I mean, in training children we give them reasoning tasks they commonly get wrong. I don't think we say they are incapable of reasoning because they get wrong answers commonly?

This is why we see improvement in GPT when chain of thought/tree of thought is used with reasoning for each step. That can't correct every failure mode, but it increases the likelihood you'll receive a more correct answer.