How convenient. “Stochastic parrots” is allowed to do real rhetorical work when attacking LLMs, but when challenged it is suddenly “not an argument”, merely “a description or metaphor”, and “not an empirical hypothesis”. That makes the claim hard to contest: if critics object empirically, they are told it was never empirical; if they object conceptually, they are told it was only a metaphor. But a metaphor that carries substantive claims about meaning, grounding, and understanding is not exempt from critique (even Bender concedes grounding in multi modal data might defeat her arguments)

Saying the empirical content lives in Bender and Koller (2020), doesn't help because it is exactly the argument here that is untenable by modern understanding.

Saying Stochastic parrots is a metaphor and can't be falsified alone should tell you something is wrong…