Note: Annex 22 is available as a draft (as of 5 October 2026). This content is not legal, regulatory or compliance advice.
Short answer
The EU GMP Annex 22 draft provides that generative AI and LLMs are not used in critical GMP applications. The 2025 consultation, however, showed support for enabling them under certain conditions. The EMA held an expert workshop on this on 30 June and 1 July 2026. The draft is being revised and the outcome is open.
What the draft says today
In its draft form, Annex 22 targets static models with deterministic outputs: models that are frozen after training and return the same result for the same input. Dynamic models that keep learning in operation, and probabilistic models such as generative AI and large language models (LLMs), should not be used in critical GMP applications.
This is not a rejection of AI as a whole. It is a boundary for one specific area: applications with a direct effect on product quality, patient safety or data integrity. For non-critical support with human review, the draft does not make this statement in the same way.
Why LLMs in particular are excluded
Three properties make generative AI difficult in a GMP environment:
- Lack of reproducibility: an LLM can give different answers to the same input. Validation, however, requires that a result can be repeated.
- Plausible errors: a language model writes fluently and convincingly even when the content is wrong. A wrong number in a well written sentence is harder to spot than an error message.
- Hard to bound scope: a model that answers almost any question has no natural edge. For a defined intended use, that is a problem.
What the EMA has been reassessing since 2026
According to the EMA, the 2025 consultation showed support for potentially enabling generative AI and LLMs. The EMA therefore held a multistakeholder workshop with expert contributions on 30 June and 1 July 2026. The aim was a risk-based approach to the use of generative AI, with control and mitigation measures (guardrails) as a central topic.
A presentation by the Irish regulator HPRA at the QP Forum 2026 describes the state of play as follows: the draft is under revision, with further targeted consultations scheduled for 2026. Where AI is used, testing and confirming the outputs is key, for example through a human in the loop or other guardrails.
What remains open
It remains open whether and under which conditions the final text will allow generative AI in more critical applications, which controls would then be required, and when the final text will arrive. We make no prediction about this. Anyone planning today should be able to handle either outcome.
How labs can prepare for either outcome
Whatever the EMA decides, some preparation pays off in any case:
- Separate language from computation: if an LLM only plans and writes, while the numbers come from reproducible tools, the critical part remains checkable. CovaSyn follows this principle: the language model plans, the tools compute, every step is traceable.
- Define human review: for each application, it is clear who checks the result before it feeds into a decision.
- Traceability from the start: model version, inputs and result are recorded so that you can later show how a result came about.
- AI literacy: since 2 February 2025, the EU AI Act has required deployers of AI systems to ensure sufficient AI literacy among their staff (Art. 4). This applies independently of Annex 22.
The article AI in the QC Lab: Use Cases With a Human in the Loop shows what this can look like in everyday QC work. The distinction from Annex 11 is explained in Annex 11 vs Annex 22: What Applies to What?. The pillar page AI in the GxP Lab and EU GMP Annex 22 gives the overview.
Frequently asked questions
May LLMs such as ChatGPT or Claude be used in GMP processes?
The draft excludes generative AI and LLMs from critical GMP applications. For non-critical support, such as drafts, research or evaluation suggestions with human review, this does not apply in the same way. Since 2026, the EMA has been examining under which controls generative AI could be permitted more broadly.
When will the final Annex 22 arrive?
There is no fixed date. As of 2026, the draft is being revised and further targeted consultations are scheduled.
What are guardrails?
Control and mitigation measures that limit and secure the use of an AI system, for example fixed limits on its scope, checks of its outputs or a human approval step.
Next step
Find out where generative AI is already in use in your lab, including unofficially, with the Annex 22 Readiness Check.
Sources
- EMA, multistakeholder workshop on Annex 22 (30 June and 1 July 2026): ema.europa.eu
- HPRA, presentation at the QP Forum 2026: tcd.ie (PDF)
- IntuitionLabs, EU GMP Annex 22 and AI compliance: intuitionlabs.ai
- Regulation (EU) 2024/1689 (AI Act), Art. 4 and Art. 113: eur-lex.europa.eu
Annex 22 is available as a draft (as of 5 October 2026). This content is not legal, regulatory or compliance advice.