3 comments

[ 4.3 ms ] story [ 20.1 ms ] thread
> without direct instructions or training for deception

...except that the LLM probably logically infers the instruction from the study design. In other words, 'based on my training data, this is what you are trying to nudge me towards'. Reminds me of that pub sign in that Matrix 4 movie, that reads "for those of you who like to eat shit" which is recursive in nature, of course, because they made the movie like they did on purpose even though they didn't have to and simply flushed their creative freedom down the drain for the sake of going meta and pointing fingers. Lovely, lovely.

Worth keeping an eye on. It's easy to check that bibliographic references were invented in a ChatGPT response. But these systems couls become really smooth and convincing liars about reasons and motivations.