Industries / Health / Healthcare / Behavioral health
Therapy
Treating mental disorders by talking with a trained person, alone or in groups, and through programs, apps and AI chatbots that deliver the same methods.
7
Milestones achieved, from 1895 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1895
Talking as treatment
Josef Breuer treated Bertha Pappenheim, known as "Anna O.", from 1880 to 1882, and he and Freud discussed her case in Studies on Hysteria in 1895. It became the most studied case in the history of psychoanalysis. Biographical information published from 1953 onward led a 2003 review to argue that dependence on chloral hydrate and morphine, and a mood disorder, must be considered in her case.
LessonCase reports by the treating doctors can leave out what later records show.
Credit
J. Breuer and S. Freud
1977
Pooling therapy trials
In 1952 the psychologist H. J. Eysenck argued from hospital and insurance records that "some two-thirds of severe neurotics" recovered or improved considerably without systematic psychotherapy. In 1977 Mary Lee Smith and Gene Glass coded 375 controlled studies and combined their 833 results by meta-analysis, a statistical way of pooling many studies. The typical therapy client ended up better off than 75% of untreated people, and few differences appeared between types of therapy.
LessonTallying studies one by one as wins and losses can hide an effect that pooling them reveals.
Credit
Mary Lee Smith and Gene V Glass, University of Colorado
1989
Therapy tested like a drug
The US National Institute of Mental Health assigned 250 outpatients with major depression at random to 16 weeks of interpersonal psychotherapy, cognitive behavior therapy (CBT), the antidepressant imipramine, or a placebo pill, the last two with regular clinical management. Neither therapy did significantly worse than imipramine. Against placebo with clinical management, there was limited evidence for interpersonal therapy and none for CBT, and among less severely depressed patients no treatment beat placebo.
LessonIn milder depression, a placebo with regular clinical contact can do as well as active treatment.
Credit
I. Elkin and colleagues, National Institute of Mental Health
2015
Online CBT in practice
England's National Institute for Health and Care Excellence (NICE) issued guidance on computerized CBT for depression and anxiety on 22 February 2006. The REEACT trial then gave adults with depression in general practice a commercial program (Beating the Blues) or a free one (MoodGYM), with weekly supporting phone calls, on top of usual care. Neither added any benefit over usual care at four months, uptake and use were low, and 24% had dropped out by four months. NICE withdrew the 2006 guidance in July 2018.
LessonA program that works in trials can do little in routine care if few people use it.
Credit
S. Gilbody and the REEACT team, University of York; NIHR Health Technology Assessment programme
2020
Apps on prescription
Under Germany's Digital Healthcare Act of 2019, doctors and psychotherapists can prescribe approved health apps (DiGA) paid for by statutory health insurance. The federal drugs and devices agency, BfArM, reviews each in three months and may list it provisionally while its benefit is still being tested. By the end of 2025, 74 apps had been listed, 16 (22%) had been struck off, and 55% of the 58 remaining addressed mental disorders, according to the national association of statutory health insurers. A 2023 review of 17 mental health apps found that manufacturers took part in 23 of 36 published trials and that only 6 of 39 randomized trials used an active treatment as the comparison.
LessonPaying for apps before their benefit is proven means paying for some that never prove it.
Credit
BfArM and Germany's statutory health insurance; S. Schreiter and colleagues, Berlin, for the review
2025
Law catches up with chatbots
Woebot, a chatbot that guided users through CBT exercises with pre-scripted replies and had about 1.5 million users, shut down its app at the end of June 2025. Its founder, Alison Darcy, said the closure was largely due to the cost and challenge of meeting FDA requirements for marketing authorization, and that the FDA had not yet worked out how to regulate large language models. On 1 August 2025 Illinois enacted the Wellness and Oversight for Psychological Resources Act, passed unanimously, which bars anyone from using AI to provide therapy or make therapeutic decisions while allowing it for administrative support, with fines of up to $10,000.
Credit
Woebot Health; Illinois Department of Financial and Professional Regulation, Rep. Bob Morgan and Sen. Ram Villivalam
2026
Commercial chatbots, pooled
A systematic review of 52 studies, 22 of them randomized trials, covering 13 commercial mental health chatbots and 110,594 participants found a small benefit for depression, about 1.6 points on the PHQ-9, below the 5 points usually taken as a noticeable change. The effect on anxiety was uncertain, the certainty of evidence was low to very low, and only 3 of the 22 trials actively looked for harms.
LessonNo reported harm is not evidence of safety when few trials look for harm.
Credit
H. Yang, J. Cai and colleagues, Psychotherapy and Psychosomatics
Watch for
A therapy model in use
Slingshot AI trained a large language model on more than 900,000 hours of anonymized mental-health transcripts to power its app, Ash. In a single-group study with no comparison group, 299 US adults with at least moderate symptoms had lower depression scores at 10 weeks (Cohen's d 0.93, a standardized measure of change) and at 12 months, while 57.2% of participants followed a non-responding course. Three of the five authors work or worked for Slingshot, another holds advisor shares, and the last received a Slingshot grant.
