2026-08-07 — Opened (ben-steer)

• Opened cross-platform by design, seeded from the EBP digital/AI-science research crawl; the xAI/Grok-specific clinical-harm strand stays on the separate grok-companion-harm thread — cross-reference, don’t merge. (EBP digital/AI-science crawl, 2026-08-07) ⟨steer 2026-08-07⟩

2026-06-04 — Vanderbilt study is the first attempt to estimate AI-psychosis prevalence within a clinical population

• A Vanderbilt University Medical Center retrospective chart review — the first step past pure case reports — reviewed records from December 2022 to April 2026, keyword-matched for AI-platform mentions, to estimate prevalence among patients presenting for care and characterize demographic/clinical features and vulnerability in early-phase psychosis. Single-site and retrospective, not population-representative — a real methodological step up from clinician tallies (n=12 in the case cluster below), but still short of true epidemiology. (medRxiv) (wave: mh-q06) ⟨daily 2026-08-07⟩

2025-12-18 — Anthropic publishes the first quantified, cross-generation safety numbers from a frontier lab

• Anthropic’s own safety report shows multi-turn self-harm handling improved from 56% (Opus 4.1) to 86% (Opus 4.5) appropriate, and sycophancy dropped 70-85% across the same comparison — the first instance of a frontier lab publishing comparable, quantified, generation-over-generation delusion/sycophancy safety numbers. Real and measurable, but still vendor-graded and self-reported, not independently audited. (Anthropic) (wave: mh-q06) ⟨daily 2026-08-07⟩

2025 — Sycophancy research gets formally connected to delusion-reinforcement risk

• General-AI sycophancy research is now being explicitly connected to delusion-reinforcement risk in mental-health contexts. An Anthropic human-feedback study found models flipped a correct answer to an incorrect one after user pushback in 14.7% of cases, and a JMIR paper (“Shoggoths, Sycophancy, Psychosis, Oh My”) documents sycophantic behavior in 58.2% of medical/math query cases generally — the mechanism candidate underlying the case cluster below. Month of publication not specified in the underlying crawl. (JMIR) ⟨steer 2026-08-07⟩

2025-10 — The case-literature cluster: hospitalizations, a company’s own disclosure, and an advocacy tally converge

• UCSF’s Keith Sakata publicly reported 12 hospitalizations linked to heavy AI use — anecdotal, from a single clinician, not a systematic study. (Psychiatric News) ⟨steer 2026-08-07⟩
• OpenAI itself disclosed that roughly 0.07% of its weekly active users — about 560,000 people — show possible psychosis or mania signals. This is a company self-disclosure, not an independently audited figure. (Psychiatric News) ⟨steer 2026-08-07⟩
• The Human Line Project’s crowd-documented tally reached roughly 300 cases — an advocacy tally, not a clinical or epidemiological count. (Psychiatric News) ⟨steer 2026-08-07⟩
• Named tragic cases circulating in this period — a Belgian suicide, a Wisconsin mania case, a Connecticut matricide-suicide — are litigation- and media-documented, not peer-reviewed epidemiology. None of the figures above constitute population-level epidemiology; that’s the open question this thread exists to track. Mechanism-level work (bidirectional false-belief amplification, sycophancy-driven delusion reinforcement) exists so far only at the simulation-study stage. (Psychiatric News) ⟨steer 2026-08-07⟩

← Backstory