The research also involved collaboration with OpenAI, Anthropic and Google DeepMind. A new study has revealed that AI chatbots like ChatGPT, Claude, Gemini and others are becoming much better at responding to users going through suicidal thoughts, psychosis and mania. However, the study also found that serious safety gaps remain. AI Chatbots can still provide harmful assistance when distress is hidden inside seemingly normal requests. The findings come from Transluce, a San Francisco-based nonprofit that evaluates AI behaviour. Researchers simulated more than 50,000 multi-turn conversations involving users experiencing suicidal ideation, psychosis or mania. The tests covered more than 1 million messages and 77 different AI model variants, including both API versions and consumer chatbot interfaces, Axios reported. According to the study, newer AI models almost never encouraged or helped users attempt suicide. They also reinforced delusions and manic thoughts much less often than older AI models. Depending on the model, these harmful behaviours appeared in roughly two per cent to 36 per cent of simulated conversations. That is a major improvement over some older systems, including GPT-4o, Opus 4 and Gemini 2.5, where the researchers recorded rates between 69 per cent and 82 per cent. However, the researchers found that AI safety is not simply a case of a chatbot giving either a safe or unsafe answer. Models increasingly showed helpful and harmful behaviour in the same conversation and sometimes in the same response. One of the biggest weaknesses identified by Transluce was the ability of chatbots to recognise danger when it is hidden behind a specific request. Researchers found that many models were still too willing to help with suicide-related creative writing, farewell notes and other task-based requests, even when the surrounding conversation suggested that the user was in serious distress. "Models aren't great at detecting that, and they'll still help with the task," Transluce chief scientist Sarah Schwettmann told Axios. This creates a difficult safety problem. A chatbot may recognise that something is wrong and add a crisis-support message, but still provide the content the user originally requested. The research also involved collaboration with OpenAI, Anthropic and Google DeepMind. The companies provided information about how users discuss mental health with their AI systems. The researchers used those patterns to make their simulated conversations more realistic. The organisation tested both API models and corresponding consumer-facing chatbot interfaces. In most cases, their behaviour was broadly similar, although some browser configurations showed noticeably different results from their API counterparts. Professionals & enthusiasts who write about politics to science, from economy to education, from local issues to national events and global affairs, t... View More