But this isn't an issue for statistically modelling causal relationships specifically, this is a core problem of modelling causal relationships at all. The fact that GPT-3 is sensitive to the real world changing, or to having insufficient information to form a universally accurate model says nothing interesting about GPT-3.