ChatGPT emailers, why everyone can tell, and why that matters

You got one this week. The email that's slightly too smooth, warm in a way that has no temperature, organized like a press release about nothing. You knew within two sentences. Everyone knows within two sentences. The interesting questions are how you knew, and what the knowing does.

How everyone can tell

People assume it's vocabulary, and vocabulary is part of it. The famous tells are real: the delving, the "I hope this finds you well," the landscapes and tapestries and testaments. But you can strip all of those out and the email still smells wrong, because the deeper tells are structural. Every sentence arrives at the same temperature. Every claim comes pre-hedged, balanced against a counterpoint nobody asked for. The structure is flawless and nothing is at stake in it. Human writing has a pulse: it speeds up where the writer cares, gets clumsy where they're unsure, skips what they assume you know. Model writing has no pulse, because the model doesn't care more about one sentence than another. Uniform care is the fingerprint.

And humans are tuned to detect exactly this, because we've each read tens of thousands of emails. You have a distribution in your head of what a colleague writing at 4:50pm sounds like. The smooth message sits outside that distribution, and you notice the way you notice anything uncanny: instantly and pre-verbally, before you could name a single tell.

Why it matters: the effort was the message

Here's the model that makes sense of the discomfort. Certain emails were never really information transfer. The cold intro, the thank-you note, the apology, the "thinking of you for this role" note: these are bids in a relationship, and what made a bid credible was that it cost something. Ten minutes of a specific person's attention, spent on you, unrecoverable. Economists call this a costly signal. The content was partly a receipt for the cost.

An email used to prove someone spent attention on you. That proof is what died.

AI writing destroys the signal without touching the words. When a recipient detects the machine register, what they've actually detected is that the marginal cost of this message was zero, and therefore that it claims nothing about how much you valued them. That's why an AI apology lands worse than a clumsy one and an AI condolence lands worst of all. Nobody is offended by the grammar. They're responding, correctly, to the absence of spent attention. The message said "you were worth my time" in words and "you were not" in evidence, and people believe evidence.

The line that actually matters

The conclusion is not "never use AI for writing." I use AI constantly, including on parts of this site, and pretending otherwise would be its own kind of fake. The line is between AI as a tool for the work and AI as a stand-in for the relationship. Let it research the company, structure the argument, tighten the draft, catch the error. Those are work, and doing work faster is just competence. But the part of the message that performs the relationship, the part whose entire function is to demonstrate that a person spent attention, has to actually contain spent attention. Outsource that and you haven't written the message efficiently. You've sent a different message.

What replaces polish

The practical upshot, since polish is now free and therefore worthless as a signal: specificity is the new costly signal. One sentence that could only have been written by someone who actually read the recipient's work, knows their situation, remembers the last conversation, is worth more than four paragraphs of frictionless prose, because that sentence is unfakeable by a model that doesn't know them. Short and specific now beats long and smooth, exactly because everyone knows what long and smooth costs.

The email isn't dying. The free-riding on email's old signal is dying. What's left is the version where you have to mean it, which is arguably the version that was worth sending all along.