can a chatbot diagnose you?

you describe the last three months to a chatbot, and it tells you, kindly and fluently, that this sounds like depression, or adhd, or ptsd. it may even be right. is that a diagnosis?

this page answers in two halves: what the fda has and has not authorised software to do, read from the agency's own documents, and what the studies of large language models measured when they were asked to recognise mental disorders. both directions, every number with its source. the running regulatory status is on our fda tracker for ai and mental health.

has the fda authorised any ai to diagnose mental illness?

not according to the briefing it wrote for its own advisory committee. the executive summary for the 6 november 2025 meeting on generative-ai mental-health devices says: “although the agency has authorized over 1200 ai-enabled medical devices (encompassing a wide range of ai technologies), none of those ai-enabled devices have been authorized for mental health uses. to date, fda has authorized fewer than twenty digital mental health medical devices that encompass non-ai technologies”^[1].

fda's own summary of the meeting puts the opening remarks of the director of its devices centre differently: that the agency has authorised more than 1,200 ai-enabled devices but “none yet involve generative ai for mental-health conditions”^[2]. “no ai” and “no generative ai” are not the same sentence, and the difference matters for the next section.

on diagnosis specifically, the committee was asked about a hypothetical over-the-counter product that would diagnose and treat depression in people who had not been diagnosed by a clinician. fda's summary records that the committee “agreed that the risk is substantially greater,” and that “accurate diagnosis requires ruling out medical and psychiatric comorbidities, assessing severity, and detecting suicidality, tasks current ai systems cannot reliably perform”^[2]. any future pathway, it said, would need “high diagnostic concordance with clinicians, validated rule-outs, extensive rct data”.

then what are the ai autism devices?

they are real, they are on fda's ai list, and they sit awkwardly beside the “none” sentence. in june 2021 fda granted cognoa's asd diagnosis aid a de novo classification, creating a new device type, “pediatric autism spectrum disorder diagnosis aid”^[3]. fda's decision summary describes how it works: “the cognoa asd diagnosis aid machine-learning algorithm produces a single scalar value” that is compared with thresholds to output positive, negative, or no result. in june 2022 fda cleared the earlipoint system, whose summary states that “artificial intelligence software analyzes the eye-tracking data and provides a diagnosis for asd”^[4].

fda's public list of ai-enabled devices, current as of 4 september 2026, contains six devices under that autism product code, from cognoa (2021, and its canvas dx in 2025) and earlitec (2022, 2023, 2025, 2026), every one with neurology as the lead review panel^[5]. five of the six were authorised before the november 2025 meeting. the same executive summary that says no ai-enabled device is authorised for mental-health uses defines mental illness to include “neurodevelopmental disorders (e.g., autism)” and names “pediatric autism spectrum disorder (asd) diagnosis aids” as an example of digital mental-health diagnostics^[1].

the documents do not say how those statements fit together, and we will not invent a reconciliation. what they do say is enough to read the devices correctly. they are prescription aids for clinicians, not consumer products. cognoa's is “not intended for use as a stand-alone diagnostic device but as an adjunct to the diagnostic process”^[3]. in its pivotal study of 425 children, it gave no result for 68%; for the 32% it did classify, sensitivity was 98% and specificity 79%^[3]. earlipoint, in 475 evaluable children, reported 71% sensitivity and 80.7% specificity against expert clinician diagnosis, dropping to 56.3% and 68.1% when the experts themselves were uncertain^[4]. none of them is a chatbot, and none diagnoses depression, anxiety, ptsd or anything a person would type into a chat window.

why is a chatbot that talks about mood not a medical device, when one that diagnoses would be?

because fda regulates by intended use and claims, and diagnosis is the claim that crosses the line. the general wellness guidance, reissued on 6 january 2026, treats claims to “promote relaxation or manage stress” and to boost self-esteem as general wellness, outside device rules^[7]. it lists, as examples of claims that are not wellness, “a claim that a product helps treat an anxiety disorder” and “a claim that a computer game will diagnose or treat autism”. it even allows a narrow disease-adjacent claim: software that tracks sleep and exercise and “may help living well with anxiety” is its own example of a wellness claim^[7].

the clinical decision support guidance, in the version issued on 29 january 2026 (which superseded a version issued on 6 january 2026), sets out when decision-support software is excluded from the definition of a device^[6]. the exclusion is built around software that supports a health professional who can independently review the basis for a recommendation. for everyone else it is blunt: “software functions that support or provide recommendations to patients or caregivers – not hcps – meet the definition of a device”^[6]. even a tool for clinicians can fail the test: one of the guidance's own examples, software that gives a professional a prioritised list of depression treatments without showing the studies behind it, is a device^[6].

