FDA AI Guidance: Drugs & Medical Devices

“FDA AI guidance” is not one document. It is shorthand for a small, actively growing family of FDA guidance documents that address artificial intelligence and machine learning from two largely separate directions inside the agency: the medical-device side, led by the Center for Devices and Radiological Health (CDRH), which regulates AI as embedded software functions in a device; and the drug/biologic side, led jointly by the Center for Drug Evaluation and Research (CDER) and the Center for Biologics Evaluation and Research (CBER), which addresses AI as a tool used to generate or analyze evidence that supports a regulatory decision, rather than as the product itself. A sponsor or research administrator working across both a device and a drug program may need to track both tracks simultaneously, because they use different vocabulary, different review pathways, and are at different stages of finality.

This guide summarizes both tracks as they stand as of this writing, with the key documents, dates, and open questions a research administrator needs to track. Because this is one of the fastest-moving areas of FDA policy, treat every status below as a snapshot rather than a permanent description, and confirm current status directly against FDA’s AI in SaMD page and CDER’s AI in drug development page before relying on it for a submission decision.

Two tracks, not one framework

It helps to keep the two tracks conceptually separate from the start:

CDRH / medical devices: governs AI when it is part of the regulated product itself — an AI-enabled diagnostic algorithm, an AI-driven imaging or triage tool, or any Software as a Medical Device (SaMD) that uses machine learning. The core regulatory question is how a device that is designed to learn and change after it reaches the market can still be safely and effectively reviewed and marketed.
CDER/CBER / drug and biological products: governs AI when it is used as a tool in the drug or biologic development and review process — for example, an AI model used to help select trial endpoints, flag adverse events, support a manufacturing decision, or otherwise generate information that will factor into a regulatory submission. Here the product being reviewed is still a drug or biologic; the AI model itself is treated as something whose “credibility” for that specific use has to be established.

Both tracks sit under the umbrella of FDA’s broader Artificial Intelligence and Medical Products initiative, and the agency has said the two centers coordinate, but as of this writing they have published separate guidance documents with separate scopes, terminology, and timelines rather than a single unified rulebook.

CDRH’s framework for AI in medical devices (SaMD)

CDRH’s approach to AI/ML-based Software as a Medical Device has developed incrementally since 2019, through a mix of discussion papers, joint international “guiding principles,” and formal guidance documents:

2019 — discussion paper. FDA published a discussion paper proposing a regulatory framework for modifications to AI/ML-based SaMD, introducing the idea of a “predetermined change control plan” as a way to handle devices designed to be updated after clearance.
January 2021 — AI/ML-Based SaMD Action Plan. FDA laid out its planned next steps, including a proposed framework for a Total Product Lifecycle (TPLC) approach to AI-enabled devices.
October 27, 2021 — Good Machine Learning Practice (GMLP), 10 guiding principles. FDA, Health Canada, and the UK’s Medicines and Healthcare products Regulatory Agency (MHRA) jointly issued ten guiding principles intended to inform the development of Good Machine Learning Practice — a machine-learning analogue to Good Manufacturing Practice or Good Clinical Practice, covering areas such as representative training data, appropriate human oversight, and performance monitoring across the product lifecycle.
April 2023 — draft guidance on Predetermined Change Control Plans (PCCPs). FDA proposed how manufacturers could pre-specify, at the time of initial marketing authorization, the types of future modifications an AI-enabled device is expected to make and the protocol for making them — intended to reduce the need for a new marketing submission every time a learning algorithm updates.
December 4, 2024 — final guidance on PCCPs for AI-enabled device software functions. FDA finalized the PCCP guidance, broadening its scope beyond machine learning specifically to AI-enabled device software functions (AI-DSF) more generally, and shifting its terminology for training data categories from “training and testing” to “training, tuning, and testing.”
January 7, 2025 — draft guidance on AI-enabled device software functions: lifecycle management and marketing submission recommendations. This is the broader, TPLC-oriented draft guidance covering what a marketing submission for an AI-enabled device should contain across its lifecycle. As of this writing it remains in draft status; industry commentary has suggested finalization is unlikely before late 2026 or 2027, but sponsors should verify current status directly with FDA rather than relying on that estimate.

Alongside these, FDA has also issued companion “guiding principles” documents jointly with Health Canada and MHRA on related sub-topics, including transparency for machine-learning-enabled devices and predetermined change control plans specifically for machine-learning-enabled devices — non-binding, principles-level documents rather than formal guidance, but useful as an indicator of where the tripartite regulators are converging.

Key CDRH vocabulary

SaMD (Software as a Medical Device): software intended for one or more medical purposes that performs those purposes without being part of a hardware medical device.
AI-DSF (AI-enabled device software function): FDA’s current umbrella term, broader than “machine learning,” for any device software function that uses AI.
PCCP (Predetermined Change Control Plan): a pre-authorized plan, submitted as part of the original marketing submission, describing the specific anticipated modifications to an AI-enabled device and the methodology for implementing them without a new submission for each change.
TPLC (Total Product Lifecycle): FDA’s framing for regulating AI-enabled devices continuously across development, premarket review, and postmarket performance monitoring, rather than only at a single point-in-time review.
GMLP (Good Machine Learning Practice): the set of best-practice principles — jointly endorsed by FDA, Health Canada, and MHRA — for how AI/ML models should be developed and evaluated to support device safety and effectiveness.

Importantly, an AI-enabled device still goes through the same underlying premarket pathways as any other device — 510(k) clearance, De Novo classification, or Premarket Approval (PMA) — the AI-specific guidance layers additional expectations onto those pathways rather than replacing them. See CASRAI’s guide to FDA medical device classification, 510(k), and PMA pathways for the underlying framework.

