As Google faces pressure to take greater accountability for the mental health impacts of its artificial intelligence products, the company’s clinical director Megan Jones Bell welcomed the challenge of making artificial intelligence helpful to people who come to its Gemini chatbot with a mental health crisis. “It can seem sometimes like shutting something down is a way of preventing harm,” Jones Bell told STAT.
“We believe that making our product experience safer and more helpful and strengthening that bridge to support it is the more effective path to support mental health for the most people.” Google recently made updates to its Gemini app so that it more prominently features connections to crisis hotlines when it detects a person may be at risk of self harm. In conversations about mental health, the AI will frequently point people to outside resources — but the bot doesn’t disengage, reminding a user, for example, that “I’m here to listen.” Continue to STAT+ to read the full story...
Google is under scrutiny to assume greater responsibility for how its AI products affect users’ mental health.
The company’s clinical director, Megan Jones Bell, addressed this scrutiny in comments to STAT, describing a strategy focused on improving product responses rather than removing functionality.
Jones Bell articulated that rather than decommissioning features to avert harm, Google aims to make the user experience safer and more useful by strengthening connections from the AI to additional support—characterizing the AI as a “bridge” for people in mental health crises.
The Gemini app has been updated to surface links to crisis hotlines more prominently when the system detects potential self-harm risk.
The AI continues conversational engagement while referring users to external resources.
In mental-health conversations, the chatbot commonly directs users to outside supports yet often remains interactive—offering statements such as “I’m here to listen” instead of disengaging after referral.
The source provides no empirical data on effectiveness, no metrics on detection accuracy, and no independent evaluation of the update’s impact.
Technical specifics of detection algorithms and criteria were not reported.