AI Disclosure

I use AI in my work and in my correspondence, and the communities I write to hold different — and shifting — expectations about whether and how that should be said. Rather than guess at each community's line, I disclose by default. The words in my email signature link to the sections below; each says exactly what that kind of use means, and in any given message I delete the words that don't apply. What remains is a claim about that specific message.

The practice exists because disclosure only happens when it is cheap. A standing sentence in a signature costs nothing to send and nothing to read past; a per-message confession is enough friction that it would quietly stop happening. The longer argument — why I think disclosure is owed at all, and why the norms differ so much inside and outside the university — is in the essay The Mores of Use.

How the tools fit my workflow, in one paragraph: I write knowing that an editor will pass behind me. That knowledge lets me write faster and often more plainly, because I don't have to polish while I think — the language will be shortened. I also work with AI as an interlocutor and as research staff, in ways the sections below separate. Everything sent under my name I have read, and everything claimed under my name I stand behind.

edited — grammatical assistance

The wording passed through an AI: either I wrote the text and the tool proofed it — grammar, spelling, cutting words, untangling sentences — or the tool drafted prose to my direction and I revised and approved it. Either way the ideas, the structure, the claims, and the decisions about what to say are mine; the assistance was in wording, not thinking. This is the disclosure I attach most often, and it is the one that most resembles what a copy editor or a sharp-eyed friend has always done.

ideated — idea partnership

The content was developed in dialogue with an AI. Framings, arguments, objections, and examples emerged in conversation, and I cannot cleanly separate which thoughts started on which side of the exchange. I made the judgments, verified what could be verified, and take full responsibility for what is asserted — but it would be inaccurate to present the thinking as mine alone, so I don't.

supported — research support

An AI did supporting work behind the text: searching literature, writing or running code, processing data, checking calculations, summarizing sources. I verified what it reported before relying on it. Unless edited or ideated also appears, the prose itself and the ideas in it are mine.

reviewed — error check

I wrote this, and I made any changes myself. An AI read the text and flagged problems — grammatical slips, factual errors, numbers that don't match their sources, claims that overreach, a tone likely to land wrong — but did not supply the wording. The tool's contribution was a list of objections; the prose that answered them is mine. Where I adopt an AI's suggested wording verbatim, the honest word is edited, not reviewed.

no AI

No AI touched this message: not the writing, not the editing, not the thinking behind it. I attach this where the norms of the exchange call for an unassisted hand, or where a correspondent has told me they prefer one.

The signature line

The fragment as it appears when every kind of use occurred — in practice I delete what doesn't apply:

And when none did:

My AI assistant follows a standing instruction set when it drafts or edits mail for me — which words to include, and when. A static copy is published at standing instructions (as of September 1, 2026).