Woebot
Woebot Health (San Francisco) — 2017–2025
At a glance: Woebot was the world’s most scientifically studied mental health chatbot: 14 randomized clinical trials, FDA “Breakthrough Device” designation for postpartum depression, 1.5 million users. Unlike ChatGPT or Claude, Woebot did not use generative AI: every response was hand-written by CBT clinicians and orchestrated by a decision tree. The application was shut down on June 30, 2025, with its founder citing the prohibitive cost of FDA regulation and the impossibility of integrating LLMs within a nonexistent regulatory framework. Its trajectory — from clinical promise to shutdown — is a lesson for the entire field of AI in mental health.
Identity
Publisher: Woebot Health (San Francisco, USA)
Period: 2017–2025 (shut down June 30, 2025)
Type: Rule-based CBT chatbot (no generative AI)
Founder: Alison Darcy, PhD (clinical psychologist, Stanford)
Pricing: Free (consumer) / Enterprise licenses
Languages: English only
Access: Was available on iOS, Android, and web — no longer accessible
Users: ~1.5 million total
Origin: Making CBT Accessible
Alison Darcy is a clinical psychologist, trained in Dublin and then at Stanford. For 20 years, she worked on digital treatments for mental health — first online support groups for eating disorders, then virtual reality in child psychiatry. Her recurring observation: validated treatments exist, but most people cannot access them.
At Stanford, she met Andrew Ng, a pioneer in AI, and led his Health Innovation Lab. From this collaboration came Woebot: a chatbot that delivers CBT techniques through structured conversations. The first trial (2017) showed a significant reduction in depressive symptoms within two weeks — a timeframe comparable to antidepressants.
What followed: $123.5 million in funding, the FDA “Breakthrough Device” designation, and in 2023, Darcy’s inclusion in the TIME 100 AI.
What Woebot Did (in plain terms)
Woebot was fundamentally different from ChatGPT, Claude, or Gemini. It was not a language model. It was a rule-based system: every response was hand-written by writers trained in CBT, IPT, and DBT, then organized in a decision tree. AI was only used to understand the user’s messages (NLP), not to generate responses.
Clinical implication of this technical choice
No hallucination risk, no unpredictable content, no sycophancy. Every intervention was controlled and reproducible — which made the tool testable in clinical trials. This is precisely what made it the most studied chatbot and the most validated. It is also what made it obsolete compared to LLMs.
Integrated approaches
- • Cognitive Behavioral Therapy (CBT)
- • Interpersonal Psychotherapy (IPT)
- • Dialectical Behavior Therapy (DBT)
Features
- • Daily mood check-ins
- • Cognitive restructuring exercises
- • Modular psychoeducation
- • Graphic mood tracking
- • Guided journaling
Therapeutic alliance with a chatbot
A study of 36,070 users found that Woebot achieved a therapeutic alliance score (WAI-SR) of 3.8/5 — comparable to group CBT (3.8) and close to individual CBT (4.0). The bond formed within 3 to 5 days and did not diminish over time.
Clinical Results
Woebot is the only mental health chatbot with a substantial body of clinical evidence: 14 randomized clinical trials published in peer-reviewed journals.
Depression in young adults (Fitzpatrick et al., 2017)
- • RCT, 70 participants (ages 18–28), 2 weeks
- • Significant reduction in depressive symptoms (PHQ-9) vs. control group
- • Moderate effect size (d = 0.44), significant after Bonferroni correction
- • Limitations: small sample, short duration, founder is co-author
Problematic substance use (Prochaska et al., 2021)
- • RCT, 180 adults
- • One-third reduction in frequency of problematic substance use episodes
- • Improved confidence and decreased cravings
- • No significant difference on general depression/anxiety
Stress and burnout (2023)
- • Exploratory study (non-randomized), 8 weeks
- • Significant reduction in perceived stress and burnout, increased resilience
- • Effect more pronounced in participants with high baseline symptoms
Postpartum depression (pivotal trial)
- • WB001: FDA “Breakthrough Device” designation (2021)
- • Multicenter, double-blind, randomized trial
- • First patient enrolled — results never published (interrupted by shutdown)
Critical reading: Woebot’s results are the strongest in the field, but remain modest. Samples are small (70–180 participants), durations short (2–8 weeks), and the founder is co-author of the initial studies. The pivotal trial on postpartum depression — which could have provided decisive evidence — was never completed. This is nevertheless far more than what ChatGPT, Claude, or Replika can show.
