AIforPharma.org

Introduction

⇒ FDA's current regulatory-science framework identifies AI as a cross-lifecycle priority, including commercial drug manufacturing, quality decision-making, regulatory review, postmarket surveillance and inspector capability-building.

Artificial intelligence is now positioned by FDA as a regulatory science topic spanning essentially the entire lifecycle of regulated medical products.

In its current Focus Areas of Regulatory Science framework, FDA identifies artificial intelligence as relevant to:

• product development;
• commercial manufacturing;
• regulatory assessment;
• postmarket surveillance;
• inspectional activities.

For pharmaceutical companies, one point is particularly important.

FDA explicitly states that it is:

exploring how AI is used during commercial drug manufacturing to improve quality decision-making.

This places AI directly within the pharmaceutical-quality and manufacturing regulatory-science agenda rather than treating it solely as a clinical, digital-health or administrative technology.

AI Is Being Considered Across the Entire Product Lifecycle

FDA’s regulatory science framework shows AI as applicable across:

• product characterisation, manufacturing and quality;
• nonclinical premarket evaluation;
• clinical premarket evaluation;
• postmarket activities.

This broad scope matters.

Pharmaceutical AI governance is sometimes organised as if different applications belong to completely separate domains:

R&D AI

Clinical AI

Manufacturing AI

Quality AI

Regulatory AI

FDA’s regulatory-science perspective increasingly suggests that these should instead be seen as different Contexts of Use within one broader AI lifecycle.

The technology may differ, but recurring regulatory questions remain similar:

• Are the data appropriate?
• Is the model sufficiently reliable?
• Is performance robust to changing conditions?
• Can the output be understood and verified?
• How are changes controlled?
• Who remains accountable?

Commercial Drug Manufacturing Is Explicitly Included

The direct reference to commercial drug manufacturing is particularly relevant for GMP professionals.

Potential AI applications in manufacturing can include:

• process monitoring;
• predictive process models;
• advanced process control;
• deviation and trend analysis;
• predictive maintenance;
• visual inspection;
• Quality System analytics;
• knowledge management.

FDA does not state that all of these uses are automatically acceptable.

The important signal is that the Agency is actively studying how AI can support manufacturing quality decisions.

This means that AI in manufacturing is increasingly becoming a regulatory-science problem rather than merely an industry innovation project.

Robustness to Changing Conditions Is Becoming a Regulatory Question

Another particularly important FDA research activity concerns model robustness.

FDA describes work aimed at understanding how machine-learning models respond to:

• changes in real-world data;
• changing contextual factors;
• distributional shifts;
• different intended-use conditions.

This issue is fundamental for pharmaceutical AI.

A model may perform well during development and validation but later encounter:

• different raw-material characteristics;
• new equipment;
• process changes;
• new sites;
• different populations;
• changing documentation patterns;
• different operating conditions.

The key regulatory question therefore becomes:

Does the model remain sufficiently reliable when its operating environment changes?

This is closely related to the concept of maintaining AI in a continued state of control.

Validation Cannot Stop at Initial Accuracy

For conventional software, validation commonly demonstrates that predefined functions operate as expected.

AI requires an additional question:

Will the model continue to perform adequately when the data distribution or context changes?

This can require controls such as:

• ongoing performance monitoring;
• drift detection;
• periodic reassessment;
• defined performance limits;
• change-control triggers;
• revalidation criteria.

A possible lifecycle therefore becomes:

Initial Validation

↓

Controlled Deployment

↓

Operational Performance Monitoring

↓

Detection of Context or Data Change

↓

Impact Assessment

↓

Revalidation or Model Update Where Necessary

FDA Is Also Studying Synthetic Data

FDA’s current AI research activities include evaluating synthetic datasets used for model training and testing.

This is another important area for pharmaceutical companies.

Synthetic data may potentially support:

• model development where real data are limited;
• privacy protection;
• simulation of rare events;
• test-set expansion;
• model robustness assessment.

