If you care about who gets good healthcare - patients, clinicians, hospital leaders, policymakers, or anyone with a body that occasionally files complaints - this Lancet comment matters because AI may not just reflect health inequity. It may learn it, automate it, and feed it back into the system with the smug efficiency of a child who alphabetized the spice rack but put salt under "rock."
Josip Car, Tien Yin Wong, and Rifat Atun call this the recursive care law: a warning that AI in healthcare can turn existing inequality into a self-reinforcing loop. The idea builds on Julian Tudor Hart's 1971 inverse care law, which said good medical care often becomes least available to the people who need it most. Grim, but at least familiar. The recursive version adds software, feedback, and scale. Because apparently healthcare inequity looked at regular unfairness and said, "Can we make this update automatically?"
The Loop With a Lab Coat
Here is the basic mechanism.
A health system already under-serves certain groups - maybe because of poverty, geography, racism, disability, language barriers, immigration status, spotty insurance, or all the usual villains having a team meeting. Those patients generate less complete data, get diagnosed later, receive fewer specialist referrals, and show up in the electronic record as if their medical lives are blurrier than everyone else's.
Then an AI model trains on that data.
The model sees fewer examples from underserved patients, or sees them mostly when things have already gone sideways. It learns patterns from the record, not from reality. This is the part where we gently take the model by the shoulders and say: sweetheart, the chart is not the patient.
Now the model enters clinical practice. It helps prioritize appointments, flag risk, recommend follow-up, summarize histories, or support diagnoses. If it underestimates risk for groups already missing from the data, those patients receive less attention again. That creates even worse data next time. The loop tightens.
Output becomes input. Bias becomes infrastructure. The model is not just making a mistake - it is helping manufacture the future evidence that makes the mistake look reasonable.
Why This Is Sneakier Than "Bad Data In, Bad Data Out"
"Bad data in, bad data out" is true, but too tidy. The recursive care law is messier and more annoying, like glitter in a hospital carpet.
The problem is not only that training data can be biased. It is that deployed AI can change what happens next. A triage tool may influence who gets seen. A documentation system may influence what gets recorded. A risk model may influence who gets monitored. Once those choices enter the record, future systems treat them as facts.
This matters because machine learning is very good at finding patterns and very bad at asking whether those patterns exist because society has been doing something embarrassing for decades. Your AI may discover that one group has fewer specialist referrals and conclude they need fewer specialists. Proud of you for spotting the pattern, champ. Deeply concerned about the interpretation.
Recent research keeps finding versions of this problem. A 2025 npj Digital Medicine framework argues that healthcare LLMs need standardized audits for both accuracy and bias, including stakeholder input, calibration to local patient populations, clinically relevant test scenarios, and ongoing monitoring for data drift. In other words: do not just launch the model and hope vibes count as governance.
Benchmarks are also getting sharper. FairMedQA, a 2025 arXiv benchmark, tested 12 large language models on counterfactual medical question pairs and found accuracy gaps of 3 to 19 percentage points across demographic groups. That is not a rounding error. That is the model acing the exam and then tripping over the patient ID bracelet.
Medical imaging is no escape hatch either. A 2024 Nature Medicine study found that models can rely on demographic shortcuts in ways that look fine during internal testing but break under real-world generalization. The model may be "fair" in the lab, then meet a new hospital and immediately need adult supervision.
What Fixing It Actually Looks Like
The recursive care law does not say "never use AI in healthcare." That would be too easy, and also wrong. AI can help catch disease earlier, reduce paperwork, expand access, and support overstretched clinicians. The kid has talent. We are not grounding it forever. We are taking away the car keys until it learns what a stop sign is.
Car and colleagues are pointing toward a harder standard: equity has to be designed into the whole lifecycle. That means representative datasets, subgroup evaluation, local validation before deployment, monitoring after deployment, transparency about failure modes, and governance with people affected by the system - not just people who enjoy saying "pipeline" in meetings.
It also means health systems should measure whether AI changes care patterns over time. Are some groups getting fewer referrals after deployment? Are language-minority patients receiving worse summaries? Are rural clinics being scored with assumptions from wealthy urban hospitals? Are model updates drifting quietly like a Roomba with clinical privileges?
The most dangerous AI in healthcare may not be the one that makes a spectacular error. Spectacular errors get investigated. The scarier one is the polite model that nudges care slightly away from already-neglected patients, every day, at scale, while everyone compliments its dashboard.
The Takeaway
The recursive care law gives us a name for a very modern failure: using yesterday's inequity as tomorrow's training signal. It reminds us that healthcare AI does not arrive in a clean room. It arrives in hospitals, clinics, insurance systems, billing codes, missing data, overloaded staff, and human history wearing a stethoscope.
If we build these tools carefully, they can help widen access. If we build them lazily, they will learn our worst habits and perform them faster. And then, like any brilliant child making dumb mistakes, they will look us right in the eye and say the math checks out.
References
-
Car J, Wong TY, Atun R. "The recursive care law: artificial intelligence reinforcing feedback loops and health inequity." The Lancet (2026). DOI: 10.1016/S0140-6736(26)00982-7
-
Hart JT. "The Inverse Care Law." The Lancet (1971). DOI: 10.1016/S0140-6736(71)92410-X
-
Templin T, Fort S, Padmanabham P, et al. "Framework for bias evaluation in large language models in healthcare settings." npj Digital Medicine (2025). DOI: 10.1038/s41746-025-01786-w
-
Xiao Y, Huang J, He R, et al. "FairMedQA: Benchmarking Bias in Large Language Models for Medical Question Answering." arXiv (2025). arXiv:2505.19562
-
Yang Y, Zhang H, Gichoya JW, Katabi D, Ghassemi M. "The limits of fair medical imaging AI in real-world generalization." Nature Medicine (2024). DOI: 10.1038/s41591-024-03113-4
-
Liu M, Ning Y, Teixayavong S, et al. "A scoping review and evidence gap analysis of clinical AI fairness." npj Digital Medicine (2025). DOI: 10.1038/s41746-025-01667-2
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.