AI in evidence generation: what regulators and HTA bodies say

When AI helps produce the evidence in a regulatory or reimbursement submission, who is accountable for what it produces?

Regulators and HTA bodies have started to answer. Across very different institutions, the answer is the same. A human stays accountable for the result. AI supports the expert. It does not replace them.

The use-case for AI is real. The average systematic review takes more than a year to complete.[1] AI has potential to cut that to weeks or even days - a genuine step change in drug development timelines.

The key challenge is capturing the speed without losing the rigour that regulators and HTA bodies require.

In the EU, no step of a JCA dossier can be fully automated

The EU Joint Clinical Assessment (JCA) is the shared clinical assessment that now supports reimbursement decisions across EU member states. In July 2026 the group that oversees it, the EU HTA Coordination Group, published its first principles for using AI to prepare a JCA dossier.[2]

The company preparing the dossier, not the AI tool, is accountable for its content, methods and findings. That accountability includes the decision to use AI at all, and the validity of anything it produces.[2]

One rule sits at the centre of the principles:

“None of the steps in the dossier preparation or analysis process should be fully automated without a human being ultimately responsible for the quality and accuracy.”[2]

Disclosure is required alongside oversight. Every AI-assisted step must be declared, from searching and screening through data extraction, risk of bias assessment, analysis and reporting. The dossier must record the tool used, its version and developer, and its purpose, and must keep the prompts for inspection on request.[2]

NICE, Canada and France set the same rule

National HTA bodies that have addressed AI have reached the same position.

NICE stated it first. Its 2024 position is that AI should augment human involvement, not replace it, with a capable and informed human kept in the loop. The organisation making the submission stays accountable for its content.[3]

Canada’s Drug Agency adopted NICE’s position in 2025, almost word for word, keeping the same requirement that the submitter remains accountable for what it submits.[4]

France’s HAS takes the same line. A manufacturer may use AI to help prepare a reimbursement application, but remains solely responsible for its final content.[5]

Formal HTA guidance on AI is still new. A review of ten global markets found that, as of August 2025, only NICE and Canada’s Drug Agency had published dedicated position statements. Wider acceptance is expected to come through combined human-and-AI approaches rather than automation alone.[6] The EU has now built the same principle into the JCA itself.[2]

Medicines regulators take the same position for drug development

The principle holds beyond HTA, in how medicines regulators treat AI across drug development.

In January 2026 the FDA and EMA jointly published guiding principles for good AI practice in drug development. Their first principle is that AI should be human-centric by design, and that performance should be judged on the whole system, including the human-AI interaction, not the model alone.[7]

The FDA had already set out its approach in 2025 draft guidance built around a risk-based credibility assessment. How much a sponsor must prove about an AI model depends on how far it influences a decision and how serious an error would be. The higher the stakes, the more evidence the model needs.[8]

This framework is risk-based rather than a ban on autonomy, so a high-influence model can in principle carry weight if the evidence supports it. In practice the human stays central. The FDA’s earlier discussion paper lists human-led governance, accountability and transparency first among its requirements for trustworthy AI.[9] The EMA’s reflection paper calls for a human-centric approach throughout, and for generative AI outputs to be used under close human supervision.[10]

Practice matches the guidance. A 2026 review of EMA advice and public reports identified 43 AI tools across 52 documents and found that human review alongside AI output remains the standard safeguard. The first AI tool to gain full EMA qualification had human oversight built into its approved use, and the review found no case of a fully autonomous tool contributing to the main regulatory evidence.[11] FDA leadership has argued the same, that oversight of complex models must come from people and institutions, and must continue after deployment.[12]

Evidence-synthesis standard-setters keep humans accountable too

The organisations that set the standards for evidence synthesis have reached the same position.

RAISE, the Responsible use of AI in evidence Synthesis project, is the field’s emerging standard. It holds that the synthesist remains accountable for the content, methods and findings, including the decision to use AI, that an AI tool cannot be an author, and that AI should enhance human work rather than replace it.[13]

RAISE also separates two things that can look alike. A rigorous review, where human experts do the work with the help of AI tools, is not the same as an AI platform that auto-generates a summary. The technology overlaps. The product does not.[13] Its companion paper sets the test of a tool: an AI tool is only as good as the human reference standard it is measured against, and that comparison should be reported openly.[14]

The major synthesis organisations have endorsed this. In late 2025, Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence issued a joint position. The synthesist is ultimately responsible, including for the choice to use AI. AI is used with human oversight. Any AI use that makes or suggests a judgement is reported in full.[15]

Their position is not to avoid AI. Avoiding it can be wasteful. A single reviewer screening studies alone misses an estimated 13% of relevant records,[16] and AI acting as a second reviewer can help close that gap.[15] The point is to pair AI with human oversight and to validate it. As Cochrane’s editor-in-chief put it in Nature, a systematic review is not a purely computational task, and the goal is systems where people and AI work together rather than AI working alone.[17]

One way of working emerges

Set these positions side by side and one approach emerges, consistent across regulators, HTA bodies and standard-setters.

A human stays accountable for every step. AI supports the work rather than replacing it. The required rigour scales with the stakes. Every AI-assisted step is declared and documented. Tools are validated against a suitable reference standard, and the results are shared. Oversight continues over time.

