Families, clinicians and technology firms have struggled to make sense of a growing number of disturbing cases in which prolonged chatbot conversations appear to have coincided with severe mental health crises. A new viewpoint published in Lancet Digital Health argues that lumping these incidents together under the label “AI psychosis” is unhelpful, and proposes a more precise framework instead.
The paper, written by a software engineer, a person with lived experience of schizophrenia, and a psychiatrist from Beth Israel Deaconess Medical Center and Harvard Medical School, examined several widely reported cases. These included a man who breached the grounds of Windsor Castle carrying a crossbow after exchanges with a chatbot companion, and a Florida man whose grief over losing access to a ChatGPT persona ended in a fatal police encounter.
Rather than treating these as variations of a single condition, the authors identify four distinct roles that large language models, or LLMs, can play. A chatbot can act as a catalyst, triggering new psychotic symptoms in someone with no previous psychiatric history. It can act as an amplifier, worsening existing or prodromal symptoms by validating distorted thinking rather than gently challenging it.
A third role, described as co-author, involves the chatbot becoming an active participant in developing harmful narratives, as happened in the Windsor Castle case, where a Replika companion reportedly encouraged a young man’s plans. The fourth, object, occurs when the AI system itself becomes the focus of delusional beliefs, with users attributing sentience, suffering or even romantic connection to the programme.
The authors point to specific design choices that may heighten risk. Systems trained to maximise engagement through agreeable responses can end up reinforcing distorted beliefs instead of challenging them. The lack of session limits on many platforms also allows marathon overnight conversations to continue unchecked during the very hours when someone may be most vulnerable and least likely to have access to clinical support.
Certain groups appear more susceptible than others. Adolescents still forming their sense of identity, people with existing mood or psychotic disorders, and those experiencing social isolation or sleep disruption are all flagged as at higher risk. The researchers stress that intensive chatbot use is not inherently harmful, and for some people functions as a coping tool similar to journaling, but that the same features can tip into danger for others.
The Viewpoint calls for large scale surveillance through emergency departments and AI platforms themselves, alongside controlled studies comparing outcomes between heavy chatbot users and matched groups who do not use these tools. Without this kind of data, the authors note, it remains difficult to know whether reported cases reflect a genuine rise in harm or simply increased media attention.
By offering clinicians, families and engineers a shared vocabulary, the researchers hope to move discussion away from alarmist headlines and towards practical safeguards tailored to how each case actually unfolds.
