EU GMP Annex 22.

The draft applies to computerized systems across the GMP environment - from production to quality processes such as deviation management - wherever AI models operate in critical applications, meaning they bear directly on patient safety, product quality or data integrity, for example by predicting or classifying data. Criticality is not a matter of direct impact alone: it also rises where errors are hard to detect and the degree of automation is high - our reading of the ongoing discussion. The models covered are machine-learning models that acquire their function through training, and the draft draws a deliberately narrow boundary: in critical applications, only static models with deterministic output are in scope. The ten sections show how all of this fits into the wider picture of AI in GxP.

What runs through all of these sections is a change in the form of the evidence: from deterministic single tests to statistical evidence. Test data that must itself meet requirements, metrics robust enough to carry claims about control and effectiveness - this is how data scientists think, and it is also the core of quality risk management under ICH Q9(R1): deciding under uncertainty. Not a break with GxP logic, but its consistent application.

For perspective: rule-based automation without machine learning remains a straightforward Annex 11 case, while static ML models with deterministic output are Annex 22 territory. In critical applications the draft leaves no room for dynamic, probabilistic or generative models. Everywhere else they remain usable - with the risk assessment the current Annex 11 requires of any computerized system, and in our view always under human oversight, which the draft makes an explicit condition for generative AI and LLMs.

The draft says yes. It does not apply to generative AI and Large Language Models, and they should not be used in critical GMP applications; outside critical applications, their use remains possible - always with human review of the outputs (Human-in-the-Loop, HITL). Responsibility for their use remains with the regulated company. This is precisely where the consultation's comment priorities, as summarized by the EMA, take issue:

• Risk instead of technology bans

  Assessment per use case instead of categorical exclusion of entire model classes.

• Define "critical application"

  A robust criticality methodology instead of ad-hoc classification.

• Harmonization

  Coherence with ICH Q9(R1), Annex 11 and the EU AI Act instead of parallel control worlds.

• Training data and bias

  Dedicated requirements for provenance, representativeness and bias assessment.

• Clarify human oversight

  Design oversight based on risk - HITL is one form of it, not the only one.

• Model lifecycle

  Rules for versioning, retraining and AI functions delivered from the cloud.

Our reading: the logic behind this critique is that of quality risk management - higher risk means stronger controls, not a blanket ban. The EMA itself has put this direction on the agenda of its expert workshop: the first of six topic areas ("Regulatory pathways for adaptive / probabilistic AI models in GMP") asks outright, "How could we accommodate adaptive and probabilistic models in Annex 22?" What is being negotiated is a pathway for these models, no longer just their exclusion. Whether the final text follows that line is for the ongoing revision to decide.