AI is now a mainstream device technology
Artificial intelligence has moved from pilot projects to a large and growing share of FDA authorizations. FDA maintains a public list of AI-enabled medical devices, “a resource intended to identify AI-enabled medical devices that are authorized for marketing in the United States,” most recently updated September 4, 2026 (FDA). One third-party analysis of the list counts 1,614 devices, with radiology accounting for 76.2% (Complizen); FDA itself does not publish a running total on the page.
The list is about to change in an important way. FDA says it “will explore methods to identify and tag medical devices that incorporate foundation models,” including “identifying those devices incorporating LLM-based functionality in a future update of this list” (FDA). Generative AI is arriving, and FDA wants it visible.
How FDA regulates AI-enabled devices
AI does not have its own premarket pathway. AI-enabled devices go through 510(k), De Novo or PMA like other devices, with the same quality system obligations under the QMSR. What differs is the guidance layered on top.
| Instrument | Status | What it covers |
|---|---|---|
| Predetermined Change Control Plan (PCCP) guidance for AI-enabled device software functions | Final (FDA) | How to pre-specify model modifications, the protocol to implement them and an impact assessment, so changes can be made without a new submission |
| AI-enabled device software functions: lifecycle management and marketing submission recommendations | Draft, January 2025 (docket FDA-2024-D-4488) (FDA) | Recommended submission content: model description, data management, performance validation, bias, transparency, monitoring |
| Good Machine Learning Practice (GMLP) guiding principles | IMDRF final, 10 principles, January 2025 (FDA) | Lifecycle practices from data quality to monitoring |
| Transparency for machine learning-enabled devices: guiding principles | June 13, 2024 (FDA) | What users need to know about a model’s purpose, performance and limits |
| Clinical decision support and general wellness guidances | Revised January 2026 (Arnold & Porter) | Which software functions are not regulated as devices |
Three practical consequences
- Plan for change from day one. A PCCP is the main tool for updating a model without a new 510(k). It must be submitted with the original marketing submission and describes changes that are specified, validated and bounded in advance.
- Document data like a design input. The draft lifecycle guidance expects detailed descriptions of training and test data, how they were separated, and how performance holds across relevant subgroups.
- Treat monitoring as part of the device. Real-world performance monitoring, drift detection and user transparency belong in the design, risk file and postmarket plan, not in a slide deck.
Does the software count as a device at all?
The January 2026 revisions to FDA’s clinical decision support and general wellness guidances narrowed the set of software functions FDA actively regulates (Arnold & Porter). Teams building AI that supports clinicians should reassess classification against the revised guidance before assuming a submission is, or is not, needed.
Generative AI and LLMs in medical devices
FDA’s Digital Health Advisory Committee met on November 20–21, 2024 to discuss “Total Product Lifecycle Considerations for Generative AI-Enabled Devices.” FDA’s executive summary noted that foundation models can be deployed “locked” or “unlocked,” warned that unlocked third-party foundation models “may raise further new and significant questions during FDA’s review,” and stated: “Importantly, FDA has also yet to authorize any unlocked AI-enabled device” (FDA). A second committee meeting on November 6, 2025 focused on generative AI-enabled digital mental health devices (FDA).
For developers, the practical message is that generative AI devices face open questions on hallucination, output variability, third-party model changes and evaluation methods. Expect detailed questions about how outputs are bounded and verified, how the underlying model is version-controlled, and how performance is monitored after release.
The EU AI Act: a later but firm deadline
AI systems that are medical devices, or safety components of them, requiring notified body assessment are classed as high-risk under the EU AI Act through its Annex I link to the MDR and IVDR. In 2026, the EU’s Digital Omnibus on AI, Regulation (EU) 2026/1744 of July 8, 2026, published in the Official Journal on July 24 and in force from July 27, postponed the high-risk deadlines. Obligations for AI in products covered by Annex I, including medical devices, now apply from August 2, 2028 (Cuatrecasas; Gibson Dunn).
