Since 2016, FDA has reviewed more than 500 drug submissions containing AI-generated or AI-supported components. Now, sponsors and CROs increasingly use large language models to draft protocols, informed consent forms, investigator brochures, and clinical study reports. The challenge: GxP frameworks were built for deterministic software, not probabilistic language models.
Recent guidance from EMA (September 2024), ICH E6(R3) (January 2025), and FDA (January 2025) has clarified what validation means for AI document tools. The key distinction: not all frameworks apply equally. FDA's 21 CFR Part 11 and ALCOA principles govern systems creating electronic GxP records. ICH E6(R3) requires validation before required use in trials and periodic review to confirm systems stay validated. FDA's draft credibility framework applies only when AI directly affects safety, quality, or study results, not routine document drafting.
The accuracy stakes are real. Studies show LLMs generate factual errors in 1.47% of clinical text and achieve only 57% to 82% accuracy for informed consent documents without structured human review. FDA's own internal AI assistant encountered hallucination problems, prompting the agency to state publicly that AI tools do not replace human judgment.
For sites, this means AI-drafted documents you receive should come with clear audit trails showing model version, reviewer identity, and validation status. Sponsors remain accountable for fitness-for-purpose regardless of vendor assurances.