How to (not) use AI for GxP Inspection Responses

In a recent blog post, the U.K. Medicines and Healthcare Products Regulatory Agency MHRA has outlined examples of responses to inspection reports that were generated by artificial intelligence (AI).

Although the use of AI can offer advantages, such as clarifying complex technical issues, providing consistent responses and speeding up the routine drafting of documents, the agency has also been presented with responses to inspection reports that do not meet the requirements and, in some cases, contain alarming errors. The following examples are cited:

• References to non-existent MHRA guidance
• Citations of inappropriate regulatory frameworks
• Responses to serious deficiencies that were misleading rather than addressing the underlying problems.

For example, one response to an inspection received by the MHRA had over 90 pages in total and failed to address the identified deficiencies. This resulted in a disproportionate amount of time being spent on reviewing and responding to it.
In at least one case, an AI-generated response to a serious deficiency contained significant inaccuracies. The content of this response required review by multidisciplinary teams. The time taken for review and response in this instance rose from around 4 hours to over 20 hours.

According to the MHRA, the improper use of AI has become an "actual, realised risk".

In principle, companies are responsible for the accuracy of their responses. Providing false information to inspectors or the authority is considered a breach of regulatory requirements. These responsibilities do not change simply because the tools used to prepare documentation have changed.

The MHRA is not concerned with whether AI is used or not, but rather that the responses are accurate, verifiable and have been prepared under appropriate supervision.
In this regard, the authority requires that all submissions/responses are:

• factually accurate and verifiable
• technically reviewed by appropriately experienced people
• signed off by someone with authority and accountability
• supported by evidence for factual claims
• appropriate to the specific regulatory context.

The MHRA now offers the option to disclose the use of AI in responses to compliance teams. This is not mandatory, but is intended to promote transparency. The blog describes in more detail what this might look like.

AI-assisted definition of actions:
Responses to inspection reports should also set out effective corrective and preventive actions (CAPA). These are based on identifying the root causes, assessing the impact proportionately, implementing measures, and verifying their effectiveness. AI tools can support parts of this process but cannot replace the technical understanding and organisational knowledge required for this, according to the MHRA.

It should also be noted that if AI compiles the response, no one within the organisation will learn anything from the observations. And if regulatory authorities then begin to use AI to produce their inspection reports, we will have a situation where AI inspection reports are responded to with AI-generated responses.