AI in Clinical Trials: How Artificial Intelligence Is Helping Researchers Test New Medicines

Last updated: October 27, 2026

Editorial note: This article draws on peer-reviewed studies in Nature Communications, Nature Medicine and Drug Discovery Today, and on official documents from the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA). Figures reported by companies are labeled as such. Helathora has no commercial relationship with any company mentioned, and this article is not investment or medical advice.

Every new medicine has to prove itself in clinical trials before it reaches patients. Those trials are slow, expensive and often unsuccessful. Researchers must find volunteers who meet long lists of eligibility criteria, recruit enough people to detect a real effect, measure results consistently across dozens of sites and keep participants safe throughout. Most candidate drugs that enter human testing never reach approval.

Artificial intelligence is now being applied at almost every one of those steps. Some uses are modest and practical, such as helping staff screen patient records faster. Others are more ambitious, like using prognostic models to sharpen trial statistics or testing molecules that were designed with generative AI. This guide explains what each approach does, what the evidence shows so far and how regulators in the United States and Europe are responding.

Two different questions: “AI in clinical trials” can mean using AI to run trials better, for example to recruit, design or measure, or it can mean testing drugs that were discovered with AI. The two are often mixed up in headlines. A drug designed by AI still has to pass the same human trials as any other medicine, and an AI tool that speeds up recruitment says nothing about whether the drug being tested works.

42.6%Less screening time for clinicians using an AI trial-matching tool in a small user study (TrialGPT)

21 of 24AI-discovered molecules from AI-focused companies that completed Phase I successfully by end of 2023 (industry analysis)

4 of 10Of those molecules that completed Phase II, a small sample in the same analysis (source)

March 2025EMA’s first qualification of an AI tool for generating trial evidence, for one defined use (AIM-NASH)

These figures come from different studies and regulatory contexts. They are not directly comparable and should not be interpreted as evidence that AI improves every clinical trial.

Quick answer: AI is helping researchers find eligible volunteers faster, design trials that may need fewer participants in the control group and measure some outcomes more consistently. Regulators have formally accepted a small number of these methods for specific, defined uses. Molecules discovered with AI have done well in early safety trials, but the limited data available do not show that they are more effective than traditionally discovered drugs. AI can make trials more efficient; it does not make an ineffective medicine work.

Where AI Enters the Life of a Trial

Stage	Task	Example of AI use	Current status
Before the trial	Choosing a drug candidate and target	Generative models proposing molecules and disease targets	Several AI-discovered molecules in Phase I and II trials
Design	Deciding how many participants are needed	Prognostic models used for statistical adjustment, which may allow smaller control groups	One method qualified by EMA for specific trial types
Recruitment	Matching patients to eligibility criteria	Large language models reading patient summaries and trial criteria	Promising research; real-world evaluation under way
Measurement	Scoring outcomes consistently	AI-assisted analysis of biopsy images, supervised by pathologists	One tool qualified by EMA for liver disease trials
Conduct and analysis	Monitoring data quality and safety signals	Automated checks of trial data and documents	Growing use; limited published evidence

Finding the Right Volunteers

Recruitment is one of the main reasons trials are delayed. Eligibility criteria can run to dozens of items covering diagnoses, previous treatments, test results and other conditions, and checking each patient by hand takes clinicians and research staff a great deal of time.

Researchers at the US National Library of Medicine and National Cancer Institute developed TrialGPT, a system that uses large language models to match patient summaries with trials listed on ClinicalTrials.gov. Published in Nature Communications in November 2024, the study found that TrialGPT could retrieve more than 90% of relevant trials while reviewing less than 6% of the initial collection. When assessing whether patients met individual criteria, it was correct in 87.3% of 1,015 checks, close to the performance of human experts. In a user study, clinicians working with the tool reduced screening time by 42.6%.

Important context: TrialGPT was evaluated on 183 synthetic patient descriptions, not on real hospital records, and the time saving comes from a small user study. It refers to the time needed to screen patients against trials, not to a proven reduction in the total time it takes to recruit a trial. The NIH team has received funding to test real-world performance and fairness, including whether the tool works equally well for underrepresented groups. Like any language model, it can make mistakes, so clinicians remain responsible for every eligibility decision.

Can AI Reduce Control Group Sizes? Understanding Prognostic Models

In a randomized trial, some participants receive the new treatment and others receive a placebo or standard care. The control group is essential, but it means many volunteers do not receive the experimental medicine, and large control groups make trials longer and more expensive.

The company Unlearn developed a method called PROCOVA (prognostic covariate adjustment). A machine learning model trained on historical patient data predicts how each participant would be expected to progress if they received the control treatment. That prediction is then used as a statistical adjustment in the analysis, which can increase a trial’s precision and may allow a smaller control group. In 2022, the EMA’s human medicines committee issued a qualification opinion stating that the procedure could increase power or precision in randomized Phase 2 and 3 trials with continuous outcomes.

