ChatGPT self-distances when admitting fault

3 pointschrisjj1 comment

E.g. when its response contained a chunk of its UI HTML code as literal text and I queried why, it said:

Because ChatGPT did not render the <iframe> tag as HTML. It treated my response as ordinary text, so you saw the literal markup.

Note the third and first person.

humans do this sort of thing too, the patterns are in the training data

I've seen many of the open weight models do this, or not, it's kind of random

A coworker commented that he thought they got dumber when they started simulating shame after getting something wrong