[michaelmrose]

Incidence of harm is a function of harm/population. It is likely that Facebook is orders of magnitude more harmful than ChatGPT and bathtubs and bikes more dangerous than long LLM conversations.

It doesn't mean something more should not be done but we should retain perspective.

Maybe they should try to detect not long conversations but dangerous ones based on spot checking with a LLM to flag problems up for human review and a family notification program.

EG Bob is a nut. We can find this out by having a LLM not pre prompted by Bob's crazy examine some of the chats by top users by tokens consumed in chat not API and flagging it up to a human who cuts off bob or better shunts him to a version designed to shut down his particular brand of crazy eg pre prompted to tell him it's unhealthy.

This initial flag for review could also come from family or friends and if OpenAI concurs handle as above.

Likewise we could target posters of conspiracy theories for review and containment.