The company describes these predictions as “digital twins”, a term that can be misleading. PROCOVA is better understood as a prognostic covariate-adjustment method than as a complete digital replica of a patient. Its potential benefits depend on the predictive performance of the model and the design of the trial.

The prediction does not replace real patients. Trials using this method still randomize participants and still have a real control group; the prediction only sharpens the statistical analysis. Unlearn reports reductions in control-group size of up to about 35%. That is a company-reported figure, not a result guaranteed by the EMA, and the actual gain depends on how well the model predicts outcomes in each new trial. FDA comments on the method noted that power calculations may be optimistic if the model performs less well in the new population than in historical data.

Measuring Outcomes More Consistently

Some trial outcomes depend on expert judgment, and experts do not always agree. In liver disease trials for metabolic dysfunction-associated steatohepatitis (MASH, formerly NASH), the key results often come from pathologists scoring liver biopsy images, and variation between readers can blur the true effect of a treatment.

In March 2025, the EMA’s human medicines committee issued a qualification opinion for AIM-NASH, an AI tool that helps pathologists analyse these biopsy scans. The EMA describes it as the first time it will consider data generated with the assistance of an AI-based tool to be scientifically valid. Crucially, the tool is supervised by a human pathologist, and the agency notes that it can help researchers obtain clearer evidence with fewer patients. The qualification covers this specific use in MASH trials; it is not a general approval of AI for measuring outcomes in other diseases.

Testing Medicines Discovered With AI

A separate question is how drugs designed with AI perform once they reach human trials. A 2024 analysis in Drug Discovery Today, written by researchers at the consulting firm BCG, tracked molecules from AI-focused biotech companies. By December 2023, 21 of 24 AI-discovered molecules that had completed Phase I were judged successful. In Phase II, which tests whether a drug actually works in patients, 4 of 10 succeeded.

Success of AI-discovered molecules that completed each trial phase (to December 2023)

Phase I (21 of 24)

87.5%

Phase II (4 of 10)

40%

Source: Jayatunga MKP, et al. Drug Discovery Today, 2024. The two bars describe small and different groups of molecules, not the same molecules followed through both phases. The authors compare these with historical industry averages of about 40–65% in Phase I and 30–40% in Phase II. The analysis covers AI-focused companies only.

The early results are encouraging, but they should be interpreted cautiously. Phase I trials primarily assess safety, tolerability and pharmacokinetics, rather than whether a drug treats a disease effectively. The reported Phase I and Phase II figures come from small, selected groups of AI-focused companies and do not represent a direct comparison with a matched group of conventionally discovered drugs. They therefore cannot establish that AI-discovered medicines are safer, more effective or more likely to succeed overall.

One example: rentosertib in pulmonary fibrosis

In June 2025, Nature Medicine published a Phase 2a trial of rentosertib, a drug from Insilico Medicine whose target and molecular design were both identified with generative AI. The randomized, placebo-controlled trial enrolled 71 people with idiopathic pulmonary fibrosis at 22 sites in China for 12 weeks. Its main goal was safety, and side effects were broadly similar across groups. At the highest dose tested, lung capacity (forced vital capacity) rose by an average of 98.4 ml, while it fell by 20.3 ml in the placebo group.

Keep this in proportion: The lung-capacity change was observed in a small trial, over 12 weeks and at one specific dose. A difference seen in a small group of patients is not enough, on its own, to demonstrate a confirmed clinical benefit. The trial was sponsored by the company that developed the drug, its primary purpose was safety, and the authors themselves called for larger and longer trials. It is an encouraging signal, not proof that the drug works or that AI-designed drugs are better.

How Regulators Are Responding

Regulators are not banning AI in drug development, but they want evidence that each tool is reliable for the specific job it does. In January 2025, the FDA published a draft guidance on using AI to support regulatory decisions about drugs. It proposes a risk-based framework for assessing the credibility of an AI model for a defined “context of use”, with more scrutiny where a model’s output has more influence on decisions and greater consequences if it is wrong.

In Europe, the EMA adopted a reflection paper on AI across the medicine lifecycle in September 2024, including clinical trials, and advises developers to talk to regulators early when AI could affect a medicine’s benefit-risk assessment. In January 2026, the EMA and FDA published joint guiding principles for good AI practice in drug development, covering everything from early research and trials to manufacturing and safety monitoring.

Regulatory guidance and qualification opinions do not constitute blanket approval of AI systems or of the medicines developed with them. Each application must be evaluated within its intended context of use.

The Risks Researchers Are Watching

AI can introduce new problems into trials as well as solve old ones. Language models can misread records or produce confident but wrong explanations. Prognostic models can lose accuracy when used in populations that differ from the ones they were trained on, which can leave a trial underpowered. Three safeguards matter most for participants.

Privacy: who sees your data, and why

Trials handle highly sensitive health data. In the European Union, the GDPR gives health data special protection, and participants are entitled to clear information about how their data are used. In practice, that means the consent documents should say whether an AI tool will process your records, who runs it (the hospital, the sponsor or an outside vendor), whether data leave the site and how long they are kept. If a recruitment tool scans hospital records before you have been contacted, the trial team should be able to explain the legal basis for that.

