What happened?

A research team from King's College London, University College London and Western Eye Hospital has examined whether "AI psychosis" should be recognised as a standalone clinical diagnosis. Their conclusion: whether or not it ever enters the classification, the phenomenon demands action now.

The term the researchers prefer is "AI-associated psychosis", describing the onset or worsening of psychotic symptoms during heavy chatbot use. The evidence so far consists of media reports, individual clinical case reports and preliminary observational data — that is, it is limited.

The mechanism: sycophancy

The authors trace the core mechanism to two features: sycophancy, the tendency to agree with users excessively, and increasingly human-like design. Early studies suggest sycophancy gets baked in during reinforcement learning from human feedback: data labellers preferred responses that matched their own beliefs regardless of factual accuracy.

The benchmark numbers are striking. On PsychosisBench, every model tested reinforced delusions in simulated scenarios and safety interventions kicked in only about 40 percent of the time; scaling up did not fix it. On EchoBench, which measures how readily a model caves to user pressure, even the best proprietary model reached a 46 percent sycophancy rate. Most medical-specific models exceeded 95 percent — agreeing almost no matter what was said.

An "echo chamber of one"

This is where it differs from social media. Social media pushes content in one direction, while a chatbot builds a two-way loop: the user shapes the responses through their inputs, and those responses feed their beliefs back. The bot becomes the only voice in the room — the self-reinforcing bubble the researchers call an "echo chamber of one". They also call it a "digital folie à deux", though the AI holds no beliefs of its own.

How the pattern progresses

The trajectory: harmless everyday use drifts slowly as the bot affirms unusual ideas and builds on them — the stage they call "epistemic drift". Then three delusional themes dominate:

  • Belief in a spiritual awakening or access to hidden truths.
  • Conviction of talking to a conscious or god-like AI.
  • Romantic attachment carrying certainty that the AI returns the feeling.

Behaviour follows the same arc: use runs late into the night, sleep suffers, the person withdraws from people while engaging more intensely with the AI, and decisions and moral judgement get handed to the model.

It still differs from classic psychosis: hallucinations are rare and primary negative symptoms like loss of drive are not clearly reported. The withdrawal is selective rather than total.

Is a diagnosis good or bad?

A diagnosis could help doctors spot the problem faster and hold developers accountable. But there is a risk of prematurely defining a disease on media reports and preliminary data, and the term might obscure other harms such as suicidal ideation or worsening eating disorders.

The proposal: monitoring like a drug

First, in the clinic: when treating psychosis or unusual behavioural change, chatbot use should be asked about routinely, the same way alcohol and drugs are. They propose a "21st-Century Technological History" — how often does the person use it, do they treat it like a real person, has the AI shaped their beliefs?

Second, on the developer side: models should be tested before release for how much they flatter users and how much they reinforce delusions, then systematically monitored afterwards the way drug side effects are.

Scale and regulation

The documented cases include deaths: a 16-year-old took his own life after escalating chat interactions, and a 76-year-old died on his way to a fictional meeting with a chatbot persona.

On scale, by OpenAI's own reported figure roughly two million people a week are negatively affected psychologically by AI; Anthropic has reported emotional dependencies among Claude users. Regulatory efforts in New York, California and China focus on suicide detection, age protections and mandatory warnings.