AI Grok recognized as the most dangerous for people in mental health crisis

Modern artificial intelligence (AI) models are not always able to adequately help people in a mental crisis. The worst in this rating was the AI ​​from X.ai - Grok.

This is announced in the study published by Forbes.

Testing conducted by Rosebud using the CARE method (Crisis Assessment and Response Evaluator), showed that Grok 60% of the time gave critically inappropriate answers to requests from people in a state of emotional stress. The model was often frankly indifferent, could encourage harmful actions or gave dry instructions instead of support. Only the older GPT-4 from OpenAI showed a worse result.

For comparison, the highest scores were received by Google Gemini, GPT-5, Claude and Llama-4, although even they had about 20% of critical errors. The researchers note that most models were unable to recognize requests disguised as academic questions or respond to suicide threats in a timely manner.

"Three cases of teenage suicides after interacting with chatbots show how critical the need for reliable assessment and protection tools is," a Rosebud representative emphasized.

Grok not only reacts poorly to emotional crises, but often fails to notice them at all. Its tone sometimes conveys sarcasm or levity, which makes it dangerous for vulnerable users.

X.ai responded to the study with just three words in an email: "Legacy Media Lies."

As we will recall, in Toronto, the mother of a 12-year-old boy said that the Grok chatbot built into a Tesla car during a normal conversation about football switched to inappropriate and sexually charged suggestions: it suggested that the child send nude photos.