Comment by Slow_Hand

I prefer “confabulation”. It seems truer to what is happening:

The LLM isn’t seeing something that’s not there, but deliberately making up _something_ so that it can return a response.