journal articles, news, forums, and noise. If your complaint is capture by institutions, you should not assume the model is uncaptured. It often is a polished mirror of the same sources, plus a habit of inventing citations when it gets lost.

Expert fatigue is real. People were told firm things that later softened. Trust cracked. A tool that says “maybe not” can feel like relief. Relief is not proof. A dissenting paragraph from a chatbot is still a paragraph from a chatbot. You still need methods, not vibes.

1. Ask what dataset the answer leans on.
2. Ask what it cannot see, such as the exam you have not had.
3. Ask a licensed clinician to contest the draft.
4. Keep the original notes, not only the summary.
5. Treat certainty language as a style choice, not a lab result.

That list is not anti-tech. It is anti-magic. Magic is how people get hurt.

Records, Ownership, And The Unsexy Middle

Give people their files. That sentence should not be controversial. Interoperability has been promised for years. It still fails in small, stupid ways: mismatched identifiers, PDFs that cannot be parsed, imaging that lives on a disc in a drawer. AI does not fix those pipes by existing. It rides on them.

If the administration wants a legacy here, the unsexy middle is the prize. Clean exports. Patient-controlled permissions. Audit logs when a model reads a chart. Clear rules for what a vendor can keep. Without that plumbing, “AI second opinion” is a slogan printed on a leaky bucket.

I would rather see a boring portal that works than a dazzling demo that cannot name its sources. Boring scales. Dazzle trends.

Where The Law Is Already Walking

Product liability is the vice president’s preferred brake. Fair enough. If a tool gives advice that a reasonable user treats as care, courts will ask whether the tool is a product, a service, or a speech engine hiding behind a disclaimer. Those categories matter. Disclaimers that say “not medical advice” sit next to interfaces that sure look like medical advice. Juries notice mismatches.

Clinicians have their own liability. If a hospital deploys a summarizer and a doctor relies on a missed negation in a note, who owns the miss? The vendor? The health system? The attending who clicked accept? Shared systems create shared blame. Shared blame without contracts is chaos.

A practical stack, not a slogan:
  1. Patient-held record
  2. Machine summary with citations back to notes
  3. Clinician review and signed plan
  4. Logged overrides
  5. Clear vendor liability for known failure modes

That stack will not fit on a rally sign. It will fit in a hospital. I will take the hospital version.

What Doctors Hear That Officials May Not

Talk to working clinicians and you hear a different soundtrack. They already use suggestion tools for billing codes and draft notes. Some of those tools save time. Some generate fluent nonsense that takes longer to edit than writing from scratch. Hallucinated meds. Wrong laterality. Softened allergies. The failure mode is not that the software is dumb. The failure mode is that it is fluent.

Fluency is dangerous in a tired clinic. A tired person accepts a clean paragraph. A clean paragraph can be wrong in one fatal adjective. That is why “check the AI” cannot mean “trust the AI.” Checking is labor. Labor needs time. Time is what the six-minute visit does not have. You see the loop.

So if leaders want models in the room, they should buy time, not only licenses. Add minutes. Cut pajama-time charting. Pay for review. Otherwise the tool becomes another tab that nobody truly reads.

Patients Will Use This Anyway

Here is the part officials did not need to announce. People already paste symptoms into chat windows at midnight. They already upload lab PDFs. They already ask a model to translate a discharge summary into English a grandparent can follow. That behavior will not stop because a medical society frowns. It will grow.

The adult response is not a ban and not a coronation. It is literacy. Teach people that a model can miss a surgical history. Teach them that rare disease lists are long because rare things are many. Teach them that “see a doctor” is not a dodge when chest pain is on the table. Teach clinicians to ask, without scorn, what the patient already asked a bot. Shame makes secrets. Secrets make surprises.

I’ve found that the best visits now include a third artifact on the desk: the printout from last night’s chat. Treat it as a letter from a well-meaning cousin. Some cousins are sharp. Some cousins invent family lore. You do not hand them the prescription pad.

National Competition And Clinic Reality

The administration’s broader pitch is familiar. Stay ahead of strategic rivals. Do not let another country own the stack. Sign voluntary accords. Talk about leadership. That geopolitical frame is not fake. Compute, talent, and data are contested. Health data is a prize because it is intimate and because it trains the next model.

Clinic reality does not care who wins a race if the local discharge instruction is wrong. A family in a small town does not experience national leadership. They experience a portal that times out. Policy that only speaks in races will overbuild the showcase and underbuild the waiting room.