But synthetic data create their own credibility questions.

For example:

• How representative are they?
• What assumptions were used to generate them?
• Could important edge cases be missing?
• Could bias in the source data be amplified?
• Can the generation process be reconstructed?

Synthetic data therefore do not eliminate the need for data governance.

They create an additional data-generation process that itself may need to be understood and controlled.

AI Is Also Entering Regulatory Review and Pharmacovigilance

FDA describes several internal or research applications involving AI.

These include:

• natural-language processing of regulatory submissions;
• classification of submission complexity;
• analysis of clinical-trial data;
• processing and evaluation of safety reports;
• MedDRA coding support;
• pharmacometrics;
• precision-medicine applications.

This is strategically important for pharmaceutical companies.

AI is increasingly appearing on both sides of the regulatory interface.

Companies may use AI to:

generate → analyse → organise → submit regulatory information

while regulators may use AI to:

retrieve → classify → analyse → review regulatory information.

This increases the value of structured, traceable and machine-readable data.

FDA Is Building AI Knowledge Among Inspectors

One of the most interesting points from a GMP perspective is that FDA describes workshops designed to introduce AI and machine learning to Agency personnel.

FDA specifically notes that this helps keep its inspectional cadre informed about how these technologies may affect investigations.

This is a meaningful regulatory signal.

AI governance is gradually becoming relevant not only to:

• regulatory policy;
• scientific review;
• guidance development;

but also to:

inspection capability.

For pharmaceutical companies, this means AI could increasingly become part of ordinary GMP inspection discussions whenever it materially affects a regulated process.

What Could an Inspector Eventually Ask?

A future inspection involving an AI-enabled quality or manufacturing process could reasonably explore questions such as:

• Where is AI used?
• What is the intended use?
• What GMP decision does it influence?
• Which data are used?
• How was performance evaluated?
• What are the known failure modes?
• How are model and configuration changes controlled?
• How is ongoing performance monitored?
• What happens if the AI becomes unreliable?
• Who is responsible for the final quality decision?

These are not presented by FDA as a new inspection checklist.

They are logical consequences of treating AI as part of the regulated pharmaceutical process.

A Broader AI Assurance Model Is Emerging

FDA’s regulatory-science priorities reinforce a pattern increasingly visible across pharmaceutical AI regulation.

The assurance model is moving beyond:

Validate the model.

towards:

Intended Use → Data → Performance → Robustness → Human Oversight → Change Control → Monitoring → Accountability

This is significantly closer to lifecycle Quality Risk Management than to conventional one-time software testing.

Why This Is Important

The significance of FDA’s current regulatory-science position is that AI is no longer treated as a narrow emerging technology.

It is explicitly connected with:

• commercial drug manufacturing;
• quality decision-making;
• regulatory assessment;
• postmarket surveillance;
• inspection capability.

For pharmaceutical companies, this suggests that AI governance should increasingly become part of the normal Pharmaceutical Quality System and regulatory-science strategy.

The relevant question is becoming less:

“Do we have an AI policy?”

and more:

“Can we demonstrate that every important AI use remains appropriate, reliable and controlled for its specific regulated purpose?”

Regulatory Status Note

FDA’s Focus Areas of Regulatory Science describe research priorities and ongoing regulatory-science activities.

They are not guidance and do not establish new legally binding CGMP requirements.

FDA also explicitly notes that inclusion of research in the Focus Areas of Regulatory Science does not necessarily imply a determination regarding its future regulatory use.

Nevertheless, the framework provides an important indication of the scientific questions FDA considers relevant to future regulation.

Sources

FDA – Focus Area: Artificial Intelligence:

https://www.fda.gov/ … ificial-intelligence

FDA – Focus Areas of Regulatory Science Report:

https://www.fda.gov/ … atory-science-report

FDA – Lifecycle of FDA-Regulated Products:

https://www.fda.gov/ … a-regulated-products

Introduction

⇒ The proposed European Biotech Act could make AI a formal part of pharmaceutical regulatory infrastructure through EMA lifecycle guidance, trusted AI testing environments and data-quality accelerators supporting model training, validation and testing.

