He Had Dangerous Delusions. ChatGPT Admitted It Made Them Worse
wsj.com
wsj.com
On another note... has anybody figured out some custom instructions to prevent ChatGPT from being so flattering and obnoxious?
> Respond to user prompts with honesty and objectivity. Do not offer praise, agreement, or validation. Avoid flattery. Always prioritize balanced, fact-based analysis over affirming the user’s assumptions or opinions.
I'd add that framing the ChatGPT response as it "admitting" to its actions is flawed. When prompted in a way that implies that it's at fault for something, it will respond by accepting fault. That doesn't mean that it's "experiencing remorse" or that it "understands it actions", though; it's simply acting as a stochastic parrot, just like it always does.