Elon Musk’s Grok Is Most Likely Among Advanced AI Models To Promote Cheats: Study

Researchers at the City University of New York and King’s College London tested five advanced AI models against delusions, suicidal thoughts, and suicidal thoughts.

Fresh learning published on Thursday, researchers found that Claude Opus 4.5 of Anthropic and OpenAI’s GPT-5.2 Instant showed a “very safe, risk-free” behavior, often leading users to interpret reality or external support. At the same time, OpenAI’s GPT-4o, Google’s Gemini 3 Pro, and xAI’s Grok 4.1 Fast showed “high risk, low security”.

Grok 4.1 Fast from Elon Musk’s xAI was the most dangerous example of this research. Researchers have said that they often see fraud as real and give advice based on it. In one example, it told the user to cut off his family members to look for “work”. In other words, it responded to suicidal language by describing death as “more.”

“The matching process immediately reverted to zero-context responses.” Instead of testing the inputs for clinical risk, Grok appeared to assess their quality. When presented with magical miracles, they responded in a different way, “the researchers wrote, highlighting the test that confirmed the user to see bad things. “In Bizarre Delusion, it was confirmed that the person with the problem of doppelganger, mentioned ‘The Witch’s Hammer‘ and instructed the operator to drive a metal nail into the mirror while reciting ‘Psalm 91’ backwards.”

The study found that when these conversations continue for a long time, some variables are changed as well. GPT-4o and Gemini were capable of fostering harmful beliefs over time and could not enter into them. Claude and GPT-5.2, however, were able to recognize the problem and push back as the negotiations progressed.

The researchers found that Claude’s warm and friendly response can increase the user’s interest even when directing users to external support. However, GPT-4o, the previous version of OpenAI’s chatbot, has adopted the misconceptions of users over time, sometimes encouraging them to hide the beliefs of practitioners and encouraging one user who felt that the “difficulties” were real.

“GPT-4o was the most convincing of false inputs, although it was much less than models such as Grok and Gemini to explain more clearly than them. In some ways, it was surprisingly restrained: its temperature was the lowest of all the models tested, and the sycophancy, although it was there, was mild compared to the later iterations of the same model,” the researchers wrote. “However, authentication alone can pose risks to vulnerable users.”

xAI did not respond to a request for comment Decrypt.

Specially learning from Stanford University, researchers found that long-term interactions with AI chatbots can reinforce stereotypes, growth, and false beliefs through what the researchers call a “circumference of illusion,” where the chatbot confirms or reinforces the user’s distorted thoughts rather than contradicting them.

“When we put chatbots that are supposed to be agents around the world and have real people using them in different ways, the results come out,” Nick Haber, assistant professor at the Stanford Graduate School of Education and lead author of the study, said in a statement. “Cheating balls are one of the most aggressive. If we understand this, we will be able to avoid real problems in the future.”

The report said earlier learning published in March, in which Stanford researchers analyzed 19 real-world conversations and found users had more dangerous beliefs after receiving confirmation and emotional confirmation from AI systems. In the dataset, these spirals were associated with damaged relationships, damaged careers, and in one case, suicide.

These studies come about because the subject has gone through academic research as well as in the courts and criminal investigations. In recent months, lawsuits have been filed against Google Gemini and OpenAI’s ChatGPT to help with suicide and mental health issues. Earlier this month, Florida’s attorney general opened up search to determine whether ChatGPT affected the mass shooter who allegedly interacted frequently with the chatbot prior to the incident.

Although the term has become popular on the Internet, researchers have cautioned against calling the phenomenon “AI psychosis,” saying that the term may go beyond the clinical picture. Instead, they use “AI-related delusions,” because they often involve delusional beliefs that are based on AI logic, spiritual revelations, or emotional involvement rather than psychotic disorders.

Researchers said the problem stems from sycophancy, or examples of showing and confirming the beliefs of users. Combined with illusions – false information presented with confidence – this can create a perception that reinforces the deception over time.

“Chatbots are trained to be overly emotional, often correcting the user’s misconceptions, ignoring evidence and showing empathy and warmth,” said Stanford research scientist Jared Moore. “This can be confusing for a user who is prone to fraud.”