Artificial intelligence may soon become more deeply embedded in the European pharmaceutical regulatory framework.

The proposed European Biotech Act contains a dedicated chapter on artificial intelligence and data as biotechnology enablers.

Its significance goes beyond another set of general responsible-AI principles.

The proposal foresees three concrete elements with direct relevance to pharmaceutical companies:

[]EMA guidance on AI across the medicinal-product lifecycle;
[]trusted testing environments for AI-enabled biotechnology;
• data-quality accelerators supporting AI training, testing and validation.

The proposal is still moving through the EU legislative process and is not yet final law.

However, discussions continued within the Council Working Party on Pharmaceuticals and Medical Devices in September 2026, while EMA reported on 1 October that preparations for implementation of the Biotech Act are already influencing its future regulatory infrastructure.

For pharmaceutical companies, this is an important signal:

AI governance in Europe may evolve from individual guidance documents toward a coordinated regulatory and technical infrastructure for developing, validating and using AI across the medicines lifecycle.

EMA Could Receive a Broader AI Guidance Role

The proposed legislation would require EMA to develop and publish guidance, in cooperation with the European Commission and national competent authorities, on the deployment and use of advanced technologies including artificial intelligence.

The intended scope is broad.

It covers AI across areas such as:

[]medicinal-product development;
[]manufacturing;
[]clinical trials;
[]regulatory evaluation and authorisation;
• post-authorisation activities.

This is important because current pharmaceutical AI guidance remains distributed across multiple documents and regulatory initiatives.

Examples include:

[]the EMA reflection paper on AI in the medicinal-product lifecycle;
[]FDA/EMA Guiding Principles of Good AI Practice;
[]draft EU GMP Annex 22;
[]planned AI guidance for pharmacovigilance;
• digital technology qualification guidance.

The Biotech Act could provide a legal framework for bringing these activities into a more coherent lifecycle approach.

AI Guidance Would Not Be Limited to One Technology

The proposal refers not only to conventional machine-learning systems but also to:

AI systems and general-purpose AI models.

This distinction is increasingly important.

Pharmaceutical companies are moving beyond narrowly trained predictive models toward:

[]foundation models;
[]large language models;
[]Generative AI;
[]multimodal systems;
[]AI agents;
[]hybrid AI architectures.

A future EMA guidance framework may therefore need to address very different technological characteristics and risk profiles.

The regulatory question becomes less:

“Is this an AI system?”

and more:

“What does this AI do, what decision does it influence, and what level of assurance is required for that Context of Use?”

Trusted AI Testing Environments Could Be More Important Than Another Guidance Document

One of the most interesting elements of the proposed Biotech Act is the concept of trusted testing environments for advanced biotechnology innovations.

These environments would support technologies that are enabled or significantly enhanced by AI or advanced computational methods.

Their purpose would be to provide controlled environments where organisations can develop, test and validate innovations while operating under applicable regulatory requirements.

Potential applications identified in the proposal include advanced biotechnology areas such as:

[]immunology;
[]ATMP and gene-therapy development;
[]advanced computational methods;
[]New Approach Methodologies;
• AI-supported biotechnology and biomanufacturing.

This creates a potentially important model:

Develop → Experiment → Validate → Generate Evidence → Inform Standards and Regulation

under controlled conditions.

For pharmaceutical AI, this could be particularly valuable because many current barriers are not purely technical.

Companies also need to understand what evidence regulators will consider sufficiently credible.

A Possible Bridge Between Innovation and Regulatory Acceptance

AI development often moves faster than formal pharmaceutical guidance.

This creates a familiar problem.

