AI Clinical Trial Recruitment: Where the Discrimination Risk Actually Sits
Trial enrollment used to be a coordinator with a list. Now it is a model that reads the whole EHR, ranks candidates by predicted eligibility and predicted adherence, and hands a site the top two hundred names. That is an algorithmic selection decision inside a federally funded health program — which is a category of decision the law already has strong opinions about.
The step that changed
Recruitment used to fail slowly and visibly. A coordinator pulled a query, called people, and the enrollment log showed who picked up. The bias in that process was real but legible: it lived in which clinics were asked and which phone numbers were current.
AI prescreening moves the filter upstream and makes it invisible. The model reads structured fields, notes, labs, and imaging reports, applies inclusion and exclusion criteria, and produces a ranked list. Nobody at the site ever sees the patients the model scored low. The people who were never contacted leave no trace in the enrollment log — which is why the enrollment log is the wrong place to look for the problem.
How a neutral model produces a skewed cohort
You do not need a protected-class variable to get a protected-class outcome. Trial matching models sit on top of the EHR, and the EHR is a record of who has had access to care. Five routine features do most of the damage:
Chart completeness
Eligibility criteria demand documented values — an ejection fraction, a staging result, a genotype. Patients whose care has been fragmented across under-resourced settings have thinner charts, so they fail eligibility checks for missingness rather than for ineligibility. The model reads absence as exclusion.
Prior specialty engagement
A feature like 'seen by this service in the last 18 months' is a strong predictor of enrollment, and also a strong proxy for insurance status, transportation, and time off work. Optimizing for it selects the already-engaged.
Geography and distance to site
Distance is a legitimate feasibility factor and a well-documented proxy for race and income in most US metros. Using it as a hard cut rather than a logistics flag converts a solvable travel problem into a screening exclusion.
Predicted adherence or retention
The most legally exposed feature in the stack. Adherence models are trained on prior no-shows and dropouts, which track the burdens of poverty far more tightly than they track willingness. Screening on predicted adherence is screening on socioeconomic status.
Language and portal activity
English-language notes and active patient-portal use both improve a candidate's score in systems that use NLP over notes or portal-based outreach. Both correlate directly with national origin.
Which law reaches it
There is no clinical-research-specific AI statute in the United States. There does not need to be. Three existing regimes overlap on the same screening step:
- Section 1557 of the Affordable Care Act — reaches health programs and activities receiving federal financial assistance, on race, color, national origin, sex, age, and disability grounds. Its treatment of patient care decision support tools puts an affirmative obligation on covered entities to make reasonable efforts to identify and mitigate discrimination risk from those tools. A prescreening model that routes patients toward or away from an investigational therapy is difficult to describe as anything other than a decision support tool.
- Title VI — reaches race, color, and national origin discrimination by recipients of federal funds, which covers most academic medical centers and many health systems running trials.
- Section 504 and the ADA — reach disability. Exclusion criteria implemented by a model can sweep in disabilities unrelated to the study question, and accessibility failures in the consent and screening pathway are independently actionable.
Layered on top, FDA's diversity action plan expectations for certain trials create a parallel record. Those expectations are not a private cause of action, but they force a sponsor to state a target population in writing. A written target plus an unexamined model is a worse position than no plan at all.
The advertising half nobody audits
Prescreening is only one input. The other is paid recruitment advertising, and it fails in a way that survives every compliance review, because the compliance review inspects the targeting settings. The targeting settings are usually clean. The delivery is not. An ad platform optimizing for cost per completed screening form will concentrate impressions on whoever converts most cheaply, and that population is rarely the one the diversity plan named. Pull the delivery breakdown by age, sex, and geography, not the campaign configuration — and keep it, because it is the only contemporaneous evidence that the audience you reached matched the audience you promised the IRB.
Who carries the liability
The site
Usually the covered entity under Section 1557 and Title VI. It cannot delegate that status to a vendor, and it is the party a patient names first. It typically has the least visibility into how the model ranks.