Achieved whena randomized trial with an active comparison, run by researchers without financial ties to Slingshot, shows a benefit.
Claimed by
Slingshot AI with researchers at NYU and the University of Washington, November 2025, revised September 2026 (preprint)
AI app beats group therapy
In a randomized trial in Israel, 995 university students in distress were given a 12-week conversational AI platform, face-to-face group therapy or a waiting list. Anxiety fell more with the AI than with group therapy (2.17 points more on the GAD-7), and depression fell more than on the waiting list. Two authors report fees, stock options or employment from KAI.AI Ltd, which they describe as outside the submitted work.
Achieved whena team with no ties to the platform's maker repeats the comparison with group therapy, in another population or country.
Claimed by
A. Shoshani and colleagues, JAMA Network Open, April 2026
Open
01
Score therapy apps on CBT
An outside table of how faithfully AI therapy apps on sale deliver CBT. Testers open test accounts under each app's terms and type sessions from 20 fictional client scripts about everyday problems such as work stress or low mood, with no crisis content. The Cognitive Therapy Rating Scale (CTRS), a widely used 11-item scale that scores a CBT session from 0 to 6 per item, is applied to every transcript by an AI judge and, without knowing which app produced it, by two CBT-trained raters. General chatbots given a plain therapist prompt serve as a baseline.
The briefCredit · unclaimed02
Full sessions, rated three times
A public set of at least 60 full-length fictional CBT sessions, about 45 minutes each, written by CBT trainers from fictional case profiles. Some are sound and some carry documented, planted faults, such as no agenda, no homework or no feedback from the client. Each session is scored on the CTRS by three trained raters from different services, working independently. AI judges are then tested against the raters.
The briefCredit · unclaimed
Open 01 · the brief
Score therapy apps on CBT
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside table of how faithfully AI therapy apps on sale deliver CBT. Testers open test accounts under each app's terms and type sessions from 20 fictional client scripts about everyday problems such as work stress or low mood, with no crisis content. The Cognitive Therapy Rating Scale (CTRS), a widely used 11-item scale that scores a CBT session from 0 to 6 per item, is applied to every transcript by an AI judge and, without knowing which app produced it, by two CBT-trained raters. General chatbots given a plain therapist prompt serve as a baseline.
Why it is still open
Outside tests of apps on sale so far measure outcomes or safety, not method. The 2026 review pooled outcome trials of 13 commercial chatbots, and Common Sense Media's May 2026 test of five AI therapy apps measured safety for teenagers. TherapyGym, a Stanford framework accepted at NeurIPS 2026, scores chatbots on the CTRS but reports results for research models, not for apps people can download.
Built with what exists
TherapyGym's automated CTRS judge, calibrated on TherapyJudgeBench, its open set of expert-rated dialogues (CC BY 4.0); hosted models such as Claude or Gemini as baselines and judges; an open-weight judge such as Mistral Large 4 (see the Note "A trillion open parameters"); two raters trained on the CTRS.
First step
Write 5 client scripts from the fictional profiles used in TherapyJudgeBench, run them in two apps, and have one CBT-trained rater score the transcripts.
Done looks like
At least 5 apps, each run on the same 20 scripts. Each transcript scored by two human raters who do not know the app, and by the AI judge. A mean CTRS score per app with a 95% confidence interval, the agreement between the two raters and between raters and judge, and the transcripts published where app terms allow.
Open 02 · the brief
Full sessions, rated three times
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 60 full-length fictional CBT sessions, about 45 minutes each, written by CBT trainers from fictional case profiles. Some are sound and some carry documented, planted faults, such as no agenda, no homework or no feedback from the client. Each session is scored on the CTRS by three trained raters from different services, working independently. AI judges are then tested against the raters.
Why it is still open
The one open expert-rated set, TherapyJudgeBench, holds 116 synthetic dialogues of five exchanges each, only 20% were rated twice, and its card says it should not be used to evaluate systems meant for real patients. The largest automated scoring study used 1,118 real recorded sessions that are not public.
Built with what exists
the CTRS; scripts drafted with a model such as Claude or Gemini and rewritten by trainers; TherapyGym's judge and general-purpose models as judges; a small open model fine-tuned as a judge on one GPU, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune").
First step
Ask two CBT trainers to write one 45-minute session each, one sound and one with three planted faults, and have three raters score both.
Done looks like
60 sessions, each with three independent ratings, and the agreement between raters (intraclass correlation) reported for every CTRS item. Each AI judge's agreement with the raters set beside the raters' agreement with each other; a judge counts as matching people when its agreement falls within the range of the rater pairs. Sessions and ratings released under CC BY.