Has anyone found a good prompt to fix this? It seems like a subtle problem because it’s 90% too agreeable but will sometimes get really stubborn.

There are plenty of sycophantic humans around, especially with regard to relationship advice.

I find there is an inverse relationship between how willing people are to give relationship advice, and how good their advice is (whether looking at sycophancy or other factors).

AI being the ultimate yes-man is probably why CEOs like it so much.

I always add the following at the end of every prompt. "Be realistic and do not be sycophantic". Which will always takes the conversation to brutal dark corners and panic inducing negative side.

"AI overly affirms users, and that's bad" - everyone nods. "Modern society overly affirms people, and that's bad" - ....

For what it's worth, that wasn't my experience at all the last time I consulted ChatGPT for relationship advice. It was supportive, but in an honest tough love way.

Not AI chatbots but Claude models. Pandering and rushed thinking is the bane of anthropic models. And since they are the most popular ones they poison the whole ecosystem.

ask ai for advice, ask it to steelman an argument, ask to replay what your situation from the other perspective (if it's involving people), push it hard to agree with you and pander to you, then push it to disagree with you, etc.

once you have all the "bounds" just make your own decision. i find this helps a lot, basically like a rubber duck heh.

I do find them cloying at times. I was using Gemini to iterate over a script and every time I asked it to make a change it started a bunch of responses with "that's a smart final step for this task! ...".

To combat sycophancy it's always good to ask the devil's advocate view of whatever the conversation was about in the end.