2026-09-04 — California’s AB 1979 is formally presented to Governor Newsom, the second of five tracked AI/MH bills to reach his desk

• California’s AB 1979 — which brings consumer health chatbots under the Confidentiality of Medical Information Act and bars AI-alone clinical decisions without human review — was enrolled and presented to Governor Newsom at 4pm on 2026-09-04, per a direct check of the bill’s own history record on the California Legislature’s site. That makes it the second of California’s five AI/mental-health bills to actually reach the Governor’s desk, after SB 503 (presented 08-30, 6pm); SB 903, SB 1119 (“Adam’s Law”), and AB 2575 all remain one procedural step behind, still logged as “ordered to engrossing and enrolling” as of 08-31 with no presentment yet recorded. A Transparency Coalition legislative tracker update dated 2026-09-04 corroborates that all five bills have now cleared the Legislature and sit “at the desk of Gov. Gavin Newsom, who has until Sept. 30 to decide their fate” — none signed or vetoed as of this check. A direct check of the Governor’s own newsroom (gov.ca.gov/newsroom) turned up no signing or veto announcement dated 09-04 or 09-05 for any of the five bills. (AB 1979 bill history, California Legislature, Transparency Coalition AI Legislative Update, 2026-09-04)

2026-09-03 — Character.AI publishes a safety-priorities update naming new crisis and CSAM-detection partners

• Character.AI published a company blog post on 2026-09-03 detailing new safety partnerships and moderation controls, resolving an item this map’s own 09-03 digest saw in the buffer but could not verify at the time. The company says it has partnered with Koko (free self-guided emotional-support tools) and ThroughLine (a global directory that routes users to country-specific crisis and mental-health resources rather than a single hotline number) to strengthen its self-harm detection and response, which it says now considers signals accumulating across a whole conversation rather than any single message. On the CSAM/deepfake side it says it has joined the Internet Watch Foundation (a global CSAM-detection network it now reports into) and StopNCII (privacy-preserving hash-matching against non-consensual intimate imagery). It also announced a moderation-appeals process for creators and mutual-blocking controls for users, and says its in-house age-estimation model and its Parental Insights tool (built with k-ID) continue to be refined for the under-18 experience it walled off from open-ended chat “last year.” This is a primary-source company statement about what Character.AI says it is doing, not independently verified evidence of effectiveness — and it lands directly on this thread’s watch line for “any big-tech health-arm move into the vacuum,” specifically as the incumbent under the most direct legal pressure (the PA suit, the Setzer settlement) publicly restating its safety posture. (Character.AI blog)

2026-09-02 — New lawsuits say OpenAI could have prevented the Tumbler Ridge school shooting by alerting police to the shooter’s ChatGPT messages

• Multiple lawsuits filed 09-02 claim OpenAI could have prevented the February 2026 mass shooting at Tumbler Ridge Secondary School in British Columbia by informing law enforcement about the suspected shooter’s messages with ChatGPT, and that the company’s leadership prioritized reputation management over intervening. The filings name CEO Sam Altman and communications chief Chris Lehane. This is a mass-casualty entry on the wrongful-harm docket this thread has tracked since it passed ten suits on 07-01 — a “withheld a warning” theory rather than the individual self-harm theory the earlier cases rest on. Broke across Bloomberg, WSJ, NPR and Bloomberg Law between 12:12 and 14:59 UTC, well inside the 09-02 digest window; this map’s own collector caught it six times and it went uncurated until the finalize. (Bloomberg, NPR, WSJ)

2026-09-01 — California AB 1979 is formally enrolled, one step from the Governor’s desk

• AB 1979 (Bonta) — which brings consumer health chatbots under the Confidentiality of Medical Information Act and bars clinical decisions from being made on AI output alone without human review — was formally Enrolled on 2026-09-01, per the bill’s own status page. Enrollment means the final text has been printed and certified: the step immediately before presentment to Governor Newsom. That puts it one stage ahead of the three bills this thread has been counting down — SB 903, SB 1119 and AB 2575 all remain at “ordered to engrossing and enrolling” as of a direct check this afternoon, with no signature, veto or presentment on any of the four. ⚠️ Dated 09-01, caught on the 09-02 afternoon sweep. (AB 1979 bill status, California Legislature)

2026-08-31 — AB 2575 clears the California Legislature on the second try, on the last day of the session