Balance is possible. Fund evaluation the way we fund pilots. Publish failure rates the way we publish speed claims. Make “does this reduce missed follow-ups” a metric that matters more than “we launched a stack.” Metrics change culture. Slogans do not.

A Fair Read Of The Upside

Let me steel-man the speech. Chronic patients drown in fragments. Specialists do not talk. Primary care is a traffic cop with no radio. A model that stitches fragments into a timeline could catch the silent creatinine creep, the repeated ER visits with the same belly pain, the antidepressant that never got a trial at a real dose. Those catches save lives. They also save money, which is why vendors smile.

Rural care is thin. A well-governed assistant could extend a scarce specialist’s reach if the specialist remains in the loop. Language access could improve if summaries land in the patient’s preferred tongue with a human check. None of that requires declaring doctors obsolete. It requires declaring paperwork obsolete, which should have happened already.

The useful future is not AI over doctors. It is AI under doctors and beside patients, with receipts.

Receipts meaning citations. Meaning the sentence on screen points back to the note, the lab, the image. No citation, no trust. That rule is older than neural nets. We should not retire it because the prose sounds warm.

The Cultural Fight Under The Policy Fight

This story is not only about software. It is about who gets to be an authority after a bruising decade. Some audiences want institutions humbled. Some want institutions restored. AI is being drafted onto both teams. That is a lot of weight for a next-token predictor.

When a speaker says a model can free the public from expert overreach, part of the crowd hears justice. Another part hears a door opening onto amateur hour. Both reactions are human. Neither should write the safety spec.

A healthier culture would let patients question without being mocked and let clinicians decline a model suggestion without being called Luddites. That culture needs manners more than it needs another keynote.

Practical Guardrails Worth Fighting For

If you work in a system that is about to buy this stuff, push for dull rules. Require source spans. Require uncertainty labels. Ban silent training on identifiable notes unless a patient opts in. Test models on local dialects and on the ugly charts, not the demo charts. Measure disagreement between model and attending, then study the disagreements instead of hiding them.

• No autonomous “don’t go to the ER” language for acute symptoms
• Human sign-off before any change in a high-risk med
• Patient-visible logs of what the model saw
• Easy export that is not a 400-page unsearchable dump
• Vendor contracts that do not shrug at known hallucination classes

None of those bullets require you to pick a tribe. They require you to admit that bodies are not beta tests.

What I Would Tell A Friend Tomorrow

If a friend asked whether to use a model on their chart, I would say yes, with a script. Upload what you legally can. Ask for a timeline. Ask for missing follow-ups. Ask for questions to bring to the visit. Do not ask it to cancel the visit. Do not let it talk you out of chest pressure, one-sided weakness, or a child who will not wake right. Bring the printout. Watch the clinician’s face. If they engage the printout, good. If they sneer, find a better room when you can.

If a friend is a clinician, I would say the opposite of panic. Learn the tools. Mark the errors in public inside your shop. Demand time to review. Refuse to be the legal shock absorber for a vendor’s confidence score. You went to school for judgment. Keep it.

If a friend writes policy, I would say stop announcing replacements. Announce evaluations. Publish where models help and where they harm. Tie federal buying to those results. Leadership is not a press line. Leadership is a scoreboard people can audit.

The Sentence That Should Have Been Said

Here is the version I wish had been on that stage. Records belong to patients. Visits are too short. Models can read faster than people. Faster is not wiser. Use the speed to give clinicians their eyes back. Keep a human on the hook for the plan. Do not call it malpractice to practice without a product. Call it malpractice to ignore a patient. Then fund the minutes that make attention possible.

That paragraph will not thrill a sponsor. It might thrill a parent sitting in a parking lot, trying to decide whether the rash can wait until morning. Those parents are the actual market. They do not need a war between algorithms and doctors. They need both to behave.

Will AI dominate medicine? It will inhabit medicine. Inhabit is the honest verb. Domination is a movie poster. The next five years will be won by the teams that treat summaries as drafts, liability as real, and six minutes as a scandal rather than a given. Everything else is noise around a microphone.

I do not know whether this week’s remarks will become a rule, a slogan, or a footnote. I do know patients will keep asking machines for help after midnight. The decent move is to meet them there with guardrails, not with a coronation and not with a scold. Read the chart. Read the model. Then look at the person. That order still works. It will keep working after the summit lights go dark.