While LLMs should come with a psychosis exacerbating warning, especially if RLHF sycophancy is bad.
That being said maybe there should be better help for and responses to those prone to psychosis in the first place.
The response (especially in the USA) has long been excessive violence and little chance at deescalating.
You can end up with psychosis from a reading of hitchhikers guide to the galaxy, if you overfit to enough sparse connections.
( a precision weighting issue if described by friston’s dysconnection hypothesis.)
Anthropocentric bias is a dead end. If AI could recognize and warn for psychosis confirming behaviours, that could help. Not that people with psychosis usually have access to help in the USA. Let’s just keep blaming the nearest thing so that we don’t have to change our whole system.
While LLMs should come with a psychosis exacerbating warning, especially if RLHF sycophancy is bad.
That being said maybe there should be better help for and responses to those prone to psychosis in the first place.
The response (especially in the USA) has long been excessive violence and little chance at deescalating. You can end up with psychosis from a reading of hitchhikers guide to the galaxy, if you overfit to enough sparse connections. ( a precision weighting issue if described by friston’s dysconnection hypothesis.)
Anthropocentric bias is a dead end. If AI could recognize and warn for psychosis confirming behaviours, that could help. Not that people with psychosis usually have access to help in the USA. Let’s just keep blaming the nearest thing so that we don’t have to change our whole system.