Two years is less time than it sounds. AI Act requirements on data governance, logging, transparency, human oversight, accuracy and robustness must be integrated into MDR technical documentation and QMS, and notified bodies will need designation for both regimes.
| Date | Milestone |
|---|---|
| Jun 13, 2024 | FDA transparency principles for ML-enabled devices |
| Nov 20–21, 2024 | FDA advisory committee on generative AI-enabled devices |
| Jan 2025 | FDA draft AI lifecycle guidance; IMDRF GMLP principles final |
| Nov 6, 2025 | FDA advisory committee on generative AI mental health devices |
| Jan 2026 | FDA revises CDS and general wellness guidances |
| Feb 2, 2026 | QMSR in effect for all devices, including AI-enabled ones |
| Jul 27, 2026 | EU AI Act Digital Omnibus in force |
| Sep 4, 2026 | FDA AI-enabled device list updated; LLM tagging planned |
| Aug 2, 2028 | EU AI Act high-risk obligations apply to AI in medical devices |
AI inside device companies, not just inside devices
AI is also changing how device companies work: drafting and reviewing documents, triaging complaints, monitoring suppliers. FDA itself launched an agency-wide generative AI tool, Elsa, in June 2025 (RAPS). When AI tools are used in production or the QMS, FDA’s Computer Software Assurance guidance applies its risk-based validation framework to “artificial intelligence/machine learning tools” (FDA CSA guidance). The same disciplines apply: define intended use, assess process risk, validate proportionately and keep humans accountable for decisions.
Medical device AI readiness checklist
For AI-enabled devices
- Classification confirmed against the January 2026 CDS and general wellness guidances
- Data management documented: sources, representativeness, train/test separation, subgroup coverage
- Performance validated on independent data, including subgroup analysis
- PCCP drafted for planned model changes, with protocol and impact assessment
- Transparency: intended use, inputs, performance and limitations communicated to users
- Postmarket performance monitoring and drift detection in the risk file and PMS plan
- For generative AI: output boundaries, hallucination testing and foundation model version control
- EU AI Act gap assessment against the August 2, 2028 deadline
For AI used in QMS and operations
- Intended use and process risk defined for each AI tool (CSA)
- Outputs traceable to sources and reviewed by accountable people
- Validation proportionate to risk and repeated after model changes
Evidence-first AI for regulatory documents
AskDona applies AI to regulatory documentation: it assesses technical files, procedures and records against requirements and cites the passage behind every finding, so your experts can verify the evidence and make the decision.
Run a sample regulatory document assessment on one section of an AI device technical file. AskDona shows the evidence; your team decides.
FAQ
How many AI medical devices has FDA authorized? FDA maintains a list of AI-enabled devices authorized for marketing, updated September 4, 2026 (FDA). One third-party count puts the total at 1,614 (Complizen).
Has FDA approved an LLM-based medical device? In late 2024, FDA said it had yet to authorize any unlocked AI-enabled device (FDA), and it now plans to tag devices with LLM-based functionality on its list.
What is a PCCP? A predetermined change control plan describes planned modifications to an AI-enabled device, how they will be implemented and validated, and their impact, so they can be made without a new submission (FDA).
When does the EU AI Act apply to medical devices? High-risk obligations for AI in medical devices apply from August 2, 2028, after the 2026 Digital Omnibus (Gibson Dunn).
References
All pages opened and checked by October 13, 2026.
| Source | Type | Date | Used for |
|---|---|---|---|
| FDA: AI-Enabled Medical Devices list | Regulator | Sep 4, 2026 | List, LLM tagging |
| Complizen: 1,614 AI devices | Third-party analysis | 2026 | Device count, radiology share |
| FDA: PCCP guidance for AI | Regulator guidance | Final | PCCP |
| FDA: AI lifecycle draft guidance | Regulator guidance | Jan 2025 (draft) | Submission content |
| FDA: GMLP principles | Regulator | Jan 2025 | GMLP |
| FDA: Transparency principles | Regulator | Jun 13, 2024 | Transparency |
| FDA: DHAC GenAI executive summary | Regulator | Nov 2024 | Generative AI |
| FDA: DHAC Nov 6, 2025 agenda | Regulator | Nov 6, 2025 | GenAI mental health |
| Arnold & Porter: CDS guidance | Law firm | Jan 21, 2026 | CDS, wellness |
| Cuatrecasas: Digital Omnibus on AI published | Law firm | Jul 2026 | Regulation (EU) 2026/1744 |
| Gibson Dunn: AI Act Omnibus | Law firm | 2026 | New deadlines |
| RAPS: Elsa launch | Professional body | Jun 2025 | FDA AI use |
| FDA: CSA guidance | Regulator guidance | Feb 3, 2026 | AI tools in QMS |