On June 26th a trial in 16 Kenyan primary-care clinics gave medical AI an unusually practical exam. A large-language-model assistant, built into the electronic record, made clinical notes more comprehensive and more likely to contain an appropriate diagnosis and treatment plan. It did not significantly reduce treatment failure within 14 days.1
A second 2026 trial supplied the warning label. Forty-four doctors in Pakistan, all trained to use AI, diagnosed six simulated cases with optional access to ChatGPT. Half were offered unmodified advice; half saw clinically significant errors in three of the six cases. The second group scored 73.3% for diagnostic reasoning, against 84.9% in the first; after adjustment the gap was 14 percentage points.2 The two studies tested different things. Together they show that software may improve the medical record more readily than the medical outcome—and can carry a mistake as efficiently as a lesson.
That mixed verdict matters because AI may change medicine before it cures anyone. Its larger promise is institutional: an error could become cheaper to find, compare with other cases and turn into an alert or protocol. A model does not learn automatically from every patient; people and institutions decide what to capture and reuse. But the useful residue of a mistake could become easier to spread. Call it the error dividend.
Aviation built an institution around the same idea. For half a century NASA's confidential reporting system has offered pilots a bargain: a qualifying report may spare them a civil penalty or licence suspension. Accidents, crimes and deliberate violations are excluded. The point is not absolution, but candour before a near-miss becomes a catastrophe.3 Medicine has national incident systems of its own. But reporting alone does not join incidents, clinical records, outcomes and corrective action across the fragmented institutions that deliver care. A buried error still harms one patient and teaches few others.
One mistake can teach the system, police the clinician—or disappear into a vendor's model.
The useful mistake
Learning starts with a trace. A complaint, review or compensation claim makes an injury visible; an event that leaves none may vanish with the chart. America's National Practitioner Data Bank records malpractice payments. New Zealand and Sweden use no-fault schemes that can compensate injury without first proving a clinician negligent.4 These are not prices on human harm. They are different ways to produce a record from which a system might learn.
The four steps distinguish a learning system from a large archive. Records that cannot be inspected are merely stored; patterns that cannot change practice are merely described. At each step, control matters. Here “ownership” extends beyond title to a patient's chart. It is a bundle of rights: to inspect incidents, audit the software, improve a model, reuse derived knowledge, move it elsewhere and share in any return.
The Kenyan trial shows how that bundle can be split. Kenyan clinicians and researchers helped design and run the study. Its prompt is open, and de-identified inputs, model outputs and clinical outcomes are due to enter a public repository. Yet the underlying GPT-4o model and the electronic-record system are proprietary, and commercial use of the full product must be negotiated separately.5 That split is ordinary. It is also a concrete example of why “who owns the learning?” has more than one answer.
The distinction matters most where adoption is quick and bargaining power uneven. A health system can gain a useful tool and local skill. But if the encounters and validation remain local while the reusable product and commercial rights travel, better care need not bring greater control. Rwanda's three-year memorandum with Anthropic covers health, education and public services, but the announcements do not spell out the rights to data or derivative knowledge. The African Union calls for data sovereignty and locally retained value; the World Health Organization urges governments to seek rights in products built through public-private health projects.6 The real choice is between arrangements that build a learning capacity and those that merely rent one.
One record, three fates
Once mistakes are recorded, they can meet three broad fates. A learning system turns them into safer practice. A control system turns them into rules for auditing clinicians. A fragmented system leaves each hospital and vendor to learn alone. The three will overlap, but they are not equivalent. The same record that warns one doctor can be used to overrule another.
Blame, incorporated
Control is already more than a scenario. Retrospective software can flag decisions that differ from a payer's preferred standard. In a teacher's hands that may improve care. In an insurer's it can make a refusal cheaper. An ongoing federal case alleges that UnitedHealth used nH Predict estimates to press for shorter post-acute stays. Internal records reviewed by ProPublica showed Cigna doctors rejecting more than 300,000 payment claims through PxDx in two months in 2022, averaging 1.2 seconds apiece. Medicare's WISeR model now uses enhanced technology, including AI, to review selected services in six states; CMS says a licensed clinician must make a negative recommendation.7 The human signature does not decide whose standard the system serves.
China offers another route to concentration. Its 2025 plan calls for broad use of imaging and clinical-decision tools in larger hospitals, backed by stronger health-data infrastructure by 2030.8 Pooling records at that scale could support formidable system-wide learning. It would also give the state great influence over who contributes, who benefits and what counts as an error. The plan describes the machinery, not the politics.
Whoever holds the record gains power to set the standard. Yet the system can also change the worker it judges. Aviation's safety engineers learned that an authority gradient may silence the person who spots a mistake; automation researchers have long warned that turning a skilled operator into a monitor creates new failures. Medicine now offers an early warning of the same problem.
In the observational study, experienced endoscopists' unaided adenoma-detection rate fell from 28.4% before routine AI use to 22.4% afterwards; the design cannot establish that AI caused the fall. In the randomised vignette trial, 20 hours of AI training did not prevent doctors from following plausible bad advice.9 “Human oversight” is a weak safeguard if the human has learned not to disagree. It describes a box in a diagram, not proof of independent judgment.
The quarrel in the machine
Most medicine is not yet a coherent learning or control system. Hospitals buy answer engines, bolt-on tools and models embedded in the record one at a time. Each may help locally; together they amount to a procurement programme without a constitution. America is the sharpest example because the software joins an administrative quarrel already in motion.
In 2024 insurers on HealthCare.gov denied 19% of in-network claims. Consumers appealed fewer than 1% of those denials, and insurers upheld 66% of the decisions that were appealed.10 Payment denials differ from clinical errors, but reveal how a consequential decision can be made at scale while the information and stamina needed to challenge it remain scarce.
AI can accelerate every side. Payers can review faster; providers and patients can draft appeals faster. The underlying quarrel survives. But an administrative arms race is also an ownership contest: each clinical note, denial and appeal adds to somebody's store of data. Fragmentation therefore contains the seed of consolidation. The institution with the broadest view can learn faster than those feeding it.
Nor does a large evidence base guarantee useful learning. A review published in March 2026, whose searches ended on December 6th 2024, found that only 24% of 50 studies of predictive clinical decision support involved prospective use; 64% reported technical measures without data about workflow.11 A system may improve a score while learning little about care. The record must include what happened to the patient and how the tool changed the work, not merely whether a model matched a label.
A commons worth having
Ownership is not predetermined. One promising arrangement would combine inspectable software, federated governance and protected reporting. These are different safeguards. Open code permits scrutiny but does not open a patient's file. Federation can leave hospitals and public systems with a say over their data. Reporting protections can reward candour without excusing reckless or criminal conduct.
That last boundary is difficult. The prosecution of nurse RaDonda Vaught after a fatal medication error showed how quickly an argument about accountability can become an argument about whether anyone will report the next near-miss.12 A learning system needs rules for blame, not a promise to abolish it.
Models and application code may become cheaper and easier to copy. A trusted learning institution will not. Its scarce assets are access to well-kept records, permission to compare them, the ability to test a change and a reputation for treating candour fairly. Those are products of governance, not properties of a model.
Aviation did not preserve the pilot by pretending automation could not fly. It gave pilots a new source of authority: stewardship of a system that learned from near-misses. Doctors have the same opening. They can collect mistakes, test machines and reward candour. Otherwise insurers, vendors and states will decide what counts as an error and keep the lesson. The machine will not heal itself. A medical system that owns its mistakes just might.