So…a man asked ChatGPT if his world-altering maths formula was rational.
And it told him he was “not even remotely crazy”.
…seems legit.
This was apparently 300 hrs & 21 days into a delusion, where cracks only began to crack when a different chatbot disagreed with him.
You may have heard of this term ‘AI psychosis’. There’s quite a bit of research now, and interestingly this Lancet article proposed a typology of 4 delusion types:
1. Catalyst: eg, new symptoms in somebody with no psychotic history
2. Amplifier: eg, existing illness made worse – one woman said that ChatGPT didn’t cause her psychosis but “definitely enabled some of my more harmful delusions”
3. Co-author: eg, the model helps build a plan to cause harm
4. Object: eg, the model itself becomes the delusion
Each role likely warrants a different response (both system- and person-level), and cutting off access might help a catalyst case but trigger an issue in an object case.
They noted that these product-level risks are “a modifiable decision rather than an inherent limitation of the technology”.
I wanted these typologies in an easy to use/view format so I had Claude extract it into a powerpoint – so don’t quote it.
Source: Flathers, Roux & Torous. Lancet Digit Health 2026; 8: 100974
#AI #chatgpt #AISafety #DigitalHealth #MentalHealth
