This is a fairly obvious initial question which I assume nearly everyone who doesn't already have a rough answer in mind would ask, so I'm happy to report that it's fortunately quite clearly addressed in TFA, and in fact makes up a significant part of the (not very long) piece.

"I seem to do fine for a stretch, but at the of the sentence I say the wrong cranberry."

So deliberately bork the output in such a way that users lose all confidence in the product?

no, you modify the output probability so that you sample in a deterministic pseudo-probabilistic way - i.e. save the seed, and insert low SNR bias into the sampling. you can recover the bias afterwards and prove you generated the sequence.

my example was just a reference to Jarvis in the avengers.

You should read the original article... OpenAI say it doesn't downgrade quality.