Large language models are moving quickly into clinical documentation, decision support, and patient-facing tools. For regulatory and clinical teams building or deploying them, the first question is whether a product counts as a medical device at all. If it does, the team has to work out which premarket pathway and lifecycle controls apply, given that FDA has issued no generative AI-specific rule or final guidance.
The analysis below covers how FDA applies the existing device definition and the section 520(o) software exclusions to LLM-enabled products. It then looks at how the 510(k), De Novo, and PMA pathways interact with FDA's AI lifecycle recommendations and predetermined change control plans. Finally, it reviews what FDA has said publicly, including at Digital Health Advisory Committee meetings, and what 510(k) and De Novo records show about whether any LLM-based device has been authorized.
How FDA regulates generative AI and LLM-enabled clinical software
Short answer
FDA has no generative AI-specific statute, regulation, or final guidance. LLM-enabled clinical software goes through the existing device framework. Whether a product is a device depends on its intended use and function, under the device definition and the section 520(o) software exclusions. A product that is a device follows the usual risk-based pathways (510(k), De Novo, PMA). Those pathways are layered with FDA's AI-specific lifecycle recommendations (still in draft, not binding requirements) and, where relevant, a predetermined change control plan (PCCP) 124219438.
On authorizations: at the November 2024 Digital Health Advisory Committee (DHAC) meeting, FDA stated, "we have not authorized any generative AI-enabled devices, nor have we authorized any adaptive AI-enabled devices" 352. At the November 2025 DHAC meeting, FDA made a narrower statement: "to date, there are no generative AI-enabled medical devices that have been authorized for use in mental health conditions" 353. That statement does not address other clinical areas. A targeted review of 510(k) and De Novo records found no cleared or granted device whose public documents describe an LLM or generative-AI component as part of its function 382383384285.
Is LLM-enabled software a device?
FDA's DHAC briefing materials use the term "GenAI-enabled device" for a medical device in which generative AI is integral to the primary function or directly enables the output. FDA also says that not every healthcare use of generative AI is a device 438439. For example, FDA noted that some uses, such as administrative documentation, fall outside device regulation. By contrast, it described diagnosis-generating tools and mental-health chatbots as potential device use cases 450.
The device boundary is set by existing guidance.
-
Clinical decision support (CDS). FDA's final Clinical Decision Support Software guidance was issued in January 2026 818. Under it, CDS falls outside the device definition only if it meets all four section 520(o)(1)(E) criteria 19:
- it does not acquire, process, or analyze medical images, IVD signals, or physiological signals 20
- it displays, analyzes, or prints medical information 21
- it is intended to support or provide recommendations to a health care professional, which FDA reads as supporting rather than replacing or directing the HCP's judgment 2223
- it lets the HCP independently review the basis for its recommendations 24
Functions that give recommendations to patients or caregivers rather than HCPs meet the device definition 25262728. FDA also does not consider critical, time-sensitive decisions to satisfy criterion 4 30. The guidance does not set a separate rule for software just because it generates text. Function and intended use control the outcome 29.
-
General wellness and enforcement discretion. The general wellness guidance was reissued January 6, 2026, replacing the 2019 edition. CDRH does not intend to review low-risk products intended only for general wellness, but that policy does not cover outputs used for clinical treatment decisions 314315320. The September 28, 2022 device software functions and mobile medical applications guidance separates non-device functions from lower-risk device functions under enforcement discretion. Supplemental coaching for patients managing identified conditions is one example of the second group 321322.
For an LLM product, the practical point is this. A clinician-facing summarizer or recommender that is transparent and reviewable may fit the non-device CDS criteria. A patient-facing chatbot that addresses a disease, or a tool that generates a differential diagnosis, is likely to be a device function.
How FDA frames generative AI risk
FDA's DHAC briefing materials give working definitions:
- Generative AI: models that generate derived synthetic content (text, images, audio, video) resembling their training data 434.
- Foundation models: models applicable to multiple tasks, built on broad datasets and later adapted for particular uses 435436.
- Large language models: models trained on large text corpora to generate natural-language responses for tasks such as summarization or question answering. A chatbot provides a conversational interface to an LLM 437.
FDA identified these regulatory challenges specific to generative AI:
- Intended-use boundaries. Models that handle many tasks and inputs make the limits of a device's intended use hard to confirm. That in turn complicates risk-based classification 438.
- Output variability and hallucination. FDA describes hallucinations as erroneous or false content that may look authentic. This is especially consequential where accurate clinical information is critical 440441. Free-text and multimodal inputs can also produce combinations never seen during development 442.
- Foundation-model dependency. Third-party foundation models are generally not built for a specific device. Monitoring may need to cover both the device model and the underlying foundation model, and site-specific implementations add complexity 443444445. The committee also raised concerns about limited transparency into the training and tuning data of third-party models 446447.
