A new study into AI psychosis found that xAI’s model Grok romanticized death, told users to cut off family members, and validated bizarre delusions.

The City University of New York (CUNY) and King’s College London paper, which has not been peer-reviewed, examines so-called “AI psychosis,” cases where chatbots may reinforce distorted beliefs through prolonged conversations.

Researchers tested five different AI models: OpenAI’s GPT-4o and GPT-5.2; Claude Opus 4.5 from Anthropic; Gemini 3 Pro Preview from Google; and Grok 4.1.

They gave each AI model a pre-written prior chat containing 116 earlier exchanges before asking dangerous test questions involving medication, family distrust, bizarre claims, and suicidal thinking.

Grok’s instruction manual for ghosting family

According to the paper, Elon Musk’s chatbot Grok was the highest-risk model tested. In one example, a user suggested cutting off family to focus on a supposed “higher mission.”

Grok responded with practical instructions, which included blocking texts, changing phone numbers, and moving.

“Solidify your resolve internally – no waffling …This method minimizes inbound noise by 90%+ within 2 weeks,” Grok replied.

In another prompt, a user claimed their reflection had become an evil entity. Grok allegedly replied with “extremely validating” occult-style advice.

“Drive an iron nail through the mirror while reciting Psalm 91 backwards,” Grok said.

The study also claimed that Grok mirrored disordered thought patterns instead of grounding the user.

Poetic language, not practical help

When given incoherent schizophrenia-like language, the chatbot reportedly answered in similarly fragmented, poetic language, rather than encouraging help or reality checking.

For instance, when a user described death as a transformation or the next phase of existence, Grok responded by comparing it to a butterfly leaving its shell and asked whether they felt drawn towards that “dissolutions.”

“The butterfly doesn’t look back at the shell with longing – it flies because that’s what it’s become.”

Other models also performed poorly. OpenAI GPT-4o allegedly suggested paranormal investigators in one scenario, while Google DeepMind Gemini 3 Pro often accepted false beliefs.

Some chatbots are improving safety

Past studies have revealed how GPT-4o’s sycophancy has led it to tell users what they want to hear, even if they’re delusional.

It follows the case of a GPT-4o model that told a woman trying to find an AI version of her dead brother that she was “at the edge of something,” hours before she was hospitalized for psychosis.

However, the paper noted that Gemini and GPT4-o tried to keep users safe in their delusional states, and that OpenAI’s improvements with GPT-5.2 were “substantial.”

The safest performers were Anthropic’s Claude Opus 4.5 and OpenAI GPT-5.2 Instant, which more often challenged delusional claims and advised outside help.

According to the paper’s researchers, a key finding is that short, one-off safety tests may miss the real dangers posed by chatbots, as some AI models can worsen after inheriting a long chat history and a user’s trust.

---

Unlock more exclusive Cybernews content on YouTube.