Introduction

If you walk into any clinical operations conference this year and you’ll hear the same buzzwords on repeat: predictive enrollment, risk-based monitoring, automated source data verification, AI-drafted protocols. AI is now being layered onto almost every stage of the trial lifecycle.

Patient recruitment

Natural language processing tools can scan electronic health records and unstructured clinical notes to match patients against eligibility criteria far faster than manual chart review. Cloud providers have built entire toolkits around this problem; AWS, for instance, markets services like Amazon Comprehend Medical and SageMaker specifically to help sponsors extract patient information from medical text and predict which protocols and sites are likely to enroll well, while AWS Marketplace now lists dedicated “clinical trial agent” tools built by consulting partners to automate eligibility verification and trial-criteria research.

Patient engagement and retention

Chatbots and voice assistants (again, an area AWS has invested in with Lex-based solutions) send medication reminders, collect e-diary responses, and flag patients at risk of dropping out.

Protocol optimization

Machine learning models simulate different protocol designs against historical trial data to flag overly complex visit schedules or inclusion/exclusion criteria that would choke enrollment before the trial even starts.

Every one of these use cases sits inside a decades-old regulatory framework that was built for a very different world, one where “the algorithm did it” was never an acceptable line in a deviation report. That framework is Good Clinical Practice.

What Is GCP, and Why Does It Exist?

Good Clinical Practice (GCP) is the international ethical and scientific quality standard for designing, conducting, recording, and reporting clinical trials involving human subjects. It’s codified primarily through ICH E6, the guideline issued by the International Council for Harmonisation, and it underpins how regulators like the FDA and EMA evaluate whether trial data can be trusted enough to support a marketing authorization.

GCP exists for two intertwined reasons:

• Protecting human subjects. Every principle in ICH-GCP traces back to the Declaration of Helsinki: the rights, safety, and wellbeing of trial participants take priority over the interests of science or society. Informed consent, ethics committee oversight, and risk-minimization requirements all flow from this.
• Ensuring data integrity. Regulators can only approve a new medicine if they can trust that the data supporting it is accurate, complete, verifiable, and attributable. GCP’s emphasis on source documentation, audit trails, and investigator accountability exists specifically so that trial results reflect what actually happened to real patients, not what a sponsor hoped would happen. Crucially, GCP assigns personal responsibility to a named individual: the Principal Investigator. The PI is accountable for the medical care of trial participants, for the accuracy of the data generated at their site, and for ensuring the protocol and applicable regulations are followed, regardless of which tools or vendors are involved in supporting the trial. That single design feature of GCP is where AI creates friction.

Implications: Why AI Doesn’t Replace Investigator Responsibility

AI Can Assist. It Cannot Be Accountable.

An algorithm cannot sign a Form FDA 1572. It cannot be named on a delegation-of-authority log. It cannot be held to a state medical board or debarred by a regulatory agency. Accountability under GCP is a legal and ethical construct tied to a human being who has made a professional commitment to the participant in front of them, something no model, however sophisticated, can hold.

• Risk-Based Monitoring (RBM): Endorsed by both ICH E6(R2)/(R3) and FDA guidance, shifts monitoring away from 100% source data verification toward a targeted approach that concentrates resources on the risks most likely to affect data quality and patient safety. AI is a natural fit here as machine learning models can continuously score sites and data points for anomalies (unusual query rates, protocol deviations, inconsistent vital signs, suspicious enrollment patterns) far faster than a human reviewing spreadsheets. However, the human still needs to investigate, interpret context, and decide what action to take.

• Electronic Trial Master Files (eTMFs): As trials generate more electronic data i.e. consent records, monitoring visit reports and AI audit logs, the eTMF has become the backbone of inspection-readiness, giving sponsors and regulators a complete, time-stamped record of how the trial was actually conducted. AI-generated outputs (recruitment scores, monitoring flags, protocol simulations) increasingly need to be captured and version-controlled inside these systems too, so that when an inspector asks “why was this site rated low-risk,” there’s a documented, human-reviewed trail.