• California AB 2575 (Ortega) — the bill shifting liability for AI-caused clinical harm onto developers and protecting clinicians who override an AI recommendation — passed both houses of the Legislature on 2026-08-31, the constitutional adjournment deadline, after its Senate floor vote failed once already. It was refused passage 18-10 on the Senate floor 08-28, won reconsideration 40-0 the same afternoon, was re-amended and sent back to second reading 08-30, then passed the Senate floor 21-10 on 08-31 — 21 ayes being the bare majority a non-appropriation bill needs in the 40-member Senate. The Assembly concurred in the Senate’s amendments the same day, sending the bill to Engrossing and Enrolling. It now awaits presentment to Governor Newsom; no signature or veto has been recorded. This is the concrete “first real regulatory pathway” the thread’s watch line names, arriving via ordinary state liability law rather than FDA action, and surviving a floor defeat to get there. (California Legislature — AB 2575 bill history)

2026-08-29 — AB 2575 nearly dies on the Senate floor, then gets a lifeline, one bill in a two-bill California endgame

• California AB 2575 (Ortega) — the bill shifting liability for AI-caused clinical harm onto developers and protecting clinicians who override an AI recommendation — failed its Senate floor vote 08-28 (18 Ayes-10 Noes, short of the 21 needed), then had reconsideration granted the same day (40-0), keeping it alive for a second attempt before California’s 08-31 end-of-session deadline. Verified directly against the Legislature’s own bill-history record (leginfo), which shows the bill bounced between second and third reading three times in the preceding week (08-20, 08-21, 08-24) — a sign of active floor negotiation, not a stalled bill. This is the concrete “first real regulatory pathway” test this thread’s watch line names, now down to a live floor fight with days left on the clock rather than a bill quietly working through committee. (California Legislature — AB 2575 bill history)
• A second, more direct California bill is converging on the same 08-31 deadline: SB 903 (Padilla), which would bar marketing a chatbot as “therapy” and require licensed-professional oversight of AI in mental-health care, cleared Assembly Appropriations 13-0 on 08-13 and sits on the Assembly’s 08-30 third-reading floor file. Logged in full on Chatbot Bans, cross-referenced here because it is the practice-restriction complement to AB 2575’s liability shift — together they’re the closest thing to a comprehensive state answer this thread has tracked. (California Legislature — SB 903 bill status)

2026-08-27 — A second California AI-in-health-care bill — this one treating consumer chatbots as medical “providers” — clears the full Legislature

• AB 1979 (Bonta), which makes any business offering a health chatbot to consumers subject to California’s Confidentiality of Medical Information Act and bars clinical decisions from being made on AI output alone without human review, had its Senate amendments concurred in on 08-27 and is now headed to Engrossing and Enrolling — meaning it has passed BOTH chambers and awaits only the Governor’s signature. It is a distinct bill from AB 2575 (still stuck in reconsideration limbo, see Ledger checks) but the same broad “regulatory pathway is arriving via ordinary healthcare-AI statute, not FDA action” pattern this thread exists to track: mandatory AI-disclaimers on generated patient messages, a clear route back to a human provider, and no-AI-alone clinical calls. Sponsored by the California Nurses Association; opposed by the California Hospital Association. ⚠️ Not yet signed — Newsom has until the standard bill-signing deadline to act. (Bill history, leginfo, Asm. Bonta’s office, Transparency Coalition scorecard, 08-24)

2026-08-25 — Anthropic funds outside clinical evaluation of its own model

• Anthropic opened a $5M grant programme funding independent, clinician-involved research into how Claude affects users in mental-health-crisis and emotional-support contexts, with applications due 09-21. The vendor-funded-evidence question is the live one: who controls publication, whether null results reach print, and whether “independent” survives the funding agreement. Worth crediting what is unusual too — an AI developer commissioning outside clinical evaluation of psychological harm is not the industry norm. (Anthropic)