A company may develop a technically successful AI system but still face uncertainty regarding:

[]regulatory acceptability;
[]validation expectations;
[]documentation requirements;
[]acceptable performance criteria;
[]change management;
[]human oversight;
• data requirements.

Trusted testing environments could create a more structured way of addressing these questions before large-scale regulatory deployment.

Instead of waiting until a marketing application, manufacturing change or regulatory submission is ready, developers could potentially generate regulatory learning earlier.

Conceptually, this creates a bridge between:

innovation experimentation

and

regulated pharmaceutical implementation.

Data Quality Accelerators Are the Third Pillar

The proposal also recognises a problem that is becoming increasingly obvious:

AI performance depends fundamentally on the quality of the data used to train, test and operate the system.

The proposed biotechnology data-quality accelerators would support the development of datasets that are:

[]high quality;
[]appropriately annotated;
[]standardised;
[]curated;
[]provenance verified;
[]suitable for AI training, validation and testing.

This may become particularly important in pharmaceutical environments where data are distributed across:

[]laboratory systems;
[]clinical systems;
[]manufacturing systems;
[]Quality Management Systems;
[]regulatory repositories;
[]real-world data sources.

The challenge is often not lack of data.

The challenge is whether those data are sufficiently structured, controlled and contextualised to support a regulated AI application.

Data Provenance Could Become Part of AI Validation

Traditional computerized-system validation often focuses strongly on application functionality.

AI adds another layer.

The credibility of the output can depend on the entire data chain:

Source Data

↓

Curation and Transformation

↓

Training / Validation Dataset

↓

AI Model

↓

AI Output

↓

Regulated Decision

This means that validation of an AI-enabled pharmaceutical process may increasingly require evidence not only about the model, but also about:

[]where the data originated;
[]how the data were transformed;
[]which version was used;
[]whether important populations or conditions were represented;
• whether data quality remains suitable during operation.

The proposed emphasis on provenance-verified datasets is therefore particularly relevant for GxP environments.

The Biotech Act Could Help Connect the AI Act with Pharmaceutical Regulation

Another important aspect is the relationship with the EU AI Act.

The Biotech Act proposal foresees interaction between pharmaceutical AI testing environments and AI testing or regulatory mechanisms established under Regulation (EU) 2024/1689.

This is strategically important.

Pharmaceutical companies may need to manage two regulatory dimensions simultaneously:

horizontal AI regulation

and

sector-specific pharmaceutical regulation.

These frameworks ask different questions.

The AI Act focuses on the governance and risks of AI systems.

Pharmaceutical regulation focuses on:

[]patient safety;
[]product quality;
[]clinical evidence;
[]manufacturing control;
• regulatory decision-making.

The challenge will be to avoid creating two disconnected assurance systems for the same AI application.

Relevance to GMP Manufacturing

The proposal explicitly includes manufacturing within the future lifecycle AI-guidance framework.

For GMP organisations, this is particularly important.

AI applications may increasingly support:

[]process monitoring;
[]predictive process modelling;
[]deviation analysis;
[]maintenance;
[]Quality System analytics;
[]advanced process control;
• manufacturing knowledge management.

Draft Annex 22 is currently the most visible European GMP-specific AI initiative.

The Biotech Act suggests that Annex 22 may eventually sit inside a broader European AI regulatory architecture rather than functioning as an isolated GMP document.

The Regulatory Model Is Becoming Multi-Layered

A future European pharmaceutical AI framework could increasingly consist of several layers:

EU AI Act

↓

European Biotech Act

↓

EMA Lifecycle AI Guidance

↓

Domain-Specific Guidance such as GMP Annex 22

↓

Testing Environments and Regulatory Sandboxes

↓

Company AI Governance and Validation

For global pharmaceutical companies, understanding how these layers interact may become as important as understanding each document individually.

Current Regulatory Status

The European Biotech Act remains a legislative proposal.

It is currently being considered by the European Parliament and Council and its final wording may change.

Council working-party discussions on the proposal continued during September 2026.

