The complaint underneath this memo is about who absorbs the work. Synthesia's leader put the mechanism in numbers, and the total is what makes it a management question rather than a matter of taste: 100 minutes of added reading set against the 10 minutes of sharpening that never happened, a tenfold transfer from one author to an audience [15]. Polish is what makes the transfer hard to see, because a reader cannot tell from the surface of a document whether its length is carrying an argument or filling space, which is the state Stevenson describes when he says the writing arrives more finished than the thinking [2][18].
The two rules in the memo carry unequal weight. Labeling an internal document as AI-drafted is a self-report, and nothing in the policy checks it [8]. The clause holding that writing you send will be treated as your own, with its comparison to a lawyer accountable for filing hallucinated case law, is enforceable through the review that already exists [10]. It is also the one rule that does not depend on how good the models get.
Polish was doing a job, and the memo removes it. A long formatted document was a costly signal of effort, and it made junior work legible to senior readers who did not have context. A five-bullet norm strips the signal out and forces judgment onto the substance, which is what Stevenson says he wants when he asks for the scrappy version over five pages of fluff [3][7]. The people who pay are the ones whose ideas were previously carried by presentation, and the managers who now have to read generously enough to find an argument that has not been dressed up.
A 200-person company's internal writing memo could be read as one executive's preference wearing the costume of policy [4]. But other firms reached for the same lever independently: Synthesia's warning about "AI sloppification" and Clay's four principles, among them "writing is thinking" and "longer is not better" [12][13]. What none of them supply is measurement. There is no figure for how often AI-drafted proposals reach a desk, and none on whether the labels get applied once the novelty of the memo wears off.
Stevenson also dates his own claim. The memo argues that AI today works best in highly constrained work such as coding or contract review, and that models are not good at unconstrained longform writing "yet" [5]. That is a specific limit written down by the chief executive of a company whose product drafts and analyzes contracts [1], which is to say the constrained side of the line he is drawing. Customers evaluating that product can read the same sentence, and it tells them exactly where the vendor thinks its own category stops working.