There is a Dunning-Kruger effect in LLMs, with stupid models not realizing that they are stupid. We are training models well: they are basically us.

https://arxiv.org/pdf/2603.09985