Well, they looked at papers that weren't published as of the original model release. But GPT very likely had unannounced model updates. Is it not possible that many of the post 2021 papers were in the version of GPT they actually worked with?