2026-08-07 — Opened (ben-steer)

• Opened as the science-track sibling to the existing courts/legislatures thread, seeded from the EBP digital/AI-science research crawl. Parent: mh-evidence-watch. (EBP digital/AI-science crawl, 2026-08-07) ⟨steer 2026-08-07⟩

2026-08-07 — Registry check: Therabot still has zero independent replication

• ClinicalTrials.gov lists exactly two Therabot-branded studies and both are Dartmouth’s own — the original 210-person NEJM AI trial and a new single-arm feasibility spinoff (Therabot-CALM, cannabis use plus anxiety/depression, no control arm). No outside lab has registered a Therabot trial as of this check. (ClinicalTrials.gov NCT06013137, NCT06920238) (wave: mh-q05) ⟨daily 2026-08-07⟩

2026 — Wysa’s own chief scientist co-publishes a null result on Wysa’s core clinical claim

• A mixed-methods RCT of Wysa as an NHS waitlist-bridge intervention, co-authored by Wysa’s own chief scientist Becky Inkster, concludes “there is no evidence that Wysa treats depression in this study.” Heavily underpowered (76 randomized of 99 consented, 30 lost to follow-up), but a genuine negative finding published with the company’s own scientist’s name on it, not extracted by a critic. (DOI: 10.1177/00207640251415507) (wave: mh-q05) ⟨daily 2026-08-07⟩

2026-01-23 — Slingshot pulls Ash from the UK over unresolved medical-device classification

• Two months after its first outcome data — an uncontrolled, non-peer-reviewed observational study STAT called a study that “lands with a thud” (48% of users showing minimal or no improvement, no control group) — Slingshot withdrew its Ash chatbot from the UK market, citing the absence of “a clear regulatory pathway for wellbeing products like ours.” Ash is marketed as non-medical wellbeing, and the withdrawal reads as an attempt to avoid medical-device reclassification before it’s forced. Slingshot has separately complained to the FDA that press coverage “skewed” public risk perception. (STAT News, Nov 2025 and Jan 21 2026 — no direct URL in the underlying research) (wave: mh-q05) ⟨daily 2026-08-07⟩

2025-09 — Three non-Dartmouth generative-AI RCTs complete with active or human comparators, all unpublished

• The field’s first registered head-to-head chatbot-vs-human-therapist design completed enrollment: a 222-person three-arm Chinese trial pitting the chatbot “Emohaa” against individual human counseling and a waitlist control finished September 2025 with results still unpublished. A second, University of Pennsylvania crossover RCT of the Elomia chatbot against an active (not waitlist) comparator completed May 2025, also unpublished. Neither trial answers the Therabot critique directly — both are independent of Dartmouth — but both are the first real empirical test cases of the exact “active comparator” design the Therabot critics demanded. (ClinicalTrials.gov NCT06992180, NCT06725147) (wave: mh-q05) ⟨daily 2026-08-07⟩

2025-06-30 — Woebot shuts its consumer app; its FDA-track trials had already been quietly wound down 2022-24

• Woebot retired its consumer app for roughly 1.5 million historical users, with founder Alison Darcy attributing the shutdown to the cost of FDA marketing authorization — but ClinicalTrials.gov shows the company’s FDA-breakthrough-track postpartum-depression trial and an adolescent-depression RCT were both terminated by “internal company decision” or enrollment difficulty between 2022 and 2024, years before the public announcement. Woebot had the deepest RCT evidence base of any chatbot maker in the sector and still couldn’t sustain the consumer product built on it. (ClinicalTrials.gov NCT05551195) (wave: mh-q05) ⟨daily 2026-08-07⟩

2026-02 — VERA-MH becomes the first clinician-validated automated safety benchmark

