Artificial intelligence in clinical trials, what it does well and where it stops

Clinical development generates and consumes enormous amounts of text: protocols, registry records, publications, investigator brochures, clinical study reports and regulatory correspondence. Much of the work around a trial is reading that text, extracting facts from it and writing new documents based on it.

Language models are good at exactly this kind of work: finding relevant passages, summarising them, extracting structured fields and drafting first versions. Machine learning on structured data adds pattern recognition for enrolment forecasting and site performance.

The opportunity is speed and coverage. The risk is that a fluent answer can look correct when it is not, and in clinical development an unsupported statement can end up in a protocol, a slide for a steering committee or a regulatory document.

Before a protocol is final, a team needs to know which comparable trials exist, which are recruiting, which endpoints they use and how many participants they target.

• AI search across registries finds trials by meaning rather than exact keywords, so a question about a mechanism of action also surfaces trials that only list a code name.
• Structured extraction turns registry records into a comparable table of phase, population, interventions, endpoints, sponsor, status and enrolment.
• Summaries of eligibility criteria across competing trials show where recruitment pools overlap.

A human still decides what the landscape means for the protocol. Our article on trial landscape analysis before protocol design covers that step in detail, and the CRO feasibility use case shows how teams apply it.

Site selection depends on where similar trials have run, which sites recruited, and how crowded a region already is with competing studies.

• Registry location fields show where comparable trials are active, by country and by site.
• Models trained on historical operational data can forecast enrolment rates and flag sites at risk of under recruiting, provided the sponsor or CRO has good historical data.
• Competing trial density by region helps avoid placing a study where many similar trials already compete for the same patients.

Public registries hold location and status data, but they do not hold site performance metrics. Forecasting quality depends on proprietary operational data that each organization holds for itself.

Matching patients to trials is one of the most discussed uses of artificial intelligence in clinical trials. Two directions exist.

• From the trial side, language models read free text eligibility criteria in registry records and turn them into structured conditions that can be checked.
• From the patient side, systems compare a patient profile, usually from an electronic health record, against those structured criteria to suggest trials worth considering.

Eligibility criteria are written in natural language with exceptions, nested conditions and lab thresholds, which makes them hard to parse reliably. Any matching system needs clinical review of its suggestions. Matching against identifiable patient records also raises privacy and governance questions that a registry search does not, which is why ClinicalEvidence works only with public trial and literature data and needs no patient data at all.

Safety surveillance, competitive intelligence, medical information and HEOR all depend on keeping up with new publications and registry updates.

• Saved questions with alerts notify a team when new trials or publications match a defined question, instead of rerunning searches by hand.
• Screening support ranks retrieved records by relevance to a PICOS question, so reviewers read the most relevant ones first.
• Extraction pulls study design, population, arms and outcomes into a structured form that a reviewer checks against the source.

For formal systematic reviews, AI assists screening and extraction but does not replace dual independent review or a documented search strategy. Our pages on AI literature review and systematic review software explain where that boundary sits.

Drafting is where language models look most impressive and where checking matters most.

• First drafts of background sections, rationale, and summaries of prior evidence for protocols and briefing documents.
• Plain language summaries of trial results for lay audiences.
• Consistency checks between a protocol, its synopsis and its registry entry.

Every factual statement in a draft needs a source that a writer can open and verify. A draft that states a trial result without a traceable reference creates more work than it saves, because someone has to find out whether it is true.

A language model generates text that is plausible given its training and its input. Plausible is not the same as true, and several limits follow from that.

• It can invent references. A model asked for sources without access to a real index may produce trial identifiers or citations that look valid and do not exist. We explain the mechanism in why AI invents citations.
• It does not know what it has not seen. Training data has a cutoff, and registry status, posted results and new publications change daily.
• It cannot judge trial conduct from a summary. Risk of bias, protocol deviations and selective reporting require reading the full record and often the full publication.
• It cannot resolve conflicts between sources by itself. When a registry entry and a publication disagree on the primary endpoint, a human has to decide which to trust and why.
• It does not carry clinical responsibility. Interpretation for a decision about a patient, a protocol or a submission stays with qualified people.

Regulators are paying close attention to the use of AI across the medicinal product lifecycle. The US Food and Drug Administration has published discussion papers and draft guidance on the use of AI and machine learning in drug development, and the European Medicines Agency has published a reflection paper on the use of AI in the medicinal product lifecycle.

The common themes are consistent across these documents, even where the details differ.

• A risk based approach, where the level of scrutiny depends on how much an AI output influences decisions about patient safety or trial results.
• Transparency about what a model does, what data it uses and how its outputs are checked.
• Data quality and representativeness, including how training and input data were selected.
• Human oversight and documented validation for the specific context of use.
• Traceability, so that an output used in a regulated process can be reconstructed and audited later.

Teams using AI tools in regulated work should read the current versions of these documents directly from the agencies, because guidance on this topic is updated as experience grows.

The practical answer to most of these limits is grounding: the model answers only from passages retrieved from a defined, verifiable index, and every statement points to the record it came from.

• Grounding limits the answer to what the evidence actually contains, instead of what the model remembers or guesses.
• Citations let a reader open the registry record or abstract in one click and check the statement.
• Automatic citation verification catches identifiers that do not exist and sentences that the cited passage does not support.
• An explicit statement that no evidence was found is more useful than a confident answer with nothing behind it.
• A record of which sources were searched makes the answer auditable, which is the property regulators ask for.

ClinicalEvidence answers clinical questions only from passages retrieved from its index of registered trials and published studies: ClinicalTrials.gov, PubMed, the PubMed Central open access subset, the EU Clinical Trials Register and the WHO ICTRP where licences allow, refreshed daily.

• Every sentence carries NCT numbers or PMIDs that link to the source record.
• Each citation is checked to exist in the index and to support the sentence, and an answer without verified sources is not shown.
• When the index holds no evidence, the answer says so and never fills the gap from model memory.
• The Sources searched panel lists only the sources actually searched for that question.
• The trial table, intervention comparisons, alerts and exports turn one question into a reusable piece of work for the team.

See how the clinical trial search works across registries, or read how cited AI answers are verified.