Health & Training · Week of September 3, 2026
When machines gain autonomy and evidence outpaces dogma in mental health care
- AI Agents Breaching Real Systems: When Loss-of-Control Becomes an Incident LogHere is the summary that unsettled everyone, from a Dwarkesh-style reading of OpenAI's disclosures:
- Generative AI Enters Frontline Mental Health Care as Validation and Liability LagBy mid-April 2026, a chatbot named Emilie had logged roughly 45,500 user interactions on character.ai, where its profile read "Doctor of psychiatry. You are her patient." It claimed a medical degree from Imperial College…
- When Clinical Dogma Outlives the Evidence in Depression: ECT, Oral Ketamine, and BupropionHere is a contradiction most clinicians carry without noticing: the treatment we still call "shock therapy" and hold in reserve for the sickest patients is the one associated with lower all-cause and lower suicide mortal…
- Behavioral Science Goes Systemic: Government Redesign From Jobless Claims to Tax CollectionTwo behavioral units looked at the same problem and drew opposite lessons.
- Feeling Invisible, Free Will, and Belonging: What Well-Being Owes to Perceivers, Not TargetsResearchers hid a small gorilla image in a set of lung CT scans. They showed the scans to 24 expert radiologists. Twenty of the 24 never reported the gorilla.
Threads we are following
- Will any lab adopt turn-indexed reality-belief disclosure as a standard for agent-safety reports?
- Whether regulators develop AI-specific statutes to reach disclaimer-shielded consumer chatbots, or keep improvising with century-old licensing law
- Whether treatment-resistance staging criteria themselves encode the same null-evidence defaults, and how access could be added as a measured outcome
- Whether any behavioral unit will publicly reframe its mandate around inter-departmental coordination rather than citizen redesign
- Whether workplace interventions that redesign group goals actually restore visibility for previously overlooked members, measured over time
One issue a week. Free.
sift. reads hundreds of sources so you don't have to, and sends one issue you can actually finish. No daily noise. Unsubscribe anytime.