Eh dunno. I've been gaslit (gaslighted?) by AI quite a few time. Along these lines: here's a design problem, how do I fix it? Oh known problem, here's the only sane way of doing it. Then I poke holes, AI tells me nonono, do like Computer say. Eventually relenting, telling me I'm right to push back, and doing a 180 turn. Then agreeing with me/adding options etc.

The RL metaoptimisation clearly sometimes pushes it to "here's one solution, end of story".