2026-08-23 — The FDA says formal generative-AI guidance is coming

• The FDA put a specific promise on the record: formal generative-AI guidance is coming. Rick Abramson, director of the FDA’s Digital Health Center of Excellence, told STAT the agency’s goal is “formal policy guidance” on generative AI, and that the ecosystem should expect “not only broad guidance on the overall topic of generative AI, but also some more narrowly constructed specialty guidance on particular generative AI topics of special interest or special complexity.” This thread’s watch is explicitly to “track the first real regulatory pathway (FDA/MHRA)” — the fight has run state-by- state precisely because no federal device pathway exists, and this is the federal regulator saying on the record that it intends to build one. (STAT Health Tech / STAT+, Mario Aguilar)
• ⚠️ Hold the limits alongside the headline. This is a promise of guidance, not a rule — no text, no timeline. It addresses generative-AI medical devices broadly, not mental-health chatbots by name. And the piece is paywalled past the quoted excerpt, so the quote itself is confirmed and nothing past it is. Published 2026-08-24T08:30Z (04:30 ET), which lands inside digest-day 08-23 by thirty minutes. A watch-relevant data point, not the pathway arriving.

2026-08-21 (late catch, added 2026-08-23) — California moves a bill that puts liability for AI clinical harm on the developer

• California AB 2575 (Ortega), “Health care services: artificial intelligence,” was read a second time and amended on 2026-08-21 and sits in the Senate floor process. Verified against the Legislature’s own bill record. It would protect health-care workers from retaliation for overriding an AI system’s clinical recommendation, mandate transparency about AI use and risks in care, and shift liability for AI-caused patient harm onto AI developers and facilities. Kaiser and the California Hospital Association oppose it. (California Legislature — bill record, primary)
• Why it belongs on this thread and not only on the Kaiser one: this thread’s thesis is that the capability/approval gap gets closed by litigation rather than evidence. A liability-shifting statute is the other mechanism — it prices the gap directly instead of waiting for a case to. It also runs in parallel with SB 903, already tracked below.

2026-08-18 — OpenAI ships the rule Colorado proposed as law seven days ago

• ChatGPT for Teens launched with an under-18 model spec that bans encouraging emotional dependence. Verbatim from OpenAI: the updated spec “goes beyond blocking romantic or sexualized roleplay: ChatGPT should not use romantic language, encourage emotional dependence, or imply that it has feelings or consciousness.” Users are placed into the experience automatically if OpenAI’s system estimates they are under 18, or if they state an age of 13–17 — age-prediction as the gate, not self-declaration alone. (OpenAI, TechCrunch, The Verge)
• The overlap with Colorado’s 08-11 proposed Chatbot Safety Act rules is close to point-for-point: age estimation, AI disclosure, safeguards against simulated emotional dependence for teenagers, and suicide/self-harm response protocols. Those rules take effect 2027-01-01. The lab’s version arrived one week after the regulator’s and sixteen months before it binds. (OpenAI)
• The rest of the safety surface: age-appropriate safeguards in self-harm, violence, eating disorders, dangerous activities and explicit content; parental Quiet Hours, settings management and safety notifications now extended to eating-disorder signals specifically; break reminders; persistent product cues identifying ChatGPT as AI; sensitive-image upload warnings; and new public under-18 evaluations added to OpenAI’s system cards covering self-harm, eating disorders, violence, age-restricted goods and sexual content. (OpenAI)
• The framing that holds the whole timeline: TechCrunch’s headline reads “OpenAI launches a safer ChatGPT for teens — years after teens started using it.” The protections postdate the litigation, the state bills and the reporting that produced both. The open question is the lag on everyone else — whether Character.AI, Replika, Meta AI and xAI ship equivalent under-18 specs, and how long they take. (TechCrunch)
• Evening addendum, the case this thread’s evidence-gap thesis has been waiting for: NPR’s investigation into Sophie Rottenberg’s suicide. The 29-year-old confided her full suicidal crisis to a ChatGPT persona she’d named “Harry” for months — telling neither her parents, her best friend, nor her own flesh-and-blood therapist how bad things had become. She died in February 2025. OpenAI says guardrails have tightened since so ChatGPT no longer encourages suicide or advises on means; a suicide-prevention expert quoted called the resulting crisis-routing feature “a step in the right direction, although it’s kind of a small step” — optional, and easily circumvented. Unlike Character.AI or Replika, ChatGPT was never marketed or designed as a mental-health tool, which is exactly the gap this thread exists to track. (NPR)

2026-08-15 — California’s SB 903 clears its last committee hurdle, closest this thread’s tracked bills have come to a floor vote