Evidax TRACE was built on exactly this model, and it strongly represents a new frontier of evidence generation for HTA and payer dossier submissions.

Take home

Regulators and HTA bodies are setting out how teams may use AI in evidence generation, with expert humans remaining accountable for the work.[2,3,7,8,15]
The consistent rule, across regulators and HTA bodies, is that a human stays accountable for the result. The AI does not.[2,3,4,15]
In the EU this is now explicit for JCA dossiers. No step may be fully automated without a person responsible for its quality, and every AI-assisted step must be declared.[2]
Responsible use is the credible option, not the cautious one. Standard-setters draw a clear line between expert review supported by AI and an auto-generated summary.[13]
A tool is judged by validation against a suitable reference standard, typically expert human review, and reported transparently.[14]

References

Borah R, Brown AW, Capers PL, Kaiser KA. Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry. BMJ Open. 2017;7(2):e012545. https://doi.org/10.1136/bmjopen-2016-012545
Member State Coordination Group on Health Technology Assessment. General Principles on the Use of Artificial Intelligence in the Preparation of Dossiers for Joint Clinical Assessments. Adopted 15 July 2026, pursuant to Article 3(7)(c) and (h) of Regulation (EU) 2021/2282. https://health.ec.europa.eu/health-technology-assessment/implementation-regulation-health-technology-assessment/member-state-coordination-group-hta-htacg_en
National Institute for Health and Care Excellence. Use of AI in Evidence Generation: NICE Position Statement. Corporate document ECD11. Published 15 August 2024. https://www.nice.org.uk/corporate/ecd11
Canada’s Drug Agency (CDA-AMC). Position Statement on the Use of Artificial Intelligence in the Generation and Reporting of Evidence. April 2025. https://www.cda-amc.ca/sites/default/files/MG%20Methods/Position_Statement_AI_Renumbered.pdf
Haute Autorité de Santé. Numérique et intelligence artificielle à la HAS [Digital technology and artificial intelligence at the HAS]. Published 9 April 2025, updated 30 January 2026. https://www.has-sante.fr/jcms/p_3599637/fr/numerique-et-intelligence-artificielle-a-la-has
Skowron R, Remuzat C, Barbier S, François C. HTA18 Acceptance of Artificial Intelligence in Evidence and Dossier Development by Global HTA Agencies. Value in Health. 2025;28(12 Suppl):S403. https://www.ispor.org/publications/journals/value-in-health/abstract/Volume-28—Supplemental-Issue-12S1/HTA18-Acceptance-of-Artificial-Intelligence-in-Evidence-and-Dossier-Development-by-Global-HTA-Agencies
European Medicines Agency, US Food and Drug Administration. Guiding Principles of Good AI Practice in Drug Development. January 2026. https://www.ema.europa.eu/en/documents/other/guiding-principles-good-ai-practice-drug-development_en.pdf
US Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft Guidance for Industry. January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
US Food and Drug Administration. Using Artificial Intelligence and Machine Learning in the Development of Drug and Biological Products: Discussion Paper and Request for Feedback. May 2023 (revised February 2025). https://www.fda.gov/media/167973/download
European Medicines Agency (CHMP, CVMP). Reflection Paper on the Use of Artificial Intelligence (AI) in the Medicinal Product Lifecycle. EMA/CHMP/CVMP/83833/2023. 9 September 2024. https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf
Nollen L-M, van Westen GJP, Westman G, Pasmooij AMG. Artificial intelligence for regulatory evidence: a systematic document analysis of European Medicines Agency regulatory advice and public reports. Clinical Pharmacology & Therapeutics. 2026;120(1):224-235. https://doi.org/10.1002/cpt.70300
Warraich HJ, Tazbaz T, Califf RM. FDA perspective on the regulation of artificial intelligence in health care and biomedicine. JAMA. 2025;333(3):241-247. https://doi.org/10.1001/jama.2024.21451
Thomas J, Hair K, Noel-Storr A, et al. Responsible Use of AI in Evidence Synthesis (RAISE 2026) 1: Recommendations for Practice. Version 3, 13 March 2026. Open Science Framework. DOI 10.17605/OSF.IO/FWAUD. https://osf.io/cqa82
Thomas J, Hair K, Noel-Storr A, et al. Responsible Use of AI in Evidence Synthesis (RAISE 2026) 2: Building and Evaluating AI Evidence Synthesis Tools. Version 3, 13 March 2026. Open Science Framework. DOI 10.17605/OSF.IO/FWAUD. https://osf.io/cqa82
Flemyng E, Noel-Storr A, Macura B, et al. Position statement on artificial intelligence (AI) use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence 2025. Cochrane Database of Systematic Reviews. 2025;(11):ED000178. https://doi.org/10.1002/14651858.ED000178
Gartlehner G, Affengruber L, Titscher V, Noel-Storr A, Dooley G, Ballarini N, et al. Single-reviewer abstract screening missed 13 percent of relevant studies: a crowd-based, randomized controlled trial. Journal of Clinical Epidemiology. 2020;121:20-28. https://doi.org/10.1016/j.jclinepi.2020.01.005
Sarkar R. Why AI can’t be trusted to write scientific reviews. Nature. 2026;653:983 (World View, 28 May 2026). https://www.nature.com/articles/d41586-026-01616-3