The sponsor
Owns the protocol, the diversity commitments, and the marketing application. Its exposure is regulatory rather than civil-rights in the first instance, but it holds the documents that prove what the tool was supposed to do.
The CRO or matching vendor
Increasingly exposed under the agent theory that has advanced in algorithmic hiring litigation: a tool that performs the selection can be treated as acting for the entity that would otherwise perform it. Vendors who market the model as replacing coordinator judgment are describing themselves into that theory.
A screening-tool review that produces evidence
- Measure the funnel above enrollment. Compare the demographic composition of the eligible-by-protocol population, the model-surfaced population, the contacted population, and the enrolled population. The gap between the first two is the model's contribution and the only number that isolates it.
- Inventory the features and name the proxies. Ask the vendor for the feature list in writing. Flag distance, payer, portal activity, prior utilization, and any adherence or retention score, and require a justification tied to the protocol rather than to feasibility.
- Treat missingness as review, not exclusion. Route candidates who fail a criterion for lack of a documented value into human review instead of dropping them. This single change removes the largest structural bias in most matching systems.
- Never screen on predicted adherence. Use it, if at all, to allocate support — travel stipends, flexible visit windows, interpretation — not to remove people from the list.
- Keep ad delivery reports with the trial master file. Configuration screenshots prove intent. Delivery reports prove outcome, and outcome is what gets measured against the plan.
- Contract for audit rights before go-live. Documentation, cooperation, and demographic-reporting obligations are cheap to negotiate before deployment and effectively impossible to obtain after a complaint lands.
The uncomfortable summary
Every party in the chain currently assumes another party is checking the model. The sponsor assumes the CRO validated it, the CRO assumes the site reviewed it, the site assumes the sponsor qualified the vendor, and the vendor assumes it is selling infrastructure rather than making selection decisions. That assumption stack has already collapsed once, in algorithmic hiring, and the theory that collapsed it — that the tool making the cut acts for whoever would otherwise make it — transfers to enrollment without modification.
Frequently Asked Questions
Is clinical trial recruitment covered by anti-discrimination law?
Frequently, yes. Section 1557 reaches health programs receiving federal financial assistance and addresses patient care decision support tools directly. Title VI reaches federally funded programs on race, color, and national origin grounds. Where the recruiting site is a hospital or academic medical center, the research activity generally sits inside the funded program. Voluntary enrollment does not put the screening step outside the statute.
How does an AI prescreening tool become discriminatory?
It learns from who enrolled before. Past enrollment skews toward insured, English-speaking patients near academic centers who were already engaged with specialty care, so a model that predicts enrollment ranks similar patients highest. No race or ethnicity field is required — ZIP code, payer type, prior specialist visits, portal activity, and chart completeness carry the signal on their own.
Do FDA diversity expectations create liability on their own?
Not as a private claim. They matter indirectly. A sponsor that commits to enrollment targets in writing and then runs an unexamined prescreening model has created a documentary record of a stated goal and no verification — which is precisely what an investigation or demand letter asks for first.
Who is responsible when a vendor built the matching model?
All three parties, in different ways. The site remains the covered entity, the sponsor remains responsible for study conduct, and the vendor can face exposure as an agent where its tool effectively makes the selection. Audit, documentation, and cooperation clauses should name all three, because whoever is sued will need records held by the others.
Does recruitment advertising raise the same issue?
A parallel one. Platforms optimize delivery toward whoever converts most cheaply, which can narrow the reached audience well below the approved one without any exclusionary targeting setting. The evidence that matters is the delivery breakdown by demographic, not the campaign configuration.
What is the single highest-value change to make first?
Stop treating a missing documented value as an exclusion. Route those candidates to human review instead. Missingness tracks fragmented access more tightly than any other feature in the stack, and this change is usually a configuration decision rather than a model rebuild.