Human oversight: what people must still decide

In the methods regulators have accepted so far, AI supports decisions rather than making them. A matching tool may flag that a patient seems eligible, but an investigator confirms each criterion before anyone is enrolled. In the MASH example, the AI tool assists with biopsy scoring, but a pathologist supervises the result. Decisions about safety, such as pausing a treatment after a side effect, remain with the clinical team and independent safety monitors.

Representativeness: checking who gets left out

Recruitment tools trained on historical data may reproduce existing gaps in who is invited to take part, at a time when many trials already struggle to include older people, women and ethnic minorities. One practical risk: a tool that relies on detailed electronic records may miss patients whose records are incomplete, who are often those with less access to care. Good practice is to test a tool’s accuracy separately for different age groups, sexes and backgrounds, and to compare the people it flags with the population the trial is meant to serve.

What Has Been Shown So Far

Application	What the evidence shows	What it does not show
AI patient–trial matching	High accuracy and time savings in a controlled study with synthetic patients	Faster enrolment or more diverse participants in real trials
Prognostic covariate adjustment (PROCOVA)	Regulatory qualification for specific trial types; can improve statistical precision	That control groups can be removed, or a fixed size reduction for every trial
AI-assisted outcome measurement	Accepted by EMA for liver biopsy scoring under pathologist supervision	Acceptance for other diseases or without human oversight
AI-discovered molecules	High Phase I success in a small, early industry analysis	Better safety or efficacy than traditional drugs, in Phase II or overall

Patient recruitmentPrognostic modelsTrial designAI drug discoveryFDA and EMA guidanceData protection

If You Are Invited to Join a Trial

Ask whether AI tools were used to identify you or will be used to analyse your data.
Ask how your health data will be stored, shared and protected, and for how long.
Check that the trial is registered, for example on ClinicalTrials.gov or the EU Clinical Trials Information System.
Make sure you understand the chance of receiving a placebo or standard care instead of the new treatment.
Discuss the decision with your own doctor, and remember that you can withdraw at any time.

Related reading on Helathora

Limitations

Much of the evidence on AI in clinical trials comes from method studies, company reports and regulatory documents rather than head-to-head comparisons of trials run with and without AI. The analysis of AI-discovered molecules is based on small numbers, was carried out by consultants and covers only AI-focused companies. Regulatory qualifications apply to narrowly defined uses. Guidance is changing quickly, so readers should check the FDA and EMA websites for the latest status of the documents described here.

The Bottom Line

AI is making parts of clinical research faster and more precise: screening volunteers, sharpening statistical analysis and standardizing some measurements. Regulators have accepted a few of these methods for specific purposes and are building frameworks for the rest. What AI has not yet shown is that it can make the hardest step easier: proving that a new medicine actually helps patients. That still depends on well-designed human trials, and on the volunteers who take part in them.

Frequently Asked Questions

Can AI replace clinical trials?

No. Regulators still require evidence from studies in people. AI can help design trials, find participants and analyse results, and in some cases reduce the number of people needed in a control group, but it does not replace testing a medicine in humans.

Has an AI-discovered drug been approved?

Several AI-discovered molecules have reached Phase II trials, and some have shown encouraging early results. Approval requires larger confirmatory trials, so readers should check regulators’ announcements for the latest status of any specific drug.

Are “digital twins” fake patients?

Not in the methods regulators have accepted so far. In PROCOVA, the term refers to a prediction of how a real participant would be expected to progress on the control treatment, used to improve the statistical analysis. It is not a complete digital copy of the person, and the trial still includes real participants in both the treatment and control groups.

Could AI make trials less fair?

It could if tools trained on past data reproduce existing gaps in who is invited to take part. That is why researchers and regulators emphasize testing performance across different groups and keeping humans responsible for decisions.

Scientific Studies and Regulatory Sources

Jin Q, et al. Matching patients to clinical trials with large language models. Nature Communications, November 2024. DOI: 10.1038/s41467-024-53081-z.
National Library of Medicine. NIH-developed AI algorithm matches potential volunteers to clinical trials, 2024.
Xu Z, et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine, June 2025. PubMed ID: 40461817.
Jayatunga MKP, et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024.
European Medicines Agency. Artificial intelligence (AIM-NASH qualification, reflection paper and EMA–FDA principles).
European Medicines Agency. Reflection paper on the use of AI in the medicinal product lifecycle, September 2024.
EMA and FDA. Guiding principles of good AI practice in drug development, January 2026.
US Food and Drug Administration. Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products (draft guidance), January 2025.
Unlearn. PROCOVA and the EMA qualification opinion (company source).

Disclaimer: This article is for general information only and does not constitute medical, legal or investment advice. Information about investigational medicines describes research results and does not mean those medicines are approved or available. Anyone considering joining a clinical trial should discuss it with their doctor and the trial team. Helathora is not affiliated with any company or product mentioned.