What's the point of a GP visit now that patients arrive with an AI-generated differential diagnosis?

The shift

Turning a patient's symptoms and history into a ranked differential diagnosis goes from scarce — a trained clinician's minutes, gated by an appointment — to abundant, produced by the patient on their phone before they walk in. The synthesis step that the visit used to exist to deliver now happens pre-visit, for free, at any hour.

The axioms

• The GP is the gateway to the differential: reaching a plausible list of what-could-this-be requires access to scarce clinical cognition.
• The visit is where information gets delivered — the patient comes to be told what's likely wrong and what to do.
• A licensed human must own the accountable diagnosis, the prescription, and the referral; nothing downstream moves without that signature.
• Assessment and treatment require physical action on a body — palpation, auscultation, looking in an ear, a swab, a hand on an abdomen.
• Judgment on ambiguous, atypical, or red-flag presentations is scarce, because it rests on pattern exposure and the willingness to be answerable for a call under uncertainty.
• Trust and continuity are built through repeated contact with the same accountable person over time.

Invalid axioms

1. The GP is the gateway to the differential. Generating a literature-grounded, reasonably-ranked list of possibilities is now something the patient does before the appointment, not something the visit exists to supply. The habit-trap: appointments are still scoped, timed, and triaged as if the first job is to produce the differential from scratch — the 10-minute slot is built around delivering an answer the patient may already have on their phone.
2. The visit is where information gets delivered. The classic arc — patient describes symptoms, doctor explains what's likely and why — assumed the patient arrived uninformed. Many now arrive pre-briefed, sometimes better-read on their specific presentation than a generalist can be across all of them. The habit-trap: visit structure, patient-education leaflets, and "what the doctor tells you" scripting still treat information transfer as the core deliverable, when the patient's need has moved to having their own worked-up answer confirmed, corrected, or acted on.

Unchanged axioms

1. A licensed human must own the accountable diagnosis, the prescription, and the referral. The differential can be free and still change nothing on its own: the prescription pad, the referral letter, the sick note, and the legal answerability when a call is wrong all require a licensed name. This is an institutional and legal fact, not a capability gap, so it doesn't move as models improve — the authority to act, and the liability for acting, stay with the clinician.
2. Assessment and treatment require physical action on a body. A model can rank pneumonia against bronchitis; it can't listen to the chest, feel the lymph nodes, look at the eardrum, or take the blood. The physical exam frequently reorders or kills the AI differential precisely because it adds signal the patient couldn't feed the model. Ordering and interpreting tests sits in the same place — the clinician has the authority to order, and the exam-plus-test loop is where a plausible list meets ground truth.
3. Judgment on ambiguous and atypical presentations stays scarce. Confident-plausible is the model's default failure mode, and it is worst exactly where it matters most: the patient who doesn't fit the pattern, the vague presentation hiding a red flag, the symptom that reads benign but isn't. Deciding when a tidy AI answer is wrong, and being answerable for that decision, is not something the abundance of differentials touches. (Calibrated to mid-2026: models are getting better at flagging red flags and their own uncertainty, and this is the axiom most exposed to further capability gains — but the accountability for the call, per axiom 1, does not transfer even if the reasoning does.)
4. Trust and continuity are built through repeated contact with an accountable person. Disclosure of embarrassing symptoms, adherence to a plan, and knowing this patient's baseline over years run on a relationship, not on the quality of an explanation. A patient argues with a chatbot's differential and drops it; they are more likely to accept a correction from a clinician who has known them and is on the hook for the outcome.

New axioms

1. Patients now anchor on a confident-but-wrong AI differential, and the visit has to unwind it. Pre-work-up used to be absent; now it arrives with unearned confidence and a ranked list the patient has already emotionally committed to. Talking someone off a wrong self-diagnosis — especially a frightening one the model over-weighted, or a reassuring one that buried a red flag — is a new and harder conversation than starting from a blank slate, and nothing in the visit is scoped for it.
2. The visit risks collapsing into AI-adjudication without anyone resourcing that as the actual work. If the patient brings the differential, the clinician's job shifts toward auditing someone else's reasoning — agreeing, correcting, or overriding it, and owning the result. That is real cognitive labor, but it's invisible in a system that still measures and pays for the visit as if the doctor produced the answer. Who is accountable, and how the encounter is timed, for adjudicating a machine's output is unsettled.
3. Triage of the worried-well, now armed with AI, floods the front door. Free, always-on differential generation lowers the bar to being alarmed: more people arrive convinced they need to be seen, citing a plausible list. Sorting who genuinely needs the scarce physical exam and accountable call from who has been spooked by a confidently-worded model output is a volume problem no current triage model was built for — and getting it wrong in either direction has a physical cost.

Where it breaks

"The visit is where information gets delivered" (invalid) collides with "the visit is now AI-adjudication" (new): the 10-minute slot was designed for a clinician to gather, synthesize, and explain — but the actual work has become auditing a differential the patient brought, unwinding their anchor on it, and owning the corrected call. The encounter is being asked to do harder work in the same time built for the easier, now-obsolete task, and neither scheduling nor reimbursement has noticed the job changed.

A second collision: "the GP is the gateway to the differential" (invalid) meets "triage of the AI-armed worried-well" (new). Once anyone can generate a plausible reason to be seen, the gateway function the GP used to hold — being the first synthesizer — inverts into a filtering function nobody has resourced: the scarce act is no longer producing the list, it's deciding whose list warrants the physical exam and the accountable signature, at a volume the old gateway was never sized for.

ai-axioms.vercel.app/axioms/healthcare/medicine-gp-visit-ai-axioms