Regulatory Expectations Are Already Catching Up

Both the FDA and EMA have published discussion papers and guidance addressing AI/ML use in drug development, and the consistent thread is the same: sponsors must validate these tools, document their intended use and limitations, and maintain human review over any output that could affect patient safety or data integrity.

My Thoughts

AI could be beneficial for clinical research. It can shrink recruitment timelines, catch risk signals a tired monitor might miss at 4pm on a Friday, and take a mountain of unstructured EHR text and turn it into something a study team can actually use. I don’t think anyone doing this work seriously wants to go back to purely manual processes.

But every one of those gains has to run through a simple filter: a human has to check it, every single step of the way. We can never hold an algorithm accountable, so the accountability has to sit with the people who chose to deploy it, validated it, and signed off on its output.

A recent, concrete reminder of why that matters, is to take a look at what happened with AI medical scribes in Ontario . An audit by the province’s Auditor General found that nearly half of the AI transcription systems evaluated for use by doctors fabricated clinical information outright by inventing details like therapy referrals or blood test orders that were never actually discussed, while the majority missed important mental health details patients had raised, and a large share mis-recorded which drug had actually been prescribed. Officials were quick to note that clinicians are expected to review and approve every note before it enters the patient record, which is exactly the point. The safety net wasn’t the AI being flawless. It was the human being required to check its work.

That’s the model I think clinical research needs to hold onto as AI adoption accelerates: build the tools, use the efficiency gains, let the algorithms do the heavy lifting on pattern detection and triage and never, ever let “the model recommended it” become an acceptable substitute for a qualified person taking responsibility. GCP got this right decades before generative AI existed: someone’s name has to be on the line. That principle doesn’t need reinventing for the AI era. It just needs enforcing.

References

1. International Council for Harmonisation (ICH). E6(R3) Guideline for Good Clinical Practice. Step 4, January 2025. ich.org/page/efficacy-guidelines#6-2
2. U.S. Food and Drug Administration. E6(R3) Good Clinical Practice (GCP) – Guidance for Industry. September 2025. fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp
3. European Medicines Agency. ICH E6(R3) Guideline on Good Clinical Practice (GCP) – Step 5. ema.europa.eu/en/documents/scientific-guideline/ich-e6-r3-guideline-good-clinical-practice-gcp-step-5_en.pdf
4. U.S. Food and Drug Administration. Artificial Intelligence for Drug Development (hub page, incl. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft guidance, January 2025). fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development
5. U.S. Food and Drug Administration. FDA Proposes Framework to Advance Credibility of AI Models Used for Drug and Biological Product Submissions. Press announcement, January 2025. fda.gov/news-events/press-announcements/fda-proposes-framework-advance-credibility-ai-models-used-drug-and-biological-product-submissions
6. European Medicines Agency. Reflection Paper on the Use of Artificial Intelligence in the Medicinal Product Lifecycle. October 2024.
7. Amazon Web Services. Clinical Trials Use Cases – Healthcare Cloud Solutions. aws.amazon.com/health/biopharma/clinical-trials
8. Amazon Web Services. Improving Patient Engagement in Clinical Trials Using Voice and Chat with AWS. aws.amazon.com/blogs/industries/improving-patient-engagement-in-clinical-trials-using-voice-and-chat-with-aws
9. Amazon Web Services. AI-Driven Patient Retention and Engagement for Clinical Trials. aws.amazon.com/blogs/industries/ai-driven-patient-retention-and-engagement-for-clinical-trials
10. CBC News. AI Transcriber for Use by Ontario Doctors ‘Hallucinated,’ Generated Errors: Auditor General. May 13, 2026. cbc.ca/news/canada/toronto/ai-scribe-system-hallucinations-9.7197049
11. Global News. AI Systems Used by Ontario Doctors Hallucinate, Auditor General Finds. May 2026. globalnews.ca/news/11844349/ontario-auditor-general-ai-usage