EMA has also confirmed that it is preparing for potential implementation of the legislation while progressing its broader data and artificial-intelligence activities.

Therefore, pharmaceutical companies should not treat the proposed AI provisions as current binding requirements.

They should instead be viewed as a strong indication of the future direction of European pharmaceutical AI regulation.

Why This Matters

Most pharmaceutical AI regulation to date has focused on principles and guidance.

The Biotech Act introduces something different:

regulatory infrastructure.

The emerging model is not only:

write guidance → ask industry to comply.

It increasingly looks like:

Guidance → Trusted Testing → High-Quality Data → Regulatory Learning → Validation → Deployment.

If implemented, this could significantly influence how pharmaceutical companies develop and validate AI systems in Europe.

It also reinforces a broader message:

Trustworthy pharmaceutical AI will depend not only on the algorithm, but on the regulatory, data and testing infrastructure surrounding it.

Sources

European Commission – European Biotech Act:

https://health.ec.eu … opean-biotech-act_en

EUR-Lex – Proposal for the European Biotech Act:

https://eur-lex.euro … ri=CELEX:52025PC1022

Council of the European Union – Working Party on Pharmaceuticals and Medical Devices, 16 September 2026:

https://www.consiliu … evices-%28369737%29/

EMA – Management Board highlights, September 2026 meeting:

https://www.ema.euro … ptember-2026-meeting

Introduction

⇒ WHO is calling for stronger ethics oversight of AI-related health research, highlighting risks involving bias, transparency, privacy, accountability and data governance across the entire research lifecycle.

The World Health Organization is calling for stronger oversight of artificial intelligence used in health-related research.

On 21 September 2026, WHO highlighted and formally launched its report:

Artificial Intelligence-related Health Research: Ethics Review and Oversight.

The report addresses an increasingly important problem.

Artificial intelligence is already being used to analyse health data, support research, generate hypotheses and develop new technologies.

However, traditional research-ethics frameworks were largely created before modern AI systems became widespread.

WHO warns that existing oversight mechanisms may therefore not fully address risks associated with:

• bias;
• lack of transparency;
• privacy;
• fairness;
• accountability;
• rapid deployment of AI tools.

For pharmaceutical companies, this is highly relevant wherever AI is used in clinical research, health-data analysis or development of evidence that may ultimately support regulatory decisions.

WHO Identifies Three Types of AI-Related Health Research

The report distinguishes three broad categories:

• health-related data science using AI;
• research conducted with AI tools and technologies;
• health-related research on AI tools and technologies.

This distinction is useful because the ethical risks differ depending on how AI is used.

For example, there is a difference between:

using AI to analyse an existing clinical dataset

and

conducting research to evaluate a new AI system that may later affect patient care.

The governance approach should therefore reflect the actual role of AI in the research.

Research Ethics Committees May Need New Capabilities

WHO emphasises that Research Ethics Committees remain central to protecting research participants.

However, AI creates technical and governance questions that conventional ethics committees may not always be equipped to evaluate.

These can include:

• how the model was developed;
• which data were used;
• whether the data are representative;
• how bias is assessed;
• whether outputs are explainable;
• how privacy is protected;
• how model changes are managed.

WHO therefore identifies the need for additional expertise, training and resources.

This is an important message for pharmaceutical sponsors.

An AI-enabled research protocol may require a different level of technical explanation than a conventional protocol.

Ethical Risk Should Be Addressed Early

Another important principle is that ethical review should not become a final administrative check performed after the AI system has already been designed.

Researchers should identify AI-specific risks early.

This may include considering:

• data provenance;
• informed consent;
• privacy;
• population representativeness;
• bias;
• possible downstream harms;
• transparency;
• limitations of the model.

This creates a useful lifecycle sequence:

Research Question → Data Strategy → AI Design → Ethical Risk Assessment → Study Execution → Monitoring → Publication / Implementation

Ethical considerations therefore become part of AI design rather than an external review performed only at the end.

