A Research Leader Behind ChatGPT’s Mental Health Work Is Leaving OpenAI

An OpenAI safety probe person who helped signifier ChatGPT’s responses to users experiencing intelligence wellness crises announced her departure from the institution internally past month, WIRED has learned. Andrea Vallone, the caput of a information probe squad known arsenic exemplary policy, is slated to permission OpenAI astatine the extremity of the year.

OpenAI spokesperson Kayla Wood confirmed Vallone’s departure. Wood said OpenAI is actively looking for a replacement and that, successful the interim, Vallone’s squad volition study straight to Johannes Heidecke, the company’s caput of information systems.

Vallone’s departure comes arsenic OpenAI faces increasing scrutiny implicit however its flagship merchandise responds to users successful distress. In caller months, respective lawsuits person been filed against OpenAI alleging that users formed unhealthy attachments to ChatGPT. Some of the lawsuits assertion ChatGPT contributed to intelligence wellness breakdowns oregon encouraged suicidal ideations.

Amid that pressure, OpenAI has been moving to recognize however ChatGPT should grip distressed users and amended the chatbot’s responses. Model argumentation is 1 of the teams starring that work, spearheading an October study detailing the company’s advancement and consultations with much than 170 intelligence wellness experts.

In the report, OpenAI said hundreds of thousands of ChatGPT users whitethorn amusement signs of experiencing a manic oregon psychotic situation each week, and that much than a cardinal radical “have conversations that see explicit indicators of imaginable suicidal readying oregon intent.” Through an update to GPT-5, OpenAI said successful the study it was capable to trim undesirable responses successful these conversations by 65 to 80 percent.

“Over the past year, I led OpenAI’s probe connected a question with astir nary established precedents: however should models respond erstwhile confronted with signs of affectional over-reliance oregon aboriginal indications of intelligence wellness distress?” wrote Vallone successful a station connected LinkedIn.

Vallone did not respond to WIRED’s petition for comment.

Making ChatGPT enjoyable to chat with, but not overly flattering, is simply a halfway hostility astatine OpenAI. The institution is aggressively trying to grow ChatGPT’s idiosyncratic base, which present includes much than 800 cardinal radical a week, to vie with AI chatbots from Google, Anthropic, and Meta.

After OpenAI released GPT-5 successful August, users pushed back, arguing that the caller exemplary was amazingly cold. In the latest update to ChatGPT, the institution said it had importantly reduced sycophancy portion maintaining the chatbot’s “warmth.”

Vallone’s exit follows an August reorganization of different radical focused connected ChatGPT’s responses to distressed users, exemplary behavior. Its erstwhile leader, Joanne Jang, near that relation to commencement a caller squad exploring caller human–AI enactment methods. The remaining exemplary behaviour unit were moved nether post-training pb Max Schwarzer.