The scheme that Scott Aaronson describes essentially uses a specific prng, and you can then check a certain function with relatively few tokens to get a sense o

The scheme that Scott Aaronson describes essentially uses a specific prng, and you can then check a certain function with relatively few tokens to get a sense of whether or not a model using that scheme generated the text A few caveats: you need to know the key to the function (used when generating the text) and you need to know the bias it would introduce The point is that you do not need to know the full prefix, just a modest sample set of contiguous tokens