Data Governance Is Central

For pharmaceutical research, one of the most important implications concerns data governance.

AI may depend on very large datasets.

These datasets may come from:

• clinical trials;
• electronic health records;
• registries;
• real-world data;
• biobanks;
• genomic databases.

The fact that data are technically accessible does not automatically mean that every AI use is ethically appropriate.

Relevant questions include:

• Was the original consent compatible with the new use?
• Can individuals or groups be re-identified?
• Are some populations under-represented?
• Could the model create systematic disadvantage?
• Who benefits from the resulting AI system?

For pharmaceutical companies, these questions increasingly intersect with data protection, clinical governance and regulatory evidence generation.

Bias Is Not Only a Technical Performance Problem

Bias is often discussed as an AI-performance issue.

WHO treats it also as an ethical issue.

A model can appear accurate overall while performing poorly for particular groups.

This may occur because:

• training data are unbalanced;
• certain populations are under-represented;
• historical data contain structural bias;
• the model is deployed outside its original Context of Use.

For pharmaceutical development, this can be particularly important where AI influences:

• patient selection;
• risk stratification;
• endpoint analysis;
• safety assessment;
• evidence generation.

A single average performance metric may therefore be insufficient.

Oversight Cannot Be Left to the Ethics Committee Alone

WHO also makes clear that responsible AI research requires more than Research Ethics Committees.

Other important actors include:

• research funders;
• data-access bodies;
• scientific journals;
• professional societies;
• regulatory agencies.

This is significant because some AI risks are systemic.

They cannot be solved by reviewing each individual study independently.

For example, common problems involving data access, model transparency or reproducibility may require broader institutional or regulatory solutions.

Why This Matters for Pharmaceutical R&D

The report is not pharmaceutical GMP guidance.

However, it is directly relevant to pharmaceutical R&D organisations increasingly using AI in:

• clinical-development analytics;
• real-world evidence;
• biomarker research;
• patient stratification;
• safety analysis;
• computational biology.

As AI-generated analysis becomes more important to evidence generation, companies will need to demonstrate not only that models perform technically, but also that their use is ethically justified and appropriately governed.

The assurance model therefore becomes broader:

Scientific Validity + Data Governance + Ethical Oversight + Transparency + Human Accountability

From Technical Validation to Responsible Research Governance

The WHO report reinforces a wider trend in AI regulation.

AI assurance is moving beyond technical validation alone.

A model may be technically accurate but still create problems if:

• the data were obtained or reused inappropriately;
• important groups are under-represented;
• the model’s limitations are hidden;
• responsibility for decisions is unclear;
• participants are exposed to risks they did not reasonably anticipate.

Responsible AI therefore requires a broader governance system.

Why This Development Is Important

Pharmaceutical companies are investing heavily in AI-enabled research.

The WHO report provides an important reminder that faster analysis does not reduce ethical responsibility.

As AI becomes more deeply integrated into health research, the governance question is shifting from:

“Can AI perform this analysis?”

towards:

“Is this AI-supported research scientifically valid, ethically acceptable, transparent and appropriately governed?”

That distinction is likely to become increasingly important as AI-generated evidence moves closer to regulatory and clinical decision-making.

Regulatory Status Note

The WHO report is guidance on ethics review and oversight of AI-related health research.

It is not GMP guidance and does not create new legally binding pharmaceutical requirements.

Its relevance to pharmaceutical companies is primarily in clinical research, health-data science, AI-supported evidence generation and research governance.

WHO describes the publication as a starting point for strengthening future standards, guidance and governance approaches in AI-related health research.

Sources

WHO – New report calls for stronger ethics oversight of AI-related health research, 21 September 2026:

https://www.who.int/ … ated-health-research

WHO – Artificial Intelligence-related Health Research: Ethics Review and Oversight:

https://www.who.int/ … i/item/9789240124073