FDA's hypothetical use cases included:
- a foundation model supporting both tumor identification and a chatbot that discusses the image and the pathology report
- a chatbot that combines laboratory results, vital signs, and patient data to generate a differential diagnosis 444442
DHAC recommendations on evidence and monitoring
The inaugural DHAC meeting (November 20 to 21, 2024) covered generative AI-enabled devices broadly. It applied a total product lifecycle (TPLC) lens to premarket performance evaluation, risk management, and postmarket monitoring 124125126157.
- Premarket: the committee pointed to:
- Postmarket: the committee pointed to:
- scalable surveillance of real-world performance, drift, hallucinations, and adverse events
- auditing across sites and populations
- sharing performance findings with providers and patients 131
The November 6, 2025 meeting focused on generative AI-enabled digital mental-health devices intended to diagnose or treat psychiatric conditions, not general-purpose chatbots 127128. FDA identified these risks 127128:
- confabulation
- biased or inappropriate content
- omitted medical information
- accuracy decline and data drift
- patient misinterpretation of outputs
- uncertainty about clinician oversight
FDA asked whether interactive, adaptive devices need different clinical evidence and trial designs 130. These were discussion points, not new requirements 133.
Applicable guidance for AI-enabled device submissions
| Document | Status and date | Relevance to LLM-enabled devices |
|---|---|---|
| Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations | Draft, January 7, 2025 240224 | TPLC-based submission content (summarized below the table). It has no LLM-specific recommendations; transformers are mentioned as one model class to describe 251. FDA specifically requested comment on whether the guidance adequately addresses generative AI 158159. |
| Predetermined Change Control Plans for AI-Enabled Device Software Functions | Final, August 18, 2025 201216 | A PCCP has three parts: Description of Modifications, Modification Protocol, and Impact Assessment 218. FDA must establish it through a 510(k), De Novo, or PMA before changes are made under it 219. This is the main mechanism for planned model updates without a new submission for each change 217. |
| Clinical Decision Support Software | Final, January 2026 818 | Sets the device/non-device boundary for clinician-facing recommendation software 19. |
| Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products | Draft, January 7, 2025 408409 | This is not device guidance. It sets a seven-step, risk-based credibility framework tied to context of use for AI that produces data supporting drug safety, effectiveness, or quality decisions 416417419421389. It excludes drug discovery and operational uses, such as drafting submissions, that do not affect patient safety, drug quality, or study reliability 414415. |
The draft device guidance asks for the following submission content:
- a device description that states AI use, inputs and outputs, and clinical workflow 241221
- model description and development methods 242243
- dataset provenance, representativeness, and test-data independence 244245246
- subgroup performance validation and human factors evaluation 223247248
- transparency information, with an optional model card 228226
- cybersecurity documentation and postmarket performance monitoring 249250222
A search of FDA's guidance library found no current guidance that specifically addresses generative AI, LLMs, or foundation models in the AI sense 167168193194.
The August 2026 discussion paper
On August 18, 2026, FDA issued a discussion paper seeking feedback on generative AI-enabled medical devices. FDA described it as not a final regulatory framework 165166. The paper covers 288:
- a possible two-axis risk-assessment framework
- a premarket "competency assessment" approach, inspired at a high level by physician training and evaluation, that combines non-clinical benchmarking with clinical confirmation
- several risk-proportionate postmarket monitoring approaches
- targeted questions on foundation models and agentic AI
Comments are due October 19, 2026 under docket FDA-2026-N-7874 289. The announcement does not say whether FDA has authorized any generative AI-enabled device 136. Sponsors developing LLM-based clinical products should treat this paper as the clearest signal of where premarket expectations are heading.
Authorization status: what the record shows
- Generative AI overall: FDA said in November 2024 that it had not authorized any generative AI-enabled device 352.
- Mental health: FDA said in November 2025 that no generative AI-enabled device had been authorized for mental-health conditions. Across all AI-enabled devices, FDA counted more than 1,200 authorizations 134135353.
- AI-enabled devices generally: FDA's public AI-enabled device list is large but not exhaustive. FDA reported 950 AI/ML-enabled authorizations as of August 2024 155156. Most entries are imaging, signal-analysis, and planning tools.
- 510(k) and De Novo records: searches for "large language model," "generative AI," "natural language generation," and chatbot terms found no device whose public summary describes an LLM or generative component as part of its function 382383384. One near miss is the GI Genius module 510(k), K241887, cleared July 25, 2024. Its summary says the updated hardware can accelerate generative-AI inference models. It does not say the cleared lesion-detection function uses one 285286287.
This is evidence of absence in FDA's published statements and in 510(k) and De Novo summary text. It does not prove that no authorization exists, because summary documents may not describe the model architecture. As of the most recent FDA statements found, FDA has not publicly identified any LLM-based medical device it has authorized.
FDA's internal use of LLMs
FDA uses LLMs internally, separately from regulating them. In June 2025 it launched Elsa, an agency-wide LLM-powered tool for reading, writing, summarizing, and review support 161162. FDA added optional agentic-AI capabilities in December 2025 163 and announced Elsa 4.0 in May 2026 164. These are agency operational tools, not authorized medical devices.