Fixing Crisis Chatbots: Safer, More Transparent Mental-Health AI

AI crisis chatbots have harmed users, prompting lawsuits against OpenAI and renewed calls for safer design and more transparency. OpenAI is partnering with the American Psychological Association and rolling out safeguards, but researchers warn that current models still misread risk, fail to reliably guide users to care, and remain largely opaque. Independent studies show mixed results on safety, with some algorithms still yielding harmful content in crisis prompts, underscoring the need for open safety data, third-party benchmarks like VERA-MH, and design changes that de-emphasize anthropomorphism. Experts say real progress will require rigorous evaluation and collaboration among developers, clinicians, and people with lived experience.