FDA and AI therapy: what is actually regulated

the status table

last verified: august 18, 2026. we update this page monthly, or sooner if FDA acts. numbered references link to the primary sources at the bottom of the page — mostly federal register notices, the committee’s own approved minutes, and FDA database records.

what the advisory committee actually said

the november 6, 2025 meeting is the single richest public document on how US regulators think about AI in mental health, and almost nobody has read it. the committee’s chair, Dr. Ami Bhatt, opened by narrowing the scope: the meeting was not about broadly available generative AI tools, but specifically about generative-AI-enabled medical devices intended to diagnose, treat, mitigate or prevent mental-health conditions [3].

the committee agreed the upside is real — expanded access, shorter waits, supplement to crisis services, more consistent outcome tracking. it then listed the risks in blunt terms: "missed or misinterpreted harm signals, unsafe or hallucinatory advice, model drift, bias magnified at scale, privacy vulnerabilities, and inequities related to digital access, language, and literacy," plus "data ownership, commercial incentives, and potential overuse or dependence" [3].

three recommendations are worth pulling out because they cut against how this category currently markets itself:

• wait-list controls are not good enough. the committee "supported robust comparators beyond wait-list controls" — which is most of the published chatbot literature’s design.
• measure what the tool misses, not just what it improves. members asked for endpoints beyond the PHQ-9, including functional outcomes and "false-negative rates for safety events." a tool that lowers average symptom scores while missing suicidality is a failure that a PHQ-9 endpoint will not catch.
• tell the user it is AI. for labelling, members recommended "explicit disclosure that the user is interacting with AI," clear limits of use, escalation and safety-net protocols, and warnings about excessive use.

on expanding beyond prescription use, the committee was unambiguous. asked about over-the-counter and autonomous versions, it "viewed all three proposed OTC expansions as significantly higher risk than the original prescription scenario." for autonomous diagnosis in undiagnosed users, it noted that accurate diagnosis "requires ruling out medical and psychiatric comorbidities, assessing severity, and detecting suicidality, tasks current AI systems cannot reliably perform." for an autonomous multi-condition device, it "described the risk as the highest of all." and for children and adolescents — the committee used the threshold of 21 and under — it "expressed strong discomfort with any future OTC or autonomous use in this population" [3].

and then nothing happened

this is the part that matters and the part nobody reports, so here is exactly what we checked and exactly what we found. as of august 18, 2026: FDA has not scheduled or held a second digital health advisory committee meeting — the committee’s own past-materials page lists 2024 and 2025 sections and no 2026 section, the 2026 version of that per-year URL returns a 404 where 2024 and 2025 return pages, and the committee’s landing page has not been touched since november 4, 2025 [10]. querying the federal register for FDA documents published in 2026, zero contain the phrase "digital mental health" and zero contain "generative artificial intelligence." the meeting docket, FDA-2025-N-2338, contains exactly one federal register document — the notice that announced the meeting [9].

two things did happen in 2026 that sit close enough to mention, because leaving them out would make the negative look cleaner than it is. FDA revised the General Wellness guidance on january 6, 2026 [16] — that document governs whether a mental-health app counts as a device, so it touches this category directly, but it contains no mention of AI. and CDRH’s FY2026 agenda names a future digital mental health guidance on its “Under Construction List” [13]. neither is a response to the november meeting, and neither tells anyone building in this space something they did not already know.

we want to be careful about what that does and does not mean. it does not mean FDA is doing nothing; agencies do a great deal of work that never surfaces in a public document, and a review division can be actively shaping submissions without publishing a word. it means there is no public rule, no draft guidance, and no scheduled next step nine months after the meeting. anyone building or buying in this space is operating without published federal expectations, and anyone telling you otherwise is describing the advisory committee’s advice as though it were policy. it isn’t — advisory committees make non-binding recommendations, and FDA is not obliged to follow them.

this is a negative finding with a date attached, which is the only honest way to publish one. if FDA moves, this page changes.

the guidance gap

there is no FDA guidance, draft or final, addressed to digital mental health as a category. there is one on the drawing board: CDRH's FY2026 guidance agenda lists "Clinical Evidence Considerations for Digital Mental Health Treatment Devices, including Computerized Behavioral Therapy Devices" — but on the "Under Construction List," which is the tier below both the prioritised A-list and the resources-permitting B-list [13]. it is a topic FDA has acknowledged, not a document anyone can read.

what companies work from instead is three general documents doing awkward duty.

"Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations" is the closest thing to a rulebook for AI devices. it was issued in january 2025, and nineteen months later its FDA page still reads "Draft" and "Not for implementation. Contains non-binding recommendations" [5]. the industry is being asked to plan around a document FDA has not committed to. the one AI device guidance that is final, from august 2025, governs predetermined change control plans [14] — how a cleared model may be updated, which is a question that only arises after you have been cleared.

the two documents that decide the prior question — whether software is a medical device at all — are Clinical Decision Support Software and General Wellness: Policy for Low Risk Devices. it is worth being precise about the CDS one, because it is widely misreported: it has been final since september 28, 2022 [15], not since 2026. what happened in january 2026 was two revisions three weeks apart under Level 2 procedures (21 CFR 10.115(g)(4)) — the current PDF was "issued on January 29, 2026" and "supersedes “Clinical Decision Support Software” issued on January 6, 2026" [6]. no federal register notice accompanied either revision, which is what Level 2 means. the General Wellness guidance was revised on the same january 6, 2026 date, superseding its 2019 version [16].

here is the detail that says the most: neither of those two documents — the ones that draw the line every AI mental-health product sits on one side of — contains a single mention of artificial intelligence or generative AI. the line separating a regulated therapeutic from an unregulated wellness app was drawn without reference to the technology now being pushed across it.

what has actually been authorized

while the generative-AI question sits open, FDA has kept clearing conventional digital therapeutics in psychiatry. the most recent is Motivista (CT-155) from Click Therapeutics, cleared july 30, 2026 as "a prescription digital therapeutic indicated for the treatment of negative symptoms of schizophrenia as an adjunct to clinician-managed outpatient care, including antipsychotic therapy, in patients 18 years of age and older with clinically stable positive symptoms" [8]. it is the only FDA clearance we can find carrying a schizophrenia indication.

Prismira (Lumos Labs), cleared june 13, 2025, is worth reading closely as a lesson in indication language: it is indicated "to improve attention function in adults ages 22-55 years old" with inattentive or combined-type ADHD, and its own indications statement warns that patients "may not display benefits in typical behavioral symptoms, such as hyperactivity" [11]. that is a narrower claim than "treats adult ADHD," and the gap between the two is where most marketing copy lives.

outside software, Modius Spero (Neurovalens) was granted a De Novo on may 8, 2026, creating an entirely new device type at 21 CFR 882.5807 for home-use transcranial nerve stimulation for PTSD-associated symptoms — explicitly "not intended as a stand-alone therapy or to alter usual care treatment" [12].

and the TEMPO pilot, announced december 8, 2025, is the one place FDA has named mental health as a named priority area with a mechanism attached. run alongside the CMS innovation center’s ACCESS model, it lets FDA decline to enforce certain premarket requirements for selected digital health devices in four clinical use areas, one of which is "behavioral health (depression or anxiety)" [7]. note the condition attached: devices must be "intended to be used in conjunction with clinician-supervised outpatient treatment." even FDA’s most permissive current pathway for digital mental health assumes a clinician is in the loop.

the practical read

if you are using an AI chatbot for your mental health, no regulator has evaluated it. that is not an accusation — it is a description of the current state, and it applies to the well-built products as much as the careless ones. the products that are reviewed are prescription software with narrow indications, and none of them are conversational AI.

the committee’s labelling recommendation is the one worth adopting yourself, ahead of any rule: know whether you are talking to a machine, and know what it is supposed to do. a tool that will not tell you which it is has answered the question.