Why Woebot Shut Down: Lessons for the Field
On June 30, 2025, Woebot shut down its consumer application. Alison Darcy explained: “AI moves faster than regulators.” The causes are multiple.
1. The regulatory impasse
The FDA has a framework for rule-based chatbots, but no framework for LLMs. Woebot wanted to integrate generative AI to improve its product, but could not do so without regulatory approval — which did not exist. An FDA advisory committee only convened on the topic in November 2025, five months after the shutdown.
2. The cost of rigor
Running clinical trials, drafting FDA submissions, maintaining evidence standards is expensive. Meanwhile, ChatGPT reached hundreds of millions of users with no clinical validation or regulatory approval — simply because it is not marketed as a health tool.
3. The ethical chatbot paradox
The safest and best-validated application shut down, while tools with no clinical validation (ChatGPT, Replika) are massively used for emotional support. Illinois and New York have begun legislating, but the regulatory gap remains wide at the federal level.
For the clinician: Woebot illustrates a structural paradox. Tools that seek to prove their efficacy and obtain regulatory approval are penalized by the cost and slowness of the process. Tools that do not claim to be therapeutic escape regulation while being used as such. This disconnect is the central problem of the field.
Our Analysis
Woebot deserves a card on this site precisely because it shut down. Its trajectory is the clearest demonstration that clinical rigor and commercial success do not coincide in the field of AI in mental health — at least not yet.
For clinicians, Woebot represents what a “well-made” digital mental health tool could look like: based on validated approaches (CBT, IPT, DBT), written by clinicians, tested in randomized trials, submitted to regulation. It is also the demonstration that a rule-based chatbot can form a measurable therapeutic alliance (3.8/5 on the WAI-SR), even without generative AI.
But Woebot’s story also reveals the model’s limits. A rule-based chatbot cannot finely adapt to each user’s language, cannot capture subtle emotional nuances, cannot integrate the context of a long conversation. LLMs can — but without the safety and reproducibility guarantees of a rule-based system. The field’s challenge is to find how to combine the two.
As Scott Wallace (psychologist specializing in AI and mental health) writes: when “the most ethical and the most science-based” shuts down, it says something about market incentives. The tools that generate the most revenue are not the ones that heal best. This observation should guide clinicians’ thinking about the entire field.
Related Concepts on This Site
Why we treat computers as social actors
Attributing human qualities to machines
What Woebot avoided by design
FRA risk Woebot structurally eliminated
FRTherapeutic alliance with a rule-based chatbot
Structured CBT vs. unconditional validation
References
Fitzpatrick, K. K., Darcy, A. & Vierhile, M. (2017). Delivering CBT to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot): A RCT. JMIR Mental Health, 4(2), e19.
Prochaska, J. J. et al. (2021). A therapeutic relational agent for reducing problematic substance use (Woebot): Development and usability study. Drug and Alcohol Dependence.
Darcy, A. et al. (2023). Changes in stress, burnout, and resilience associated with an 8-week intervention with Woebot. Internet Interventions.
STAT News (2025). Woebot Health shuts down pioneering therapy chatbot.
Wallace, S. (2025). R.I.P. Woebot: When the most ethical and science-based AI mental health tool shuts down. Advances in AI for Mental Health (Medium).
Palmer, M. (2025). Digital Mental Health Tools and AI Therapy Chatbots: A Balanced Approach to Regulation. Hastings Center Report.
IEEE Spectrum (2024). Woebot, a Mental-Health Chatbot, Tries Out Generative AI.
Last updated: February 2026