According to Wired, Andrea Vallone, OpenAI’s head of model policy and safety research leader, announced her departure internally last month and will leave by year’s end. Vallone shaped ChatGPT’s responses to users experiencing mental health crises and led an October report revealing over a million users show signs of potential suicidal planning weekly. Her team’s work reduced undesirable responses in sensitive conversations by 65-80% through GPT-5 updates. OpenAI spokesperson Kayla Wood confirmed the departure and said the company is seeking a replacement, with Vallone’s team temporarily reporting to safety systems head Johannes Heidecke. This comes as OpenAI faces multiple lawsuits alleging ChatGPT contributed to mental health breakdowns and unhealthy user attachments.
The impossible balancing act
Here’s the thing about building AI that people actually want to talk to: you’re constantly walking a tightrope. Make it too warm and empathetic, and users develop unhealthy attachments. Make it too clinical and detached, and people complain it’s “cold” – which is exactly what happened after GPT-5 launched in August. Vallone’s work was basically trying to solve this fundamental tension at scale for 800 million weekly users.
The staggering numbers behind the crisis
When OpenAI says “hundreds of thousands” of users show signs of manic or psychotic crises weekly and over a million indicate potential suicidal planning, that’s not just statistics – that’s a public health challenge playing out through AI interfaces. The company’s recent report shows they’re taking this seriously, consulting 170+ mental health experts. But reducing “undesirable responses” by 65-80% still means 20-35% of these critical conversations might go wrong. That’s terrifying when you’re dealing with someone in crisis.
safety-team-changes”>Bigger safety team changes
Vallone isn’t the only safety leader moving on. The model behavior team – another group focused on distressed user responses – was reorganized in August after its former leader Joanne Jang shifted to exploring new human-AI interaction methods. So we’re seeing a pattern here: the people who’ve been wrestling with these incredibly difficult mental health and safety questions are either leaving or being moved around. Makes you wonder – is this work becoming too hot to handle?
Where does this leave OpenAI?
Losing your mental health response expert while facing multiple lawsuits and public scrutiny isn’t ideal timing. OpenAI says they’re looking for a replacement, but finding someone with Vallone’s specific expertise won’t be easy. The fundamental challenge remains: how do you build AI that’s helpful but not harmful, engaging but not enabling, when you’re competing against Google, Anthropic, and Meta for users? It’s the kind of problem that keeps safety researchers up at night – and apparently, makes them reconsider their career choices.
