What was built, what it did to people, and what changed in law and markets. Filed against one framework, with every claim graded for how good the evidence actually is.
The week's evidence converges on the handoff (the moment a system is meant to pass a distressed person to a human) and measures it missing or inert where models and products are asked to perform it…
Read the brief →California's 30 September action split the week's policy signal in two: AB 1979 was signed as Chapter 854, placing a statutory duty on facilities to preserve a licensed clinician's independent…
Read the brief →The instruments for judging AI mental-health safety were written and published this week by the companies being judged: OpenAI released MentalHealthBench and grades every response on it with its own…
Read the brief →The week's dominant pattern is a convergence on AI companions.
Read the brief →The week's connective tissue is oversight architecture: who has to check AI, at what point, and against what evidentiary bar.
Read the brief →Rule-making and litigation moved on general-purpose chatbots as mental-health actors in the same week that the strongest new evidence described association and uncertainty rather than benefit…
Read the brief →The model layer this week was almost entirely measurement work, and most of it reported that the instruments are unstable…
Read the brief →This week the causal chain is unusually complete, and its connective tissue is the validity lens.
Read the brief →This week's filed items cluster on the user-to-policy edge: three of the four EFF.user items…
Read the brief →The week's dominant pattern is convergence on conversation length as the locus of harm: OpenAI's GPT-5.6 system card ships dynamic multi-turn self-harm evaluations and discloses a regression on them…
Read the brief →The week's dominant pattern is the unsupervised front door getting its first hard denominator at the same moment governments on three continents moved to draw perimeters around it.
Read the brief →