Can the experiment be summerised by saying that training the model is a kind of probabilistic pre-calculation that converts the Stockfish expert system into a different, rather distinct representation that is worse than Stockfish, but still quite good?

It is a very visible indicator of the quality of the whole. If the spelling is frequently not correct, which a reader can detect relatively easily, how many more mistakes are hidden in the content, which a reader can not detect easily? Are these completely independent variables? I do not think so. Therefore, I also assess the reliability of an article based on the frequency of careless mistakes.

What is an even larger warning sign, are cliches used to spice up an article. Ars Technica is hardly to blame here, but the Smithsonian magazine is full of it.