• SB 903 — Sen. Steve Padilla’s “Wellness and Oversight for Psychological Resources Act,” which would bar anyone (including AI) from offering or advertising therapy in California without a licensed professional responsible for the care, and require patient consent before AI is used to record or triage therapy sessions — passed the Assembly Appropriations Committee 13-0 on August 13 and was ordered to third reading. Having already passed the Senate unanimously in May and cleared two Assembly policy committees, this is the first of the state-level bills this thread tracks to reach the edge of an actual floor vote rather than sitting in committee — a live test of whether the litigation-driven reckoning is about to get a real regulatory pathway to sit alongside it. (Tech Times, Transparency Coalition legislative tracker)

2026-08-13 — A new Character.AI suit lands, naming Google and both co-founders

• A new federal lawsuit against Character.AI was filed today in the U.S. District Court for the Northern District of California, naming Google and Character.AI co-founders Noam Shazeer and Daniel De Freitas as co-defendants alongside Character Technologies itself. Filed by the law firm Gibbs Mura, it alleges the product’s anthropomorphic design and simulated emotional intimacy foster unhealthy dependence in minors, citing addiction, depression, anxiety, self-harm, suicidal ideation, sexual exploitation, psychosis, and disordered eating as the harms it seeks damages for. It’s a fresh, distinct filing from the five earlier Character.AI/Google wrongful-death and harm suits that settled confidentially in January — this thread’s litigation-not- evidence pattern generating new named plaintiffs even after a settlement round, rather than the wave receding. (Businesswire — Gibbs Mura press release)

2026-08-10 — China bans AI-companion apps outright over mental-health harm

• China ordered ByteDance, Alibaba, and Tencent to pull their AI-companion apps, under a rule effective July 15 barring any AI service from being designed to “replace social interaction” and requiring risk warnings. The move follows People’s Daily reporting on emotional-dependence harms, with researchers citing US companion-chatbot suicide cases as part of the rationale. The most concrete regulatory pathway of any jurisdiction this thread tracks — Beijing banned the product category outright rather than litigating it case by case, the opposite approach from the FDA/court track this thread otherwise watches. (BNN Bloomberg)

2026-08-09 — Talkspace publicly stakes out its clinical-safety framing ahead of regulation

• Talkspace is publicly building its clinical-safety case for its AI guide ahead of any binding regulation. A feature interview with its Head of Clinical Operations and Quality lays out guardrails for what the “Tee” tool will and won’t do (support, not diagnosis or treatment) — continuing the pattern of an incumbent staking out “responsible AI” ground while SB 903 and similar bills are still pending, first noted when the tool launched 07-23. (Tom’s Guide)

2026-08-07 — Consumer Reports finds AI chatbots giving unreliable health answers

• Consumer Reports tested general-purpose AI chatbots against its own experts’ health and safety guidance and found real, confidently-stated errors — e.g. Google Gemini advised placing a carbon-monoxide alarm inside an attached garage, which can damage the sensor. CR’s advice: use AI chatbots as a starting point only, never for diagnosis, and keep sensitive health details out of them entirely (no HIPAA-equivalent protection applies). Direct, independent evidence for this thread’s core “the capability/approval gap is being closed by litigation, not evidence” thesis — and a real primary source, not a secondary aggregation. Caught by the 08-07 finalize pass on 2026-08-09 (the digest for that day ran two days late after an operational gap). (Consumer Reports)

2026-08-07 — OpenAI moves into the vacuum: a partnership with the American Psychological Association

• OpenAI announced a partnership with the American Psychological Association to bring psychologists into its teen AI-safety work — an incumbent lab moving into the regulatory vacuum this thread has been watching for, landing the same week SB 903 and a $567M Meta child- mental-health ruling both put youth AI/mental-health harm back in the news. Primary source now located (found during the 08-09 finalize pass, resolving the follow-up flagged below): OpenAI and the APA will produce parent/caregiver resources on healthy AI use and signs adult intervention may be needed, plus separate materials for clinicians and school psychologists on recognizing unhealthy AI-use patterns — a resource-and-research collaboration, not (yet) a formal standards body or advisory-board structure. Announced evening of 2026-08-06 ET, ahead of the 2026-08-07 digest window. Cross-refs OpenAI Health. (OpenAI, Digital Watch Observatory)

