The MHRA inspectorate (Peter Brown, June 29, 2026 blog post) reported receiving AI-generated inspection responses that cited non-existent guidance. The agency’s position: the company is responsible for the accuracy of everything it submits, no matter which tool drafted it, and responses should get technical review from experienced personnel. Disclosure of AI use is encouraged and heading toward expected. That review is Human verification of AI output must be explicit because document creation no longer carries implicit verification, implemented as Human-in-the-loop must be a named hard gate, not rubber-stamp oversight.
This is the regulatory version of a pattern the whole conference kept hitting. The first AI-related CGMP warning letter went to a company whose defense was that the AI did not tell them a step was needed. The regulator hears that as the company’s failure, because it is. Human error, not AI failure, drives GxP warning letters is the citation record; handing drafting to a model transfers none of the accountability. The inspection-room version of the same rule is Never guess in an inspection; say you will double check. The asymmetry tightens because Regulators are becoming AI users themselves: companies drafting AI-assisted responses are answering reviewers who run AI on their end.
Related
- 2026-09-16-1107-gxp-ai-ml-implementation-framework-lecture
- 2026-09-15-1024-gorani-ai-governance-human-risk-gxp
- Human error, not AI failure, drives GxP warning letters
- Home
- Human verification of AI output must be explicit because document creation no longer carries implicit verification
- Human-in-the-loop must be a named hard gate, not rubber-stamp oversight
- Regulators are becoming AI users themselves
- Never guess in an inspection; say you will double check