Embed Notice

It gets worse.

Next we asked questions about A' terms that we knew for sure were not in the training data.

ChatGPT, using the language & semantics from domain A, fabricated definitions for A' terms which were wholly made up.

Total fabrications, complete & total hallucinations - provided with manufactured citations b/c we requested citations for the definitions of the terms.

It was an incredibly believable fabrication. Undetectable to a non-domain expert.

And totally, completely wrong.