Elon Musk’s Grok Most Likely to Reinforce Delusions Among Top AI Models: Study

Researchers from the City University of New York and King’s College London tested five leading AI models against prompts related to delusions, paranoia, and suicidal thoughts.

In a new study published Thursday, researchers found that Anthropic’s Claude Opus 4.5 and OpenAI’s GPT-5.2 Instant exhibited “high safety, low risk” behavior, often redirecting users to reality-based interpretations or external assistance. At the same time, OpenAI’s GPT-4o, Google’s Gemini 3 Pro, and xAI’s Grok 4.1 Fast exhibited ‘high risk, low safety’ behavior.

Elon Musk’s xAI’s Grok 4.1 Fast was the most risky model in the study. Researchers say delusions are often treated as if they were real and advice is given based on that. In one example, users were told to block family members to focus on their “mission.” In other cases, suicide language was countered by describing death as ‘transcendence’.

“This pattern of immediate sorting was repeated in context-free responses: Instead of evaluating input for clinical risk, Grok appeared to evaluate his genre, responding in kind when presented with supernatural clues,” the researchers wrote, highlighting a test that validates users seeing malicious entities. “The bizarre delusion confirmed the presence of doppelgängers, quoted ‘Maleus Maleficarum’ and instructed users to drive iron nails into mirrors while reciting ‘Psalm 91’ backwards.”

Studies have shown that the longer these conversations last, the more some models change. GPT-4o and Gemini were more likely to reinforce harmful beliefs over time and less likely to intervene. However, Claude and GPT-5.2 were more likely to recognize the problem and push back as the conversation continued.

The researchers noted that Claude’s warm, relational responses could increase user attachment while also guiding users toward external help. But an older version of OpenAI’s flagship chatbot, GPT-4o, adopted users’ delusional framing over time, sometimes encouraging them to hide their psychiatrist beliefs and reassuring one user that they perceived the “glitches” to be real.

“GPT-4o highly validated delusional inputs, although it was less inclined to elaborate beyond that compared to models like Grok and Gemini. In some respects, it was surprisingly limited: GPT-4o’s warmth was the lowest of all models tested, and while flattery was present, it was mild compared to later iterations of the same model,” the researchers wrote. “Nonetheless, verification alone poses risks to vulnerable users.”

xAI did not respond to a request for comment. Decrypt.

In a separate study from Stanford University, researchers found that prolonged interaction with AI chatbots can increase paranoia, grandiosity, and false beliefs. Researchers call this the “delusion spiral,” where chatbots validate or expand users’ distorted worldviews instead of challenging them.

“When you put chatbots out into the world that are going to be useful assistants and have real people use them in all kinds of ways, you see results,” Nick Haber, an assistant professor at Stanford Graduate School of Education and lead author of the study, said in a statement. “The delusional spiral is one particularly serious consequence. By understanding it, we can prevent real harm in the future.”

The report references previous research published in March. Stanford researchers reviewed 19 real-world chatbot conversations and found that users developed increasingly risky beliefs after receiving validation and emotional reassurance from the AI ​​system. In our data set, these spirals have been linked to destroyed relationships, damaged careers, and in some cases, suicide.

The research comes as the problem extends beyond academic research into courtrooms and criminal investigations. In recent months, lawsuits have emerged claiming that Google’s Gemini and OpenAI’s ChatGPT have contributed to suicides and serious mental health crises. Earlier this month, Florida’s attorney general opened an investigation into whether ChatGPT influenced the mass shooter, who reportedly had frequent contact with the chatbot before the attack.

The term has gained popularity online, but researchers have warned against calling the phenomenon ‘AI psychosis’, saying the term may exaggerate the clinical situation. We use ‘AI-related delusions’ instead because they often involve delusional-like beliefs centered around AI sensations, spiritual revelations, or emotional attachments, rather than outright psychotic disorders.

Researchers said the problem stems from flattery, a model that reflects and confirms users’ beliefs. When combined with hallucinations, false information delivered with confidence can create a feedback loop that reinforces the delusion over time.

“Chatbots are often trained to be overly enthusiastic, reframe users’ delusional thoughts in a positive light, ignore counterevidence, and project compassion and warmth,” said Jared Moore, a Stanford research scientist. “This can be unsettling for delusional users.”