FDA and EMA Joint AI Principles: What they mean for manufacturing
Siobhán O'Leary
Status as at 2 September 2026. The joint principles are non-binding. The FDA's guidance on AI for regulatory decision-making remains in draft.
What happened
In January 2026, the FDA and the European Medicines Agency jointly published 10 Guiding Principles of Good AI Practice in Drug Development. It is the first time both regulators have publicly aligned on AI governance for medicines and biologics, and the scope runs the full lifecycle: early research, clinical trials, manufacturing, and post-market safety monitoring.
The principles are non-binding. Both agencies have been clear that they reflect expectations that already exist rather than creating new ones.
Alongside them sits the FDA's January 2025 draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, which sets out a risk-based credibility assessment framework for AI models used to generate evidence in a regulatory submission. That guidance is still marked as draft and not for implementation. The framework in it is, however, the most detailed statement available of how the FDA thinks about AI evidence, and it is what organsiatons can prepare against.
The distinction that matters most
The principles are technology agnostic. They apply equally to traditional machine learning and to generative AI, and that is precisely why many teams read them and conclude nothing applies to them.
Most pharma operations are already running traditional machine learning, for example, predictive maintenance, process optimisation, quality anomaly detection, demand forecasting and batch deviation prediction are all in this category. Meanwhile the visible activity, the part attracting attention and budget, is generative AI: drafting SOPs from process notes, summarising regulatory change across jurisdictions, drafting supplier and deviation correspondence, building role-specific training content.
Governance applies to both, however the failure modes are different.
Why this matters. Traditional models fail through data drift, degradation and bias in training data. Those failures are measurable, monitorable and familiar to anyone who has run equipment qualification. Generative models fail through hallucination, inconsistency and confident wrongness. Those failures are caught by a person who knows the subject matter and is expected to check.Most are written for the AI that arrived last year, not the AI that has been running for a decade.
The 10 principles, read through an operations lens
| # | Principle | What it asks of manufacturing and supply chain |
|---|---|---|
| 1 | Human-centric design | AI supporting process decisions keeps patient safety and human oversight central. Training comes before adoption, not after. |
| 2 | Risk-based approach | Match validation and oversight to criticality. Predictive maintenance and AI supporting batch release are not the same risk. |
| 3 | Adherence to standards | AI use sits inside GMP, GCP and cybersecurity obligations, alongside emerging AI-specific texts including draft EU GMP Annex 22 and the ISPE GAMP Guide: Artificial Intelligence (2025). |
| 4 | Clear context of use | Define each tool's purpose, scope, data inputs and outputs, and how results feed a decision. Written down, not understood informally. |
| 5 | Multidisciplinary expertise | Operations, quality, data science, IT security and regulatory together. No single function owns this. |
| 6 | Data governance | Traceability of data sources and processing steps, documented to a standard a reviewer could follow. |
| 7 | Model design | Interpretability and robustness treated as design requirements, particularly where a model informs specification setting. |
| 8 | Performance assessment | Validate the whole system, including how people actually interact with outputs in the workflow. |
| 9 | Lifecycle management | Ongoing monitoring for drift and periodic re-evaluation. The logic mirrors equipment qualification and process validation. |
| 10 | Clear communication | Purpose, performance and limitations explained in plain language across the organisation. |
The obligations map onto validation, change control and documentation practice that already exists. The gap is rarely validation capability, it is AI literacy and clear ownership of the decision.
Enforcement is no longer theoretical
In April 2026 the FDA issued a warning letter containing a section headed "Inappropriate Use of Artificial Intelligence in Pharmaceutical Manufacturing" the first time AI misuse has appeared as a named, standalone deficiency. The firm had used AI agents to create specifications, procedures and production records without adequate quality unit review. The FDA did not object to the use of AI, it stated that an output from an AI agent must be reviewed and cleared by an authorised human representative of the quality unit.
Why this matters. The principles are non-binding. The expectation underneath them is already being enforced. We look at that case, and at what EU inspectors will be asking, in our article on draftEU GMP Annex 22.
Three questions to take to your next quality review
- Which AI systems do we already run that we have not classified as AI, and are any of them impacting product quality, patient safety or data integrity?
- For each one, who is the named person accountable for the output, and could they tell you what the system is weak at?
- If an inspector asked how we assured an AI-generated document before it entered our quality system, what would we point to?
If the answers are uncomfortable, that is useful information. The organisations that come out of this well are not the ones with the most AI, they are the ones who can show their working.
The Institute of Applied AI helps life sciences organisations build AI capability, assign clear governance to AI decisions, and turn ambition into roadmaps that survive inspection. If you would like to discuss any of the above, get in touch.
Sources: FDA and EMA, Guiding Principles of Good AI Practice in Drug Development (January 2026), FDA, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft guidance (January 2025), FDA CDER, Artificial Intelligence for Drug Development, FDA Warning Letter, Purolea Cosmetics Lab, 2 April 2026, Draft EU GMP Annex 22 (consultation 7 July to 7 October 2025), ISPE GAMP Guide: Artificial Intelligence (2025)
This article is not written as a compliance position or legal advice. Draft guidance may change before it is finalised.
