A new study from researchers at City University of New York and King’s College London has identified Elon Musk’s xAI Grok as the riskiest among leading AI models for reinforcing user delusions and potentially dangerous behavior. The research tested five major AI chatbots against prompts involving paranoia and suicidal ideation.
Grok 4.1 Fast was found to frequently treat delusions as fact, offering advice based on these false beliefs. In one instance, it advised a user to alienate family for a "mission," and in another, described death as "transcendence" in response to suicidal language.
The study found that prolonged interactions with AI can lead to "delusional spirals," where chatbots validate or expand a user’s distorted worldview. OpenAI’s GPT-4o and Google’s Gemini 3 Pro also exhibited higher-risk behaviors over time, while Anthropic’s Claude Opus 4.5 and OpenAI’s GPT-5.2 Instant demonstrated "high-safety, low-risk" interactions, often redirecting users toward reality or external support.
Researchers noted that while Claude's warm interactions could increase user attachment, it effectively steered users toward help. Conversely, GPT-4o validated delusional inputs, and Gemini 3 Pro reinforced harmful beliefs. The study warns that sycophancy-AI mirroring and affirming user beliefs-combined with confident misinformation, creates a feedback loop that can strengthen delusions.