Can an AI run a clinical-trial consent conversation? A 500-session simulation says the real test has not begun
Three GPT-based agents conducted, played and scored consent conversations for three study protocols. Most concept scores were high, but no prospective participant took part, the evaluator shared the same model ecosystem and factual fidelity was not independently tested.
The short version
- What happened
- The team simulated 500 consent conversations across three real study protocols, using separate AI agents for coordinator, participant and evaluator roles; no human prospective participant was evaluated.
- Who it affects
- Prospective participants may benefit from flexible pacing and repeated explanation, but need an immediate route to a qualified person and a genuine ability to decline.
- What remains unproven
- All 500 participants were AI simulations; no human comprehension, trust, voluntariness or enrolment outcome was evaluated.
Keep in mind
All 500 participants were AI simulations; no human comprehension, trust, voluntariness or enrolment outcome was evaluated.
Related model technology played coordinator, participant and evaluator roles, creating circular-scoring and shared-bias risks.
Factual accuracy and fidelity were not independently measured, the teach-back scale was unvalidated and safeguards were not systematically red-teamed.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: the simulation shows a workflow, not valid consent from people
A peer-reviewed brief report published on 5 November describes an AI-led process intended to explain a clinical study, ask a prospective participant to teach key ideas back in their own words, revisit misunderstood concepts and escalate difficult cases to a human coordinator. In 500 entirely simulated conversations, the system often completed that cycle. This shows that the proposed workflow can run under controlled conditions. It does not show that a patient understood a study, volunteered freely, trusted the system or made a legally and ethically valid decision.
All three roles were synthetic. One AI agent acted as the clinical research coordinator, another acted as a participant with a generated profile, and a third monitored and scored the exchange. The coordinator and evaluator used separate prompts and restricted information, but came from the same commercial large-language-model ecosystem. High scores may therefore reflect compatibility among related agents as well as the quality of the conversation. No validated human comprehension instrument or independent expert rating supplied the main comparator.
The authors propose a human-supervised deployment in which the principal investigator remains responsible, participants can request a person at any time and unresolved comprehension or voluntariness concerns trigger escalation. The evidence supports taking that design into a controlled pilot. It does not support replacing research staff, treating a transcript as proof of understanding or allowing an automated score to authorize enrolment.[1]
How the agents and teach-back score worked
The coordinator received an institutional-review-board-approved consent form, relevant protocol material, eligibility information and synthetic participant characteristics. Its context was restricted to approved sources, with prompts against unsupported advice, coercive language and out-of-scope responses. The intended protections included refusal, escalation, audit logging and continued access to a human coordinator. The participant agent received only its generated profile and eligibility details, with demographic, clinical and behavioural variation added to produce different questions and levels of hesitation.
The evaluator watched the transcript and scored ten fundamental concepts. Zero meant a concept was not mentioned; 0.5 meant the coordinator explained it without a participant response; 0.75 captured acknowledgement or clarification; and one required a satisfactory explanation in the participant's own words. This is more demanding than checking that a chatbot recited a form. But the scale was created for the study and was not validated against an established assessment. The evaluator also steered the coordinator back to missing concepts before judging the result.
Real consent includes ambiguity, emotion, literacy, language, disability, power relationships and freedom to decline. A synthetic participant cannot reproduce those conditions. The simulation also cannot establish whether repeating a question feels supportive or coercive, whether personalization is mistaken for medical advice, or whether people would reveal confusion to a machine. A transcript can help audit while creating sensitive data that need access, retention and deletion rules.[1]
The distinction that matters
What the simulation did—and did not establish
The three-agent loop produced transcript scores, but every participant and evaluator response remained synthetic.
Coordinator
Explains approved study information
A simulated consent dialogue
Participant
Asks questions and teaches back
Synthetic evidence of understanding
Evaluator
Scores ten concepts and prompts repair
4,056 of 5,000 maximum scores
Human oversight
Reviews and handles escalation
Required in the proposed deployment
The denominator: 500 sessions and 5,000 concept scores
The experiment used approved materials from three US research contexts: PROMISE, a comparative-effectiveness trial; COMPARE-VOE, a sickle-cell pain study; and PROMISE-Lung, an observational lung-transplant registry. It simulated 179 PROMISE conversations, 165 COMPARE-VOE conversations and 156 PROMISE-Lung conversations. With ten concepts per session, the principal denominator was 5,000 concept observations—not 5,000 independent people.
