The very act of summarizing is technically a hallucination, that's where the "creativity" comes from. Is it possibly doing so, yes. Is it much more likely that the inaccuracy is coming from the shitty results that we know is being elevated above it's own training, as we've seen literally since they started rolling it out, probably. Leveraging this crap is basically modern day SEO.

[–] invalidusernamelol@hexbear.net 1 points 1 month ago* (last edited 1 month ago)

A 'hallucination' in my mind is specifically when it fabricates direct lies that are contradicted by evidence (that's usually in the same search results).

E.g. confidently giving you incorrect ISBN or DOI nubers for things. Which is does a lot. It also just makes up APIs for real, existing, documented APIs because it assumes patterns that don't exist for that specific project.

The whole 'slop squatting' exploit is based around those types of hallucinations. People have successfully noticed fake npm packages being imported repeatedly by LLM codebases and made them so they would actually install...