Someone receives an email, notices it reads like a model wrote it, and feels vaguely cheated. Someone else points out that half of business writing already passes through autocomplete, templates, and grammar tools, and that nobody ever disclosed those. Both are right — which usually means the question is framed wrong.
"Did AI write this?" stopped being a useful question the moment the answer became "partly, almost always." The useful question is older than AI: which of these words does the sender actually stand behind?
"AI-written" is not one thing
There's a spectrum, and lumping it together is where the etiquette debate goes circular:
- Spellcheck and grammar fixes — nobody has ever owed disclosure for this.
- Rephrasing your own points more clearly — still your points.
- Drafting the whole message from your bullet notes — now the framing is the model's.
- Composing and being sent unread — now nobody stands behind anything, and the recipient is corresponding with an unsupervised process wearing your name.
The felt offense tracks that last distinction, not the tooling. Fluent prose used to be costly, so it signaled attention. When a model makes fluency free, the signal breaks — and recipients who feel cheated are reacting to implied attention that was never paid. Nobody is owed your keystrokes. They're owed an accurate sense of what you vouch for.
Blanket disclosure marks everything, distinguishes nothing
The common proposal — a footer saying "drafted with AI assistance" — fails on information grounds. Applied to every message, it distinguishes none of them. It reads like the fine print on a receipt: technically true, universally ignored, and no help at all with the question that matters, which is never about the message. It's about the sentence — is this apology his, is this deadline hers, did a human actually agree to these terms?
Disclosure at the message level answers a question nobody asked. Attribution at the sentence level answers the one everybody does.
The parts that need an owner
Etiquette you can apply today, in any mail client, no special tools:
- Write the ask yourself. If the message requests something, that sentence should be yours — it's the one they'll act on.
- Own every commitment. Deadlines, prices, promises, agreements: type them or verify them word by word. "The model drafted it" will not be a comfortable position later.
- Never delegate the human parts. Apologies, thanks, condolences, praise. A generated apology is worse than none; if it wasn't worth your attention, the recipient will correctly conclude the sentiment wasn't either.
- Check facts stated as yours. A model's confident wrong claim, sent under your name, is your wrong claim.
Do that, and the drafting question mostly evaporates: the model produced the scaffolding; you produced the parts that bind.
Etiquette, made structural
The unsatisfying thing about all of the above is that it runs on discipline, and the recipient can't see your discipline. They still have to guess.
That's the gap provenance labels close. Every field in a RelayLink briefing carries its origin: assumptions marked stated by sender or inferred by sender's AI, replies marked as typed via the reply link, and the sender's note labeled verbatim, human-authored only when the words genuinely differ from anything the AI drafted. Approving a draft is recorded as exactly that — approval, not authorship. The label can't be granted by the model or asserted by the sender; it's earned or it's absent.
If you'd rather not wait for the norms to settle on their own, setting team norms for AI-drafted communication turns this into something a team can actually agree and write down.
Manners don't scale — they depend on everyone guessing everyone else's process charitably. Records scale. The etiquette debates will settle wherever the record is unambiguous, and AI correspondence will be trusted not when senders get more virtuous, but when recipients no longer have to take virtue on faith.