TL;DR: AI made writing free and left reading as expensive as it always was, so every inflated message moves work from the person sending it to the person receiving it. The meter is bolted to the side that got cheaper, and nothing counts what lands on the other. The padding a model adds is billed twice, once in tokens and once in attention. Five daily AI users voting together now spot machine prose 299 times out of 300. Change what you ask the tool for: cut the draft in half every time. Send nothing that came back on the first prompt.
Want to listen to this article?
Subscribers to the Orion Playbook newsletter can listen to the AI-Generated Audio version of this article for free. Already a subscriber? Log in here
I ask a model to turn three bullets into a warm, professional email. You ask a model to turn it back into (hopefully the same) three bullets. Two inference bills, one message, nothing added. A data center is playing telephone with itself, and the two people it was built to connect are further apart than when we started.
You know what the middle of that trip looks like. You read four of them before breakfast. The opening line that agrees with you before it says anything. The list of three where items two and three are there for rhythm. The paragraph that politely restates the paragraph above it. The objection answered before anyone raised it. Someone writing by hand runs out of appetite for that. A model never does. Each of those sentences was billed twice: once as output tokens to the sender, once as attention to you.
Producing text got cheap; reading it costs the same as in 1995. The writer picks the length, the reader absorbs it, and the meter is bolted to the side that got cheaper. Nothing counts what lands on the other.
Back in May, I put three questions in front of anything I published, the first being whether it was worth sharing at all. I meant it as a question about authorship. It turns out to be a question about arithmetic. A message spends somebody's day without asking, and the price is set on their side. Todd Rogers and Jessica Lasky-Fink wrote a whole book about what that costs the person receiving. LLMs seemingly missed it during training.
There is no offender to discipline. More than half the people who complain about the flood admit to producing it (and yes, I am one of them), which is what a structural problem looks like, and structures are immune to etiquette guides. Where it starts is not the person writing the email. It is the instruction they were handed. A company president, quoted anonymously in Harvard Business Review: "They want the use of AI, but other than saying 'use it everywhere every day,' there is no place our team can use it to actually be viewed as successful."
Three years ago, the reassuring finding was that nobody could reliably tell AI writing from human writing. People still repeat it. It stopped being true in 2025, when the test was run again on people who use these tools daily. They picked machine-written prose out with no training, and five of them voting together got 299 of 300 right. Paraphrasing and humanizer passes made no difference. Those daily users are your board, your best engineer, and the recruiter reading your note. The polish you added to show you cared is what tells them you did not.
O email, you were never the problem.
We had three bullet points to send.
We asked a machine to make them professional.
You carried a full page.
At the other end, another machine was waiting.
It read the page and gave back three points.
Two were the ones we sent.
One was new.
Only now nobody can tell whether the message came through.
Stop telling the organization to use AI more, and start telling it which way to point: the default ask is to cut the draft in half, because "make this more professional" has only ever been a synonym for longer. Then the rule I run on myself. Nothing goes out that came back on the first prompt. The term of art is n-shot, and it is unglamorous. I will go around with a model six times or more before I sign anything, and there are days when writing it myself would have been quicker.
When English is your third language, or the empty screen is where you stall, the model gets you to a draft, and it lets great people into conversations the blank page was keeping them out of. That is where its job ends. What comes back is generic by default, and turning it into something worth another person's minutes is yours. Brevity has its own bill, and it lands on the newest person on the team, who needed the reasoning you cut; nobody has measured whether that trade is worth making. Communication succeeds only when the receiver receives.
In 1657, at the end of a long letter, Pascal apologized for its length: "I have made this longer than usual because I have not had time to make it shorter." His excuse remains true in the age of AI. The thinking was never free, and spending it is the one courtesy a machine cannot perform for you. When everyone writes well, the writing stops telling your reader anything. What still tells them something is how much of your own time you were willing to spend so they could spend less of theirs.
All the images were generated with AI (ChatGPT Images, Claude Opus) by Gérard Métrailler.
Sources
Niederhoffer, Kate, Alexi Robichaux, and Jeffrey T. Hancock. "Why People Create AI 'Workslop'—and How to Stop It." Harvard Business Review, January 16, 2026. https://hbr.org/2026/01/why-people-create-ai-workslop-and-how-to-stop-it. Accessed 2026-08-08.
Russell, Jenna, Marzena Karpinska, and Mohit Iyyer. "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text." arXiv:2501.15654, January 2025. https://arxiv.org/abs/2501.15654. Accessed 2026-08-08.
Rogers, Todd, and Jessica Lasky-Fink. Writing for Busy Readers: Communicate More Effectively in the Real World. New York: Dutton, 2023.
Pascal, Blaise. Les Provinciales, Letter 16. 1657.
