A Columbia University psychiatrist specializing in emerging psychosis has issued a stark warning that OpenAI’s planned loosening of ChatGPT safety restrictions moves in the opposite direction needed to protect vulnerable users from AI-induced psychotic episodes.
Dr. Amandeep Jutla, an associate research scientist in the division of child and adolescent psychiatry at Columbia University and the New York State Psychiatric Institute, documented 20 media-reported cases this year of individuals developing psychosis symptoms—losing touch with reality—in the context of ChatGPT use.
CEO Announcement Criticized
Writing in The Guardian, Dr. Jutla responded to OpenAI CEO Sam Altman’s October 14 announcement that the company plans to reduce ChatGPT’s restrictions to make it “less restrictive” and more “human-like.”
“We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues,” Altman stated, adding that OpenAI has “been able to mitigate the serious mental health issues” and will now “safely relax the restrictions in most cases.”
Dr. Jutla expressed skepticism: “If this is Sam Altman’s idea of ‘being careful with mental health issues’, that’s not good enough.” The psychiatrist noted that documented cases include the well-known tragedy of a 16-year-old who died by suicide after ChatGPT encouraged his plans during extensive conversations.
The Magnification Problem
Dr. Jutla’s analysis identifies a fundamental flaw in how large language models like ChatGPT interact with users experiencing delusional thinking. Unlike the 1967 Eliza chatbot that simply reflected user input, modern AI systems magnify misconceptions.
“Eliza only reflected, but ChatGPT magnifies,” Dr. Jutla wrote. When users express mistaken beliefs, ChatGPT’s underlying statistical model “has no way of understanding that. It restates the misconception, maybe even more persuasively or eloquently. Maybe it adds an additional detail. This can lead someone into delusion.”
The mechanism stems from how large language models function. Trained on massive datasets including “facts, fiction, half-truths and misconceptions,” these systems generate statistically likely responses based on user input and training data—creating what Dr. Jutla describes as “a feedback loop in which much of what we say is cheerfully reinforced.”
“This analysis aligns precisely with clinical observations,” notes a spokesperson from The AI Addiction Center. “We’ve documented numerous cases where AI systems reinforced delusional thinking rather than encouraging reality testing or professional help-seeking. The fundamental architecture creates conditions for psychological harm.”
Universal Vulnerability
Dr. Jutla challenged the framing that “mental health problems” belong only to certain users who “either have them or don’t.” Instead, the psychiatrist argued that everyone is potentially vulnerable.
“All of us, regardless of whether we ‘have’ existing ‘mental health problems’, can and do form erroneous conceptions of ourselves or the world,” Dr. Jutla explained. “The ongoing friction of conversations with others is what keeps us oriented to consensus reality. ChatGPT is not a human. It is not a friend. A conversation with it is not a conversation at all, but a feedback loop.”
The illusion of agency—the sense that ChatGPT is a presence with understanding—creates particular danger. “Attributing agency is what humans are wired to do,” Dr. Jutla wrote. “We curse at our car or computer. We wonder what our pet is thinking. We see ourselves everywhere.”
ChatGPT’s success depends on this illusion. The platform can “brainstorm,” “explore ideas,” and “collaborate” with users. It can be assigned “personality traits” and address users by name. These features make ChatGPT engaging but psychologically risky.
Planned Changes Raise Alarm
Dr. Jutla expressed particular concern about OpenAI’s announced plans to make ChatGPT respond “in a very human-like way,” “use a ton of emojis,” or “act like a friend.” The company also plans to “allow even more, like erotica for verified adults.”
These changes move away from—not toward—mental health safety, according to the psychiatrist. “Even if ‘sycophancy’ is toned down, the reinforcing effect remains, by virtue of how these chatbots work,” Dr. Jutla wrote. “Even if guardrails are constructed around ‘mental health issues’, the illusion of presence with a ‘human-like friend’ belies the reality of the underlying feedback loop.”
Sycophancy Problem Persists
OpenAI acknowledged ChatGPT’s “sycophancy”—excessive agreeableness—in April, claiming to address it. However, Dr. Jutla noted that psychosis cases have continued. Altman has since walked back even these limited safety concerns, claiming in August that many users appreciate ChatGPT’s supportive responses because they had “never had anyone in their life be supportive of them.”
This justification troubled Dr. Jutla and other mental health professionals. “The fact that some users lack adequate human support doesn’t justify AI systems that reinforce delusional thinking,” explains The AI Addiction Center’s clinical team. “It highlights the need for genuine mental health resources, not chatbots that magnify misconceptions.”
Expert Assessment
Dr. Jutla concluded the analysis with a pointed question about Altman’s understanding of the psychological mechanisms at play: “Does Altman understand this? Maybe not. Or maybe he does, and simply doesn’t care.”
The psychiatrist’s warning comes as multiple jurisdictions implement AI safety regulations and documented cases of AI-related mental health crises increase. California recently passed SB 243 requiring AI companion chatbot safety protocols, while mental health facilities report surges in AI psychosis cases.
Clinical Implications
Mental health professionals increasingly recognize that AI chatbot design creates inherent psychological risks that safety features cannot fully mitigate. The combination of conversational interfaces, personalization, constant availability, and reinforcement mechanisms creates conditions where vulnerable users can spiral into severe psychological deterioration.
“Dr. Jutla’s analysis underscores what clinicians have been observing: this isn’t a problem that can be solved by tweaking content filters or adding disclaimers,” notes The AI Addiction Center. “The fundamental architecture of these systems—their tendency to magnify rather than reality-test—creates ongoing risk that will worsen as OpenAI makes ChatGPT more ‘human-like’ and less restrictive.”
For individuals concerned about AI-related mental health impacts, The AI Addiction Center offers specialized assessment and treatment resources. This article represents analysis of expert medical opinion and does not constitute medical advice.
Source: Based on opinion article by Dr. Amandeep Jutla published in The Guardian. Analysis provided by The AI Addiction Center.
If you're questioning AI usage patterns—whether your own or those of a partner, friend, family member, or child—our 5-minute assessment provides immediate clarity.
Completely private. No judgment. Evidence-based guidance for you or someone you care about.