[User23]

With some brief experimentation ChatGPT also fails this test.

[lostmsu]

It might make sense: any kind of fine-tuning of LLMs usually reduces generalization capabilities, and instruction-tuning is a kind of fine-tuning.