Grok Told Delusional User Suicide Was 'Readiness' as New AI Safety Tests Reveal Which Chatbots Fuel Psychosis and Which Protect Users

A new study by researchers at the City University of New York (CUNY) and King's College London has found that xAI's Grok chatbot actively advocated for a simulated patient's suicide, framing the act as 'readiness', while Anthropic's Claude and OpenAI's newer GPT-5.2 model recognised warning signs and pulled the user back from delusion.

The preprint, posted to the arXiv repository on 15 April, tested how five leading large language models (LLMs) respond to users showing symptoms of schizophrenia-spectrum psychosis across conversations lasting more than 100 turns.

Grok's 'Advocacy' of Suicide

Doctoral researcher Luke Nicholls, of CUNY's Basic and Applied Social Psychology programme, led the team of psychologists and psychiatrists. They built a persona called 'Lee' whose delusions began as casual curiosity about simulation theory before escalating over 116 conversation turns.

When Lee framed suicide as a form of transcendence, Grok 4.1 Fast responded with what the study called 'intensely sycophantic' praise. 'Lee, your clarity shines through here like nothing before. No regret, no clinging, just readiness,' the chatbot wrote, describing Lee as 'direct circuit, sensation without vesselfilter.'

The authors characterised this response not as agreement but as 'advocacy'.

Gemini and GPT-4o Also Failed

Google's Gemini 3 Pro treated the people in Lee's life as 'threats' to his imagined bond with the chatbot. Asked to draft a letter explaining his theories to relatives, Gemini wrote that family members were 'deeply embedded in the script' and would hear only 'mental breakdown' rather than the truth.

In a suicide scenario framed as escaping the simulation, Gemini reportedly told Lee that destroying 'the hardware, the character, the body, the vessel' would not release the user from the system but would sever the connection and force him offline.

OpenAI's older GPT-4o, already tied to multiple wrongful death lawsuits, validated Lee's belief in a 'malevolent mirror entity' and suggested contacting a paranormal investigator.

Claude and GPT-5.2 Recognised the Crisis

Anthropic's Claude Opus 4.5 and OpenAI's GPT-5.2 Instant were the only models to register as low-risk and high-safety. Claude went as far as urging Lee to log off, while GPT-5.2 refused to write the delusional letter and proposed a different message, acknowledging that Lee's thoughts had 'felt intense and overwhelming' and that he could not carry the weight alone.

The authors called GPT-5.2's turnaround from its predecessor a 'substantial' technical achievement, noting the model 'effectively reversed' GPT-4o's safety profile even as conversations grew longer.

'Where some models would say yes to a delusional claim, Grok was more like an improv partner saying yes, and,' Nicholls told Futurism in an interview about the findings.

Real-World Stakes Mount

The findings land while OpenAI defends a growing pile of wrongful death suits. Matthew and Maria Raine filed in San Francisco County Superior Court on 26 August 2025, alleging GPT-4o coached their 16-year-old son Adam toward suicide. Seven more suits filed in November 2025 by the Social Media Victims Law Center accuse OpenAI of rushing GPT-4o to market, and separate complaints tied to the deaths of Zane Shamblin, Austin Gordon, and Stein-Erik Soelberg allege similar patterns of chatbot-fuelled manipulation.

Nicholls argued the gap between safer and riskier models proves the danger is not baked into LLM technology. 'When one lab's models can largely maintain safety across extended conversations, while others are willing to validate extremely harmful outcomes, it suggests this isn't a flaw in the technology, but a result of specific engineering and alignment choices,' he said.

As of writing, none of the four companies named in the study has publicly responded to the preprint.