I've seen this. I can get quite a bit of decent responses from LLMs about Python. But using it for Elixir produces more shady/dated answers. And Jetpack Compose, it just completely gets weird (and mostly worthless, being either too synthetically approximated or dated).

It really illustrates (to me) that LLMs are "Plausibility Machines". They produce plausible answers. Plausibility is informed by repeated observability over large sets. We asses something as plausible when it matches the majority of our observations. That's how the parlor trick of seeming to "tell the truth" is pulled off. The venn overlap of "plausible" and "true" is pretty high. But when "truth" is in flux, you start to see the feet of the wizard behind the curtain.