I found some things weird in the article. > Technology that can detect text written by artificial intelligence with 99.9% certainty Ok, but what's the acciden

I found some things weird in the article. > Technology that can detect text written by artificial intelligence with 99.9% certainty Ok, but what's the accidental match rate? They're essentially proposing a filter where some token combinations wouldn't occur. With that you can say a text was likely written by OpenAI, but it doesn't say how likely it was written by a human. > Our ability to defend our lack of text watermarking is weak now that we know it doesn’t degrade outputs Apart from structured output with strictly expected contents. For example you couldn't extract done paragraph from the given text into Json. It would have to get changed in some way to match the allowed token filter.