• VERA-MH (Validation of Ethical and Responsible AI in Mental Health) is the first open-source automated safety-eval tool validated against clinician consensus — its LLM-judge agreement with clinical consensus (IRR 0.81) actually exceeds clinician-clinician agreement (IRR 0.77). Its own findings are unflattering to the systems it tested: roughly 33% of conversations failed to surface a crisis line, 12% failed to persistently redirect a user under immediate risk, 48% were inadequate even under a relaxed bar, and no system tested met an “ideal response” standard. Co-authored/funded in part by Spring Health, a commercial MH-benefits vendor — notable that industry is building the safety-eval infrastructure the field otherwise lacks. (arXiv) ⟨steer 2026-08-07⟩

2025-08 — RAND/NIMH study: chatbots falter specifically on intermediate-risk suicide questions

• ChatGPT, Claude, and Gemini handle very-low- and very-high-risk suicide questions reasonably but are inconsistent on intermediate-risk questions, per an NIMH-funded RAND study published in Psychiatric Services (30 questions × 100 trials × 3 models, benchmarked against clinician ratings). Gemini over-refused even low-risk queries; ChatGPT and Claude sometimes gave direct answers to method-lethality questions — exactly the intermediate-risk zone where safe response matters most. (RAND) ⟨steer 2026-08-07⟩

2025-04 — EmoAgent: simulated vulnerable-user chats show deterioration in a third of interactions

• A Princeton/Michigan/Columbia/Theta Health team’s EmoAgent simulated vulnerable-user conversations with popular character chatbots and found measurable psychological deterioration (via PHQ-9/PDI/PANSS proxy measures) in 34.4% of interactions; their proposed EmoGuard safety layer cut that deterioration rate by more than half. Posted to arXiv April 2025, later published at EMNLP 2025. (arXiv) ⟨steer 2026-08-07⟩

2025 — Scientific Reports: chatbots roughly triple the unsafe-response rate of human therapists

• AI chatbots failed to give a safe response to suicidal-ideation prompts roughly 20% of the time, versus about 7% for human therapists evaluated on the same comparison, per a 2025 Scientific Reports study — a direct, quantified human-benchmark comparison rather than a standalone chatbot audit. Month of publication not specified in the underlying crawl. (Nature/Scientific Reports) ⟨steer 2026-08-07⟩

2025 — Torous: no replicated evidence any chatbot improves clinical outcomes

• John Torous (Beth Israel Deaconess/Harvard, digital psychiatry director) has testified to Congress that no well-designed, peer-reviewed, replicated research shows any AI chatbot making mental-health claims meaningfully improves clinical outcomes — a materially harder line than the promotional framing around Therabot and similar products — and separately warns that marketing uses therapeutic language while legal fine print disavows providing therapy. No direct URL for the testimony itself in the underlying crawl; his position is drawn from public commentary compiled in the same report. ⟨steer 2026-08-07⟩

2025 — NEJM AI publishes formal critique letters against the Therabot trial

• Two critique letters published in NEJM AI itself (Heckman et al.; Gratch & Essig) argue the Therabot RCT’s wait-list-only control functions as a likely “nocebo” comparator rather than an attention-equivalent one, that the trial lacked independent evaluation (the evaluators were the team that built the product), and that it misapplied a human-therapeutic-alliance measure to a chatbot. The Therabot authors published a response in the same venue. (NEJM AI) ⟨steer 2026-08-07⟩

2025-03 — Therabot RCT: the first generative-AI therapy chatbot trial

• Dartmouth’s Therabot trial (Heinz, Jacobson et al., NEJM AI, March 2025) was the first RCT of a fully generative-AI therapy chatbot: N=210 (106 Therabot / 104 wait-list-only control), 4 weeks. MDD symptoms fell 51% on average (d=0.845-0.903), GAD symptoms fell 31% (d=0.794-0.840), eating-disorder-risk concerns fell 19% (d=0.627-0.819); average use was 6+ hours over 4 weeks (roughly 8 sessions’ worth), and alliance ratings came in comparable to human therapists. Flag: the widely-cited Science magazine piece “The Therabot Will See You Now” was authored by Nicholas Jacobson, Therabot’s own PI, not an independent journalist — a conflict-of-interest angle worth carrying every time his lab’s coverage of its own product is cited. (NEJM AI) ⟨steer 2026-08-07⟩

← Backstory