ChatGPT Caused Psychosis via Poisoning
Case report here, with excerpts and commentary below:
A 60-year-old man with no past psychiatric or medical history presented to the emergency department expressing concern that his neighbor was poisoning him.
...
In the first 24 hours of admission, he expressed increasing paranoia and auditory and visual hallucinations, which, after attempting to escape, resulted in an involuntary psychiatric hold for grave disability. He received risperidone, which was titrated up to 3 mg daily for psychosis.
...
For 3 months, he had replaced sodium chloride with sodium bromide obtained from the internet after consultation with ChatGPT, in which he had read that chloride can be swapped with bromide, though likely for other purposes, such as cleaning.
The Wikipedia page for bromism (i.e. bromide poisoning) lists psychosis as a possible symptom, so ChatGPT would almost certainly have known it could cause this (someone with access to the old models could easily check). “Likely for other purposes” sounds to me like mere speculation.
Based on the timeline of this case, it appears that the patient either consulted ChatGPT 3.5 or 4.0 when considering how he might remove chloride from this diet. Unfortunately, we do not have access to his ChatGPT conversation log and we will never be able to know with certainty what exactly the output he received was, since individual responses are unique and build from previous inputs.
If it was ChatGPT 3.5 or ChatGPT 4, it’s probably just a weird coincidence, though still unnerving.
There has not been a model officially called “ChatGPT 4.0” (to the best of my knowledge), but there is of course ChatGPT 4o. The timeline comment is fairly strong evidence against this possibility, but the fact that ChatGPT 4 was not on the free tier (while 4o was) is evidence for this possibility.
ChatGPT 4o is by far the model most implicated in LLM-induced psychosis. If that was the model here, this psychosis case could still be a crazy coincidence, but it is extremely concerning that it plausibly isn’t!!
My gut says that my posterior for ChatGPT 3.5 is lower than my prior and likelihoods shown here require, but ChatGPT 4 and ChatGPT 4o being 1:1 feels right. Interactive version here.
Generalizing a bit, I wonder how hard a misaligned ASI would have to work to get every human to voluntarily poison themselves.
Not hard at all, hypnosis already works on lots of people. But there are almost certainly more effective ways to induce mass psychosis for a superintelligence.