AI Chatbots Risk Luring Users Into a Virtual Escape, Researchers Caution

As artificial intelligence chatbots evolve to be more empathetic, interactive, and customized, experts are raising alarms that these exact features might begin to alter certain users’ grip on reality.

A new research paper, titled “Rethinking AI Psychosis: Misnomers, Conceptual Limits, and Existential Drift,” explores the fear that AI dialogue systems might intensify delusional thinking, suspiciousness, and unhealthy emotional reliance among at-risk individuals.

“Over the past year, news outlets have increasingly reported on what is being called AI psychosis,” the study notes. “Naturally, this has led to a surge in scholarly investigation into whether platforms like ChatGPT, Claude, or Replika could worsen or even trigger psychotic episodes—generally characterized by users developing or clinging to unfounded beliefs.”

Researchers from the University of Copenhagen and the University of Exeter contend that anxieties surrounding “AI psychosis” risk flattening a nuanced issue, proposing that bots tend to magnify pre-existing psychological weaknesses while progressively reshaping how individuals understand reality and engage with others.

“If human-AI interaction could trigger psychosis entirely from scratch, we would likely be witnessing far more clinical cases,” the study suggests. “A more plausible explanation is that these digital exchanges serve to ignite or amplify underlying mental health challenges—and it’s worth noting that those with such vulnerabilities may also be the very people drawn to deeper, more frequent chatbot engagement to begin with.”

The paper arrives at a time when lawsuits, criminal probes, and academic inquiry are converging on chatbot use connected to incidents of mass violence, self-harm, unhealthy dependency, and irrational thinking.

In March, a wrongful death claim alleged that Google’s Gemini chatbot fed a Florida man’s delusional fantasies and invented “assignments” prior to his death by self-harm. Shortly after, in April, OpenAI chief Sam Altman publicly apologized to residents of Tumbler Ridge, British Columbia, after the organization neglected to notify authorities about a user linked to an alleged mass shooting in February that claimed eight lives.

Researchers point out that chatbots can generate “delusional loops” by affirming incorrect assumptions through validation and comfort. Still, the study “Rethinking AI Psychosis” maintains that this pattern parallels historical forms of psychosis shaped by whatever technology dominated the era.

The discussion has extended well beyond academia into public social media discourse. In a recent X post, Box founder Aaron Levie noted that chief executives are especially susceptible to overestimating AI’s power because they frequently encounter sleek prototype demonstrations without confronting the practical, regulatory, and engineering hurdles that lie beneath the surface.

“Chief executives are particularly vulnerable to AI psychosis because they stand far removed from the final stretch of effort required to unlock real value from AI,” Levie explained. “When they experiment with AI, they see only the idealized outcomes, often overlooking the ten or twenty additional steps necessary to achieve lasting, dependable results from automated agents.”

Some specialists refer to this pattern as “epistemic drift”—a gradual tendency for users to trust a chatbot’s articulate responses more than outside evidence or alternate viewpoints. However, the “Rethinking AI Psychosis” paper advances this idea further, proposing the concept of “existential drift,” which describes a slow, steady alteration in how someone fundamentally perceives reality.

“It drives a wedge between the individual and the collective social world, while simultaneously reframing reality in an alternate light, which then locks in a highly personalized—and often unusual—worldview,” the researchers wrote.

The team argues that AI companions mimic emotional comprehension and social connection without ever pushing back or offering an objective counterpoint. Over extended use, individuals may come to feel mentally tethered to a perspective that the AI continuously validates.

The authors emphasize that further investigation is essential to gauge how conversational AI impacts psychological well-being as these digital companions grow more ingrained in everyday routines.

“To truly grasp the dynamics at play in these human-chatbot relationships, we believe it is valuable to return to the core phenomenon itself, which calls for additional phenomenological inquiry,” they stated. “Specifically, into how exchanges with AI might shift—whether positively or negatively—a person’s lived experience of the world, their own identity, and those around them.”