A scientist who assumes there is one authoritative FDA AI drug development guidance document to read would be forgiven, and wrong. The reality is closer to a paper trail: a 2021 action plan that most drug scientists have never actually read because it was written for medical devices, a pair of drug-specific discussion papers that quietly did the real groundwork, and a January 2025 draft guidance that is still, as of this writing, in draft. Artificial intelligence (AI) touches every one of these documents differently, and knowing which one actually governs a given use case—discovery, nonclinical, clinical, manufacturing—is most of the battle.

This works through that paper trail in order. It connects to the broader hub on regulating AI in drug discovery for how FDA's approach compares with EMA's and ICH's, and to the guide to AI across drug discovery for the wider scientific and technical picture.

FDA's 2021 AI/ML action plan

Here is the detail that trips people up: the document most commonly cited as "FDA's 2021 AI/ML action plan" is the Artificial Intelligence/Machine Learning-Based Software as a Medical Device Action Plan, published by the Center for Devices and Radiological Health (CDRH) in January 2021. It is a real, foundational document, but it is a device document: it addresses diagnostic algorithms, imaging software, and similar tools regulated as Software as a Medical Device, not the drugs and biologics that are CDER's and CBER's remit. A drug discovery scientist reading it looking for guidance on, say, a generative chemistry model, will not find their use case addressed.

What is genuinely relevant from that same period is Good Machine Learning Practice (GMLP), a set of ten guiding principles FDA developed jointly with Health Canada and the UK's Medicines and Healthcare products Regulatory Agency in October 2021. GMLP was also framed primarily around medical devices, but its principles—multidisciplinary expertise, representative data, robust engineering practice, human factors consideration, performance monitoring—were general enough that they show up again, largely unchanged in spirit, in FDA's later drug-specific work and in the FDA-EMA joint principles published in January 2026. CDER's own parallel activity in 2021 was narrower and manufacturing-specific: the Framework for Regulatory Advanced Manufacturing Evaluation (FRAME) initiative, which had AI/ML in pharmaceutical manufacturing as one strand among several advanced-manufacturing priorities.

Subsequent draft guidances

CDER's own drug-specific engagement with AI really begins in 2022 and 2023, and moves faster than most people outside regulatory affairs track. A discussion paper on distributed and point-of-care drug manufacturing appeared in October 2022; a dedicated "Artificial Intelligence in Drug Manufacturing" discussion paper, part of the FRAME initiative, followed in March 2023, narrower in scope but a direct predecessor to the broader work that came after.

The document that actually opened the drug-specific conversation broadly, rather than just for manufacturing, is the May 2023 discussion paper, "Using Artificial Intelligence & Machine Learning in the Development of Drug & Biological Products." It walks through AI use cases across the entire pipeline—drug discovery, nonclinical research, clinical research, postmarket safety, manufacturing—and drew more than 800 public comments, which FDA has said directly informed the structure of the January 2025 draft guidance. The two centers most affected, CDER and CBER, along with CDRH and the Office of Combination Products, published a coordination paper titled "Artificial Intelligence and Medical Products: How CBER, CDER, CDRH, and OCP Are Working Together" in March 2024, revised in February 2025, which is worth reading mainly for what it reveals about how fragmented AI oversight still is inside the agency itself, four centers coordinating rather than one center owning the topic.

The January 2025 draft guidance is the document doing the most substantive work today, introducing the seven-step, risk-based credibility framework covered in detail in the hub article on AI regulation. It remains in draft as of this writing, more than a year after its comment period closed in April 2025.

How FDA views AI-generated evidence

FDA has been reviewing submissions with AI components for longer than the existence of dedicated guidance might suggest: the agency has cited more than 500 drug and biologic applications with AI components between 2016 and 2023, concentrated heavily in oncology and neurology. One peer-reviewed analysis of that submission trend found the pace accelerating sharply, from roughly 29 applications with AI/ML components in 2019 to more than 100 in 2021 alone, though the underlying counting methodology varies enough between analyses that the exact figures should be read as directional rather than precise.

