Drafting has always done two jobs at once, and only one of them is producing text. The other is diagnostic. Sentences fail in specific ways - a claim you cannot finish, a transition that will not hold, a paragraph that keeps restating its own first line - and each failure is information about where your understanding runs out. A model that writes fluently on the first pass removes every one of those signals. The page comes back smooth, and smoothness is indistinguishable, from the outside, from having thought it through.

Leonid Rozenblit and Frank Keil named the relevant failure the illusion of explanatory depth: people rate their understanding of ordinary mechanisms highly, then find the rating was inflated the moment they are asked to write out how the thing actually works. The rating does not fall because they learned something new. It falls because explaining is the first point at which the gap becomes visible to them. A generated draft arrives on the far side of that test, with the gap intact and unmarked. Whoever sends it on - usually its nominal author - then carries a confidence calibrated to prose quality rather than to anything they know.

The cost lands later, and somewhere else. It arrives when a question is asked that the draft never had to answer: in a review, from a client, or from the engineer who has to build the thing. What looks at that moment like a failure of the model is usually a rehearsal that never happened.