I think that using the output of an algorithm to replicate it is not a novel insight. I think that it has been applied to classifiers and regressors since forever.

I think it is valuable that EY draws attention to this effect right now though since it may affect how the LLM API landscape develops.