The AI writing policy is a coordination cost showing up on the wrong line item

Clay, the San Francisco sales intelligence startup, published an AI Writing Policy last week ^[1]. The Dutch startup Polarstep published one over the summer ^[4]. Neither is a ban, and neither is really about writing: both are attempts to bill generative AI for a cost it has been quietly transferring onto readers.

The clearest statement of the mechanism came from Synthesia cofounder Victor Riparbelli, who sent a Slack message to employees in June after noticing AI slop appearing in company documents ^[7]. "When we write too long and verbose documents we place the burden of distilling the content onto the readers," he wrote, "so instead of one person spending ten mins on sharpening a doc, we have ten people spending ten additional minutes reading slop, leading to net-negative productivity" ^[8]. Take his own numbers at face value and the ratio is 10 to 1: ten minutes of author effort avoided, one hundred minutes of reader effort created ^[9]. He has not prohibited LLM use, only asked that it be more intentional ^[7].

That is the shape of the problem. The tooling line item goes down, or at least looks flat, while the coordination line item absorbs the difference and never gets measured. Polarstep CEO Clare Jones described the symptom precisely: documents getting longer and less comprehensible, messages between team members jumbled ^[5]. She told Sifted the writing was full of "genuinely" and "directionally positive," that "quietly" appeared in everything, and that she was having to squint at documents to work out what was being said ^[6]. Squinting is not free. It is just charged to a different person than the one who saved the ten minutes.

Clay's response is an ownership rule rather than a style guide. Its policy holds employees responsible for making sure a document reflects their own thoughts before they share it ^[2]. The operative line, from a policy authored by engineer Sophie Alpert, is that if a reviewer asks "What did you mean by this line?", it is not acceptable to reply "Oh sorry, AI wrote that, ignore it" ^[3]. That is a governance sentence dressed as an etiquette note. It closes the accountability gap that opens the moment a document has no author who can defend it, which matters more when the underlying models can still muddle sentences, exaggerate facts and fabricate, even with careful prompting ^[10].

The softer argument is harder to price but worth logging. A recent Harvard Business Review piece argued that CEOs who use AI to communicate risk flattening their leadership style and losing employee trust ^[11]. Jones extends that to teams: one of Polarstep's values is dedication to craft, and she says how you communicate is part of that craft, so outsourcing it to a model that does not know you, your team or your company forfeits a lot ^[12]. Her example is a Dutch colleague's use of "cucumber time" for the slow summer months, an idiom she had not encountered and found connective, and the sort of thing that disappears when linguistic quirks get sanded off ^[13].

Notably, Jones is not anti-tool. She recorded herself delivering an annual review and used an LLM to turn it into a podcast for a colleague who does not like reading ^[14], and voice AI can take notes while a team talks through a product or sales idea aloud ^[15]. The distinction being drawn is between format conversion and authorship substitution.

What to watch: whether any of these policies acquire teeth beyond exhortation, and whether anyone starts measuring the reader-side cost that Riparbelli asserted rather than assuming it ^[8]. Sifted's read is that this regulation persists as people tire of slop ^[17]. The prior worth holding is that unenforced writing norms decay, and that the complaint already circulating online, about colleagues pasting chatbot output without reading it ^[16], is the behaviour a policy document does not by itself change.