CDER/CBER’s framework for AI in drug and biological product development

The drug/biologic side of FDA’s AI policy moved more slowly than the device side but converged into a single cross-center document in early 2025:

May 2023 — discussion paper. FDA published a discussion paper on the use of AI to support regulatory decision-making for drugs and biologics, soliciting public input. It drew more than 800 comments.
August 6, 2024 and October 7, 2025 — public workshops. FDA held hybrid public workshops to gather additional stakeholder input as it developed formal guidance.
January 6-7, 2025 — draft guidance: “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.” This is FDA’s first cross-center draft guidance specifically on AI in drug/biologic development, developed jointly across CDER, CBER, CDRH, the Center for Veterinary Medicine (CVM), the Oncology Center of Excellence (OCE), the Office of Combination Products (OCP), and the Office of Inspection and Investigations (OII). It was informed in part by CDER’s own experience reviewing more than 500 submissions containing an AI component between 2016 and 2023. The guidance proposes a risk-based framework for establishing and documenting the “credibility” of an AI model for a specific proposed context of use, rather than certifying a model as broadly “validated” once and for all.
Comment period closed April 7, 2025. As of this writing, this guidance remains in draft status; it has not been finalized. FDA has also published a companion, non-binding “Guiding Principles of Good AI Practice in Drug Development” resource page summarizing the underlying philosophy.

Key CDER/CBER vocabulary

Context of use (COU): the specific, bounded role an AI model plays in a regulatory decision — for example, flagging a specific type of adverse event in a specific patient population — rather than a general claim about the model’s capability.
Credibility assessment: the risk-based, staged process FDA’s draft guidance proposes for a sponsor to establish confidence that a given AI model is adequate for its specific context of use, scaled to how much the model’s output influences the regulatory decision and the consequence of it being wrong.
Risk-based approach: both the CDRH and CDER/CBER frameworks explicitly scale documentation and oversight expectations to the risk the AI use poses, rather than applying one uniform standard to every use of AI regardless of stakes.

Because this guidance concerns AI used to generate or interpret data that flows into a regulatory submission, it intersects with data-integrity and electronic-records expectations sponsors already have to meet under 21 CFR Part 11 and ICH GCP — an AI tool does not exempt a sponsor from those underlying record-integrity obligations, it adds an additional layer of model-credibility documentation on top of them.

What this means in practice for research administrators and sponsors

Identify which track applies, or both. A combination product, or a program that uses an AI-enabled companion diagnostic alongside a drug, may need to satisfy both the CDRH device-side expectations and the CDER/CBER drug-side expectations simultaneously.
Engage FDA early. Both frameworks emphasize early, proactive engagement — FDA has repeatedly signaled it wants to discuss AI use in a specific program before a formal submission, similar in spirit to a Pre-IND meeting for a novel drug development question.
Document context of use narrowly, not broadly. The credibility-assessment approach rewards a sponsor who can precisely bound what the AI model is being relied on to do, rather than asserting general-purpose reliability.
Track draft-vs-final status before relying on either document as binding. As of this writing, the January 2025 CDER/CBER draft guidance and the January 2025 CDRH TPLC/marketing-submission draft guidance are both still in draft status; only the December 2024 PCCP guidance for devices is final. Draft guidance reflects FDA’s current thinking and is useful for planning, but it is explicitly non-binding on either FDA or industry until finalized, and provisions can change between draft and final versions.
Expect this area to keep changing. Given the volume of comments each draft guidance has drawn and the pace of AI development generally, sponsors should build in periodic re-checks against FDA’s own guidance pages rather than treating any single summary — including this one — as a fixed, permanent framework.

Frequently asked questions

Is there a single “FDA AI guidance” document?

No. FDA has issued several separate guidance and guiding-principles documents addressing AI from different angles — medical devices (CDRH) and drug/biologic development (CDER/CBER) are the two largest tracks, each with its own documents, and neither has been consolidated into one unified AI regulation.

Which FDA AI guidance documents are final, and which are still draft?

As of this writing, the December 2024 Predetermined Change Control Plan (PCCP) guidance for AI-enabled devices is final. The January 2025 CDRH draft guidance on lifecycle management and marketing submissions for AI-enabled device software functions, and the January 2025 CDER/CBER draft guidance on AI in regulatory decision-making for drugs and biologics, are both still in draft status and have not been finalized. Confirm current status directly on FDA’s guidance pages before relying on either as final.

What is a Predetermined Change Control Plan (PCCP)?

A PCCP is a plan, submitted as part of a device’s original marketing authorization, that pre-specifies the types of future modifications an AI-enabled device is expected to undergo (for example, retraining on new data) and the methodology for making and validating those changes — allowing certain updates to occur without a brand-new marketing submission for each one, provided the change stays within the pre-authorized plan.

Does FDA’s AI guidance apply to generative AI and large language models?

FDA’s published frameworks are written broadly enough to cover AI and machine learning generally rather than being scoped narrowly to any one model architecture, and FDA has acknowledged generative AI as a live area of interest. However, neither the device-side nor the drug-side draft guidance names generative AI or large language models as a distinct regulatory category with its own separate rules; sponsors using generative AI tools should apply the same risk-based, context-of-use principles set out in the existing frameworks and confirm with FDA directly for anything genuinely novel.

Where should I check for the current status of these guidance documents?

FDA’s own guidance pages are the authoritative source: the Artificial Intelligence in Software as a Medical Device page for the device track, and the Artificial Intelligence in Drug Development page for the drug/biologic track. Both are updated as documents move from draft to final or as new documents are added.