What changes for medicine and healthcare with AI?

The shift

Expert-level synthesis of a patient's full history, labs, imaging, and the medical literature goes from scarce specialist time to abundant, fast, and near-free. Pattern-matching against symptoms, images, and genomic data at population scale — the thing that used to require years of training and a packed schedule — is now something a model does in seconds, for anyone with a phone.

The axioms

• Diagnostic reasoning is gated by scarce, expensive expert cognition (years of training, limited hours in a day).
• Access to a doctor's attention is the bottleneck on getting an answer at all.
• Medical knowledge is asymmetric — the physician knows more than the patient, and that gap justifies the visit.
• A license and a body of legal accountability must sit behind every diagnosis and treatment decision.
• Treatment requires physical action on a physical body — surgery, an exam, a needle, a hand.
• Trust in a clinician is built through relationship and repeated contact over time.
• Health systems ration care by rationing clinician time, because clinician time is the scarce input.
• Records, coding, and admin exist to compress messy patient reality into a format scarce human reviewers can process.

Invalid axioms

1. Diagnostic reasoning is gated by scarce expert cognition. A model can synthesize a patient's history against the entire medical literature instantly, for free, at any hour. The habit-trap: systems still price and staff "getting a qualified opinion" as if it required booking scarce specialist minutes — triage queues, referral waits, and consult fees built around the assumption that synthesis is expensive.
2. Medical knowledge is asymmetric in the physician's favor. Patients can now get a plausible, literature-grounded explanation of their own condition before they ever see a clinician. The habit-trap: clinical encounters still run on the assumption that the doctor must first explain what's wrong, when many patients arrive already informed — visit structure hasn't adapted to a pre-briefed patient.
3. Records and coding exist to compress reality for scarce human reviewers. Summarizing a chart, drafting notes, and translating jargon into plain language is now free and instant. The habit-trap: clinicians still spend hours on documentation as if writing it up were the expensive part, when the actual bottleneck has moved to verifying and acting on what's written.
4. Second opinions are rare because expert time is scarce. A second (and third, and tenth) pass over a case is now nearly free to generate. The habit-trap: second opinions are still gated as a special, expensive event instead of a routine cross-check run by default.

Unchanged axioms

1. A licensed, accountable human must own the diagnosis and the treatment decision. Liability doesn't transfer to a model — when a call is wrong, someone with a license and a name has to answer for it. This is a legal and institutional fact, not a capability gap, so it doesn't move even as models get better.
2. Treatment requires physical action in the physical world. Surgery, physical exams, administering drugs, inserting lines — none of this is a token-generation problem. Robotic assistance narrows the gap in specific procedures, but the scarce resource is still trained hands and physical presence, not synthesis.
3. Being confidently wrong is unusually dangerous here. Medicine is one of the few domains where a plausible-sounding wrong answer causes direct physical harm rather than a redo. The cost of a hallucinated drug interaction or a missed red flag is not symmetric with the cost of a bad email draft — this raises the bar on verification rather than lowering it.
4. Trust and judgment under novel, high-stakes ambiguity stay human. Rare presentations, conflicting test results, and patients who don't fit the pattern are exactly where pattern-matching against "everything ever written down" breaks down, because there's no clean pattern to match. Judgment calls under real uncertainty, with a real person's stakes attached, still need a human who can be held to account for the call.
5. The therapeutic relationship — being believed, being physically comforted, having someone invested in your outcome — doesn't digitize. Adherence, disclosure of sensitive symptoms, and end-of-life conversations run on trust built over time with an accountable person, not on the quality of an explanation.

New axioms

1. When a plausible diagnosis is free and instant, who verifies it at the volume patients now generate them? Patients arriving with AI-generated self-diagnoses, drug interaction checks, and symptom triage shift the clinician's job from "produce the first answer" to "audit someone else's answer" — at a volume no current staffing model accounts for.
2. When synthesis is abundant, what happens to the pipeline that trained experts by making them do the synthesis themselves? Junior clinicians historically built judgment by doing the slow, effortful pattern-matching manually. If that step gets skipped because a model does it faster, the training pathway for judgment on novel cases is untested at scale.
3. When anyone can generate a confident-sounding medical claim, how does a health system distinguish a well-grounded AI-assisted opinion from a plausible-but-wrong one, especially across languages and health-literacy levels? The asymmetry that used to protect patients (they couldn't generate false confidence on their own) is gone; nothing yet replaces it.
4. Who is accountable when a clinician's decision was substantially shaped by an AI system's output? Liability law and malpractice frameworks were built for a world where the diagnosing mind and the accountable party were the same person. Shared or AI-assisted reasoning breaks that assumption before institutions have caught up.
5. What happens to care access and equity when synthesis is free but the physical exam, the procedure, and the follow-up remain scarce and unevenly distributed? Abundant diagnostic information without abundant treatment capacity can widen the gap between "knowing what's wrong" and "getting it fixed," particularly where physical access to care is already the constraint.

Where it breaks

"Diagnostic reasoning is gated by scarce expert cognition" (invalid) collides directly with "a licensed, accountable human must own the diagnosis" (still holds): once synthesis is free, the physician's remaining job is to verify and take responsibility for an AI-generated read — but staffing, scheduling, and reimbursement still treat the visit as if the physician is the one doing the original diagnostic work from scratch. Nobody has resourced the verification-and-accountability step as its own scarce activity; it's still bundled into the old "expert time" line item, which is exactly the wrong unit to ration by now.

A second collision: "second opinions are rare because expert time is scarce" (invalid) meets "who verifies a plausible diagnosis at scale" (new). Once cross-checking is nearly free, the reasonable move is to run it by default on every case — but no system has decided who is accountable for reconciling an AI-generated second opinion that disagrees with the treating clinician, or what happens legally when it does.

ai-axioms.vercel.app/axioms/healthcare/medicine-and-healthcare-ai-axioms