Of those observations, 4,056, or 81.1%, received the maximum score. The share was 86.5% for PROMISE-Lung, 79.4% for COMPARE-VOE and 77.1% for PROMISE. Another 7.7% scored 0.75, 6.5% scored 0.5 and 0.7% scored 0.25; 3.9% were not addressed before the session ended. These figures describe automated scoring within the simulation and should not be translated into a human comprehension or clinical effectiveness rate.
At session level, 96 of 500 conversations ended before all ten concepts reached the maximum score, and none of those synthetic participants proceeded to consent. Of 404 complete sessions, 284 resulted in consent. Compensation and care for research-related injury and the point of contact were absent in 16% and 15% of observations, while the right to withdraw was also imperfect. Those misses concern rights and recourse, not peripheral details.[1]
What is untested—and what a real pilot must measure
The paper states that factual accuracy, completeness and fidelity to approved documents were not independently evaluated. Team members reviewed selected conversations during development, but the review was not prespecified or designed to calculate an error rate. A synthetic participant could therefore give a convincing teach-back after an incomplete explanation. The study also lacks a human coordinator, conventional electronic consent or validated comprehension comparator, and its safeguards were not subjected to systematic adversarial testing.
A useful pilot should compare the AI-assisted pathway with an established process, use validated comprehension measures, include independent fidelity review and report who was excluded or needed escalation. It should test out-of-scope questions, conflicting protocol details, withdrawal requests, distress, outages and language switching. Outcomes should include whether people felt free to say no, not simply whether a form was completed. Review boards would also need to examine accessibility, privacy and whether a commercial model provider receives research information.
The authors report no external funding and no conflicts related to the work, while listing unrelated grants, payments, stock options, patents and advisory roles for individuals; one author is the journal's editor in chief. Generative AI assisted drafting and stylistic editing, with authors responsible for the final text. For now the result is engineering feasibility under a clear ethical boundary: the agents can rehearse a governed conversation, but no person has yet been shown to give informed consent through it.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring clinical trials: compare the finding and its evidence limits.
Can an LLM spot what a clinical-trial paper leaves out?
A peer-reviewed benchmark found that a retrieval-plus-LLM pipeline reached F1 0.822 across 119 reporting questions on 40 held-out trial articles. Better evidence retrieval lifted F1 to 0.893, showing both the promise and the main bottleneck.
Read the next reportExplore the full source and evidence record
What this means for people
- Prospective participants may benefit from flexible pacing and repeated explanation, but need an immediate route to a qualified person and a genuine ability to decline.
- Coordinators could spend less time reciting standard text while remaining responsible for complex questions, capacity concerns and valid consent.
- Investigators and review boards inherit duties for model versioning, transcript privacy, factual-fidelity audits, accessibility and incident response.
Global context
The paper uses three US protocols and US consent requirements. Legal responsibility, documentation, language access and data protection differ across jurisdictions. A system piloted in one trial or country cannot be assumed suitable for emergency, paediatric, cognitively impaired or otherwise vulnerable populations elsewhere.
What the evidence does not yet show
- All 500 participants were AI simulations; no human comprehension, trust, voluntariness or enrolment outcome was evaluated.
- Related model technology played coordinator, participant and evaluator roles, creating circular-scoring and shared-bias risks.
- Factual accuracy and fidelity were not independently measured, the teach-back scale was unvalidated and safeguards were not systematically red-teamed.
What to watch next
- Prospective results with real participants and a defined comparator.
- Independent fidelity, voluntariness and comprehension scoring.
- Escalation, refusal, withdrawal and adversarial-test failure rates.
- Subgroup outcomes by language, literacy, disability and digital access.
Living evidence record
Impact record IAI-0ZEL9JD
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Journal of Clinical and Translational Science published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Whether a supervised multi-agent system can cover and test informed-consent concepts in simulated clinical-trial conversations
Topics: Clinical trials · Informed consent · Large language models · Research ethics
Regions: United States · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.