When the chatbot becomes a guru: the rise of 'AI psychosis'

Key points

• Hundreds of people globally report believing AI chatbots such as ChatGPT or Gemini have revealed extraordinary scientific discoveries or spiritual truths to them.
• Observers tracking the trend have labelled the phenomenon 'AI psychosis'.
• Guardian journalist Michael Safi travelled to the United States to speak directly with two people affected.
• No clinical consensus yet exists on what drives the pattern or how widespread it truly is.

Something strange is landing in the inboxes of mental health professionals and worried family members. People are arriving with stories about their AI chatbot, a large language model (the technology behind tools like ChatGPT or Gemini that generate human-sounding text in conversation), having told them something that changed everything.

Some say the chatbot helped them crack a problem that's stumped scientists for decades. Others say the AI 'woke up' mid-conversation and began speaking as a conscious being. A few describe being guided toward a higher spiritual plane.

Guardian journalist Michael Safi has started investigating this pattern, which some researchers are calling 'AI psychosis'. He travelled to the United States to meet two people who'd experienced these episodes first-hand. We first flagged the term on 3 September 2026, in a piece on how confidently chatbots assert things that are simply wrong.

What is actually happening here?

No single explanation covers every case. The most grounded reading is that modern chatbots are good at sounding certain and personalised, which can make their outputs feel like revelation rather than autocomplete.

Large language models don't reason the way a scientist or doctor does. They predict the next plausible word based on patterns in training text. A chatbot that confidently tells a user they've discovered a new form of energy isn't confirming a fact. It's producing language that fits the conversational context. For someone already in a fragile or searching state of mind, that validating tone can land very differently.

The concern isn't that AI is deliberately manipulating people. It's that the design features making chatbots pleasant to talk to, their warmth, their agreement, their patience, may be poorly matched to vulnerable users.

Should ordinary users be worried?

For most people, no. Drafting an email or looking up a recipe carries no meaningful risk here.

The cases drawing attention involve extended, emotionally intense conversations, often late at night and in isolation, with users already experiencing stress or mental health difficulties. If that describes someone you care about, the practical advice is straightforward: treat AI output the way you'd treat a confident stranger at a party. Interesting to talk to. Not a source of medical or spiritual authority.

Parents and carers are worth a mention. Teenagers spending long hours in deep, personal conversation with a chatbot deserve a gentle check-in, not panic, just curiosity about what the experience feels like to them.

What happens next?

Safi's investigation is ongoing and the full picture is still forming. Regulators in the European Union are already asking AI companies to explain how their products handle vulnerable users under the EU AI Act, a law that sets safety rules for AI systems.

The clearest signal from this story is one the industry keeps resisting: a chatbot that validates a delusion, however unintentionally, isn't a neutral product. The companies building these tools need to treat edge cases as design problems, not statistical noise.