Medicine#

← III.B - Physical · Part III · Index

The argument in four claims#

1. Diagnostics compress; employment does not. Narrow imaging is already superhuman, and radiologist headcount grew. Reading was never the whole job, and liability does not transfer to a model. → Diagnostics

2. The binding constraint is authorization, not intelligence. Licensed capacity and reimbursement rules bottleneck healthcare; adding models to a system constrained on who may act changes less than it should. → Liability and licensure

3. Consumer-side change outruns the formal system. Free, first-line triage routes around institutions faster than it reforms them - the same pattern as every prior information technology in medicine, at larger magnitude. → Delivery

4. Drug discovery accelerates preclinical; the clinical gauntlet barely moves. ~60% of failures are human efficacy/toxicity. Wet-lab throughput and trial recruitment rate-limit, not model quality. → Drug discovery

The radiology template#

That first claim is the template for every professional domain in this document: AI compresses the diagnostic and documentation layers, leaving licensure, liability, physical examination, and procedure as the human moat.

The moat is more fragile than it looks. In most of Part III it is financial - a malpractice policy exists for a licensed human and does not yet cover an autonomous system - and financial arrangements reprice on an underwriting cycle. → Insurance, Uncertainty 6

Medicine is the best test case for the template precisely because it is the domain where the moat looks most like competence and is most clearly something else. A radiologist's protection is not that no system reads images as well; on several narrow tasks systems demonstrably do. It is that the act of signing carries transferable legal risk and a market exists to price that risk for a licensed human and does not yet exist for a vendor. Anyone arguing that the profession is safe on capability grounds is arguing the wrong case, and will be surprised in the direction the liability page describes.

Why medicine is filed under physical#

The filing is a claim about which layer binds. Most of the visible AI activity in healthcare is symbol work - imaging, notes, coding, prior authorization, triage - and by verification cost that belongs in the cognitive group. What keeps medicine on the physical side is that the outcomes which would validate the symbol work are physical and slow: whether the patient did better resolves in years, cannot be simulated, and is confounded by everything the system did not record. So the domain has a cheap-label layer running at software speed sitting on top of an expensive-label layer running at biological speed, and predictions that conflate the two are the standard error in this sector.

That structure also explains the sector's characteristic pathology. Because the fast layer produces measurable outputs (reports drafted, encounters triaged, minutes saved) and the slow layer produces the outcomes anyone actually wants, adoption gets justified on the metric that is cheap to observe. Health systems can therefore accumulate real efficiency gains at the documentation layer for a decade while population health moves imperceptibly, and both facts will be true at once.

Where the four arguments sit on the clock#

Consumer knowledge and institutional action diverge. The gap between what a motivated patient can know and what the system will act on is a source of friction and political pressure through the 2030s - not a temporary lag.

Failure mode for the hub argument#

All four claims assume the system being changed is the incumbent one. The reading that would break them is an entrant reading: a payer, a retail health chain, or a national system in a country with unified purchasing rebuilds primary care around model-first triage with physicians as an escalation layer, rather than inserting tools into existing clinics. Nothing in the four claims prevents that, because such an entrant does not inherit the fee schedule that makes efficiency unprofitable or the malpractice book that makes autonomy uninsurable; it writes both from scratch. The corpus's base case is slow institutional change because incumbents dominate high-income healthcare, and that base case is weakest wherever purchasing is concentrated enough for one decision to move the whole market. Watch single-payer systems and large integrated payers, not hospital pilots.

What to watch#

High-income vs low-resource sign flip#

In high-income systems, consumer triage risks over-referral into full queues. In low-resource settings the counterfactual is often no care - same tools, opposite welfare sign. → delivery, global south. Do not export one prediction to the other geography without flipping the counterfactual.

Sections#

Related: Science · Data · Insurance · Game 3 · Global South on free diagnostic support

View markdown source