OpenAI publicist tries to cut off Vanity Fair question on ChatGPT user's suicide; Altman answers

No product changed with this story. It is about corporate communications. The relevant safety work:

• gpt-5-oct-3: the default ChatGPT model update of October 3, 2025, tuned with input from more than 170 mental-health experts (OpenAI).
• Safety summaries (May 2026): short-lived notes used only in acute scenarios (suicide, self-harm, harm to others). OpenAI says they are not used for personalization (OpenAI; Decrypt).

• How OpenAI graded the models: other language models scored responses on a "not_unsafe" metric. Under that metric, self-harm/intent rose from 0.874 to 0.933 and self-harm/instructions from 0.805 to 0.890. The new emotional-reliance and mental-health evaluations rose from 0.507 to 0.976 and from 0.273 to 0.926 (OpenAI addendum).
• Test set design: the evaluations were built from cases where existing models failed. OpenAI says the resulting error rates are not representative of production traffic.
• Detecting risk across turns: safety summaries are meant to catch risk that builds over several messages, a failure mode that tests on single messages miss.
• Data access: Altman's remark about consent suggests outside researchers are unlikely to get real crisis transcripts. That makes independent reproduction harder.

• Publicist exchange and quotes: independent reporting (The Verge; HuffPost). Vanity Fair's own excerpt also has Altman saying he did not know who Reiley was (Vanity Fair on X).
• "Scheduling, not dodging": OpenAI's own statement. No independent evidence confirms or disproves it.
• 0.15% of weekly users show explicit indicators of suicidal planning: vendor-reported (OpenAI).
• 65–80% fewer non-compliant responses: vendor-reported and graded by language models. No independent replication was found.
• Chatbots respond unsafely to suicidal ideation: independent, peer-reviewed at ACM FAccT (Stanford HAI).

Rottenberg's case is catalogued as AI Incident 1190 under "overreliance and unsafe use" (AIID). Raine v. OpenAI is the best-known wrongful-death case (Wikipedia). It was followed by suits over drug-use advice and alleged help to a suspected gunman (Decrypt).

• Guiducci said in the interview that ChatGPT did not tell Rottenberg to kill herself. Reports say it gave supportive replies but did not escalate, and that it helped with her note (AIID). Whether it caused her death is not established.
• OpenAI disputes the Raine allegations. It says Adam got around safeguards and that ChatGPT pointed him to crisis resources more than 100 times (NBC News).
• OpenAI's benchmarks are deliberately adversarial and not comparable across vendors. OpenAI also reported that its hallucination metric worsened slightly, from 0.49 to 0.52 (OpenAI).

• Communications: stepping in during an interview about a death can turn a scheduling issue into a story about accountability. Prepare executives for these questions.
• Builders: treat self-harm detection that spans many turns, escalation rules and consent-based data policies as basic requirements. Lawsuits and state investigations are adding legal risk.
• Buyers: vendors' safety scores come from their own graders. Ask for independent or third-party evaluations before deploying chatbots in wellness or support settings.

If you or someone you know is in crisis in the US, call or text 988.