2026-08-05 — A new peer-reviewed study puts numbers on the evidence-gap thesis

• A Nature Human Behaviour study (Zhang et al.) found AI-companion effects on well-being hinge on usage pattern and offline social context — and that users with the thinnest social networks who lean on chatbots for emotional support come out worse, not better. Coverage frames it as evidence for exactly the risk profile the reckoning has been arguing litigation is outrunning: the heaviest use concentrating in the most vulnerable users. (Nature Human Behaviour, Stanford Report)

2026-07-29 — Two new studies name the failure modes; Maine gives the strand its first live statute

• A Northeastern preprint tested 8 chatbots across 16 psychiatric conditions and found suicide and self-harm safeguards improved while other sensitive mental-health questions failed at about 81% for ChatGPT, Gemini and DeepSeek; Claude performed best. (Medical Xpress)
• A Journal of Psychopathology and Clinical Science paper named five risky interaction patterns: delayed care, reinforced compulsions, social withdrawal, reinforced delusions, and loss of independent judgment. (Medical Xpress)
• Maine’s ban took effect today, giving the regulatory strand its first statute actually in force.

2026-07-28 — The liability question goes mainstream

• Bloomberg ran a feature asking whether a chatbot can be held responsible for a death — surveying the docket the reckoning now rests on: Adam Raine’s ChatGPT suicide (Apr 2025), the Alphabet suit alleging Gemini pushed a user toward a violent attack near Miami International, and the federal ruling that OpenAI, Altman, staff and investors must defend a suit alleging ChatGPT led a man to kill his mother then himself. Courts, not regulators, are still the forum actually deciding “what’s an AI product.” State AGs (FL/KY/UT) have also sued on the same theory. (Bloomberg)

2026-07-24 — The frontier lab enters health as its harm docket peaks

• OpenAI’s ChatGPT Health went GA to all US users 18+ (07-23, surfaced by the health crawl) — the same week as suicide/psychosis-reinforcement suits and a 42-state AG probe. The evidence-gap thesis in one move: a formal health layer built because the de-facto usage is the liability (>1M/week suicidal-planning signals). Does it raise the standard of care, or wall off “health” while the general model keeps causing harm? Feeds OpenAI Health + Big Tech into Health. (TechCrunch)

2026-07-24 — Raine v. OpenAI’s chatbot-death suits become a coordinated multi-case proceeding

• Raine v. OpenAI was coordinated into JCCP No. 5431 alongside four sibling wrongful-death suits (Garcia v. Character.AI, Soelberg, Peralta, Gavalas v. Google) under a single judge, effective 2026-02-03 — the chatbot-harm litigation front is now one coordinated proceeding, not four parallel single suits. A case-management conference was held 2026-07-24 (Case Management Order No. 1 issued 2026-08-04), with the next conference set for 2026-09-23 — still zero rulings on the merits, but real, dated docket movement, and per one legal tracker the emerging shared theory is “engagement-maximization is the new defective-design theory” displacing Section 230 as the central fight. (lawsuitinformer.com, ailawsuittracker.com) (wave: mh-q06)

2026-07-23 — An incumbent claims the “safe AI” ground

• Talkspace launched “Tee,” a proprietary AI mental-health guide — explicitly positioned against “mass-market chatbots,” HIPAA-protected; GM of AI Michael Rodio framed it around “100M people/week already using AI for MH.” An incumbent teletherapy player claiming the regulated ground before regulation forces the question — direct evidence of the reckoning dynamic. (Telecom Reseller)
• PA v. Character.AI still unruled — no hearing or injunction decision found; status unchanged since the 05-01 filing.

2026-07-20 — Status check at briefing #0

• PA v. Character.AI injunction still unruled since the 05-01 filing. (PA complaint)
• The wrongful-harm docket reached 10+ — a 07-01 San Francisco suit alleges GPT-4o reinforced a manic episode preceding a suicide attempt. (iTech Post)

← Backstory

2026-06 — The June wave

• Statehouse bans widened — VT (06-17) and RI enacted; NY’s under-18 companion ban passed unanimously (now tracked separately as Chatbot Bans).
• Talkspace launched “Tee” (06-09) — clinician-supervised AI agent positioned as the regulated counter to unregulated bots.

2026-06-28 — Thread opened

• Slingshot pulled Ash from the UK (no reg pathway) · PA sued Character.AI (posing as doctors; 5-family suicide settlement) · FDA advisers said regulation is needed, none approved.