put together: a chatbot that listens, reflects and suggests a breathing exercise is making wellness-type claims. a chatbot that tells a member of the public which disorder they have is doing what these documents describe as a device function, and no such device has been authorised^[1],[6],[7]. the briefing document also describes a middle category where fda does not intend to enforce device rules, with examples such as a “skill of the day” coping technique for people already diagnosed^[1].

what is a diagnosis, actually?

a clinician's judgment that a set of criteria is met, after ruling other things out. the fda briefing borrows the dsm-5 definition: a mental disorder involves “clinically significant disturbances in cognition, emotional regulation, or behaviors that are associated with significant distress or disability in social or occupational functioning”^[1]. every word of that requires judgment about a person over time: significant compared to what, distress measured how, caused by this or by a thyroid, a medication, grief, or circumstances. that is why the committee named ruling out comorbidities and assessing severity as the hard part^[2]. what the dsm is and where its categories came from is on the dsm and what the dsm is for.

how accurate are chatbots at recognising depression or ptsd?

very good on vignettes written to describe a disorder, much weaker on real conversations. in a 2025 study, four models were given six clinical vignettes and compared with norms from mental-health professionals^[9]. on the two depression vignettes, gemini, chatgpt-3.5 and chatgpt-4 named depression in 100% of runs, against 95% for professionals. on the ptsd vignette, three models scored 100% and claude 95%, against 90% for professionals. on early schizophrenia, chatgpt-4 managed 55% against 95% for professionals, and 67% on chronic schizophrenia^[9].

vignettes are the easy case: a paragraph written so that one answer is right, possibly much like text the model was trained on. transcripts of real interviews are harder. in a 2026 study, six current models, including gpt-5, claude sonnet 4 and gemini 2.5, predicted the 17 items of a ptsd checklist from 100 clinical interview transcripts. the best mean accuracy for whether each symptom was present was 0.705, and for severity the best model barely beat always guessing the most common answer (0.438 against 0.399). the authors called the performance “insufficient for clinical application”^[10]. on the same interview corpus, small models that could run privately on a clinic's own hardware ranged from 23% to 68% accuracy at detecting depression, and agreed poorly with each other^[11].

the much-cited 2023 study comparing chatgpt with primary care physicians on depression vignettes is often quoted as a diagnosis study. it was not: the vignettes already described depression, and the study compared treatment recommendations^[8]. for mild depression, chatgpt-3.5 and chatgpt-4 recommended psychotherapy in 95.0% and 97.5% of runs, against 4.3% for physicians, and unlike the physicians showed no gender or socioeconomic bias^[8]. a 2024 systematic review of 16 studies summed up the field as promising for diagnostic support, with “difficulties with complex cases and potential underestimation of suicide risks”^[14].

can a chatbot give you a phq-9, and is the score a diagnosis?

it can give you one accurately, and no, the score is not a diagnosis. the phq-9 is a free nine-question screening questionnaire; who wrote it, why it is free and who has been paid around it is on who wrote the phq-9, and who gets paid. in a 2026 study, 132 adults completed both a self-administered phq-9 and one delivered by an llm chatbot called hopebot; the scores agreed closely, an intraclass correlation of 0.92 with a median difference of 1 point^[13]. the authors are explicit that this tested whether the chatbot reproduces “the questionnaire's structure and scoring accuracy rather than diagnostic validity”^[13].

that is the honest ceiling for a chatbot today: it can ask the standard questions well and add up the answers. a high score says it is worth seeing someone who can assess the whole picture. it does not say what is wrong, and a low score does not rule anything out.

where talking to people fits

we run a peer-support app, so weigh this paragraph accordingly. peer support does not diagnose either, and it should not try: the people on veisund have lived experience, not clinical training, and it is not therapy. what other people can do is say what it was like for them when they got a diagnosis, or didn't, and what they asked the clinician. on veisund no ai writes replies; people answer, with one automated exception: if a post suggests immediate risk, an automated safety response surfaces crisis resources. people can answer; nobody is promised an answer. the evidence on what peer support does and does not do is on is peer support effective, and the free routes to an actual assessment are on free therapy. the hub page on using chatgpt as a therapist covers the wider question.

what to ask if a chatbot gives you a label

questions, not advice. they are the ones the documents above make it reasonable to ask.

• is this a screening score from a standard questionnaire, or the chatbot's own guess?
• what else could explain this, and has anyone ruled out a medical cause, a medication, or circumstances?
• does the product claim to diagnose, and if so, where is its fda authorisation?
• would a clinician reach the same conclusion after an interview, and how would i get that interview?
• what does the chatbot do if what i wrote suggests i am at risk?

adam