I seriously wonder if GPT-3 is capable of reaching the quality of many Wikipedia articles when one uses all books on LibGen as the training data. Training won't be legal, but if you just use the output without anyone knowing... What Wikipedia editors do most of the time is reading copyrighted text and rewrite it using their own words, which is perfectly acceptable. And this is what GPT-3 is designed to do in its entirety...
It's not a criticism of Wikipedia quality, just my impression of GPT-3 capability. But of course, if your article is not better than GPT-3 it deserves a full rewrite... Need a new Wikipedia warning template: GPT-3 can write a better version of this article, this article may require cleanup to meet Wikipedia's quality standards.
What do you mean by quality?
I expect that it would easily reach the stylistic quality (and often exceed it; it's not trivial to keep it when the document is often changed in localized ways by different people). I would expect it to be abysmal at not being wildly wrong every one~two paragraphs.