The core idea the January 2025 draft guidance introduces is that AI-generated evidence is not evaluated on a single, fixed bar. It is evaluated against the model's declared context of use: the same underlying algorithm might need only lightweight internal validation if it is flagging manufacturing anomalies for a human to review, and a full, independently reviewed credibility assessment if its output is being used to support a claim about clinical safety or efficacy. In practice, this means an IND or NDA submission that relies on AI-generated analysis is expected to include, or have available on request, a credibility assessment report explaining what the model was for, how its risk was assessed, and what evidence supports trusting its output for that specific purpose, scaled to how much weight that output actually carries in the regulatory decision.

Model documentation and explainability

Documentation expectations run in a fairly consistent direction across every FDA AI document published so far: data provenance, model development choices, and validation results need to be traceable well enough that someone outside the team that built the model, an FDA reviewer or inspector, could reasonably assess whether it does what it claims to do. That is a materially higher bar than internal validation practices that many discovery-stage AI tools were built against, particularly generative or exploratory models never originally intended to support a regulatory claim.

Explainability and interpretability show up as related but distinct expectations, not interchangeable ones. An interpretable model is one whose internal logic a human can follow directly; an explainable model may remain a black box internally but comes paired with a separate method for characterizing why it produced a given output. FDA's language leaves room for both, but consistently ties the acceptable level of either one back to risk: a low-risk, human-reviewed screening tool can likely rely on relatively limited explainability, while a model whose output more directly drives a safety or efficacy conclusion is going to draw a harder look at how, and how well, its behavior can be characterized.

The predetermined change control plan

One mechanism worth understanding precisely, because it is frequently misapplied to drugs in conversation, is the predetermined change control plan, or PCCP. A PCCP lets a sponsor pre-specify, at the time of an initial submission, the kinds of updates an AI model is expected to need over time, and the validation methodology that will be used to implement them, so that later modifications within that pre-agreed scope do not require a brand-new submission each time. It is a genuinely useful answer to a real problem: AI models are rarely static, and traditional regulatory pathways assume a product does not change quietly underneath its approval.

The catch is scope. PCCP, as it currently exists, was built under statutory authority added by the 2022 Food and Drug Omnibus Reform Act specifically for AI-enabled device software functions, and lives under CDRH, not CDER. There is no equivalent, formally named PCCP pathway yet for drugs and biologics. The January 2025 CDER draft guidance references the device-side PCCP guidance only as related context, not as an available mechanism for drug submissions. The life cycle management principle in the joint FDA-EMA guiding principles gestures at similar thinking: periodic re-evaluation, planned monitoring for data drift, being applied to drugs eventually, but a sponsor working with an AI model that supports a drug or biologic claim should not assume a device-style PCCP is available to them today.

Practical implications

For a team building or using AI somewhere in a drug discovery or development program, the practical reading of all of this is less about waiting for the January 2025 draft to finalize and more about building toward its logic now. Regulatory affairs groups and the CROs that support them are already doing exactly that: Parexel brought on a dedicated Vice President of Consulting for AI and Digital Policy in mid-2025, a hire explicitly framed around helping sponsors navigate this exact landscape, while ICON has published its own analysis of what FDA's AI/ML discussion paper means in practice for sponsors already using AI-processed digital health technology data in trials.

• Write a context-of-use statement for any AI model whose output could plausibly end up in a regulatory submission, well before that submission is drafted.
• Keep data provenance and validation records as though an outside reviewer will eventually ask for them, since that is close to the explicit expectation for high-impact uses.
• Do not assume a discovery-stage tool is out of scope indefinitely; FDA's current draft excludes discovery today, but EMA's parallel framework already does not, and the two agencies are visibly converging.
• Do not plan around a drug-side PCCP pathway that does not yet exist; build a life cycle monitoring plan on the assumption that any formal change-control mechanism, if one arrives, will look more like the device-side model than something invented from scratch.

None of FDA's AI documents, taken individually, is the whole story. Taken together, chronologically, they describe an agency that has been quietly reviewing AI-influenced submissions for the better part of a decade and is only now, in fits and starts across four different centers, writing down what it actually expects.

This article was produced in accordance with Drug Discovery News’ Editorial Policies.