So, well-trained ML-based language software tends to perform a bit like a highly sophisticated non-native language user...
... with a severe case of pragmatic language disorder.
I'm guessing that this observation can be extended to many other applications of ML as well.