AI disclosure

“AI-assisted” is the disclosure nearly every policy I know would accept, and it is close to useless. It satisfies the letter of the rule while saying nothing about what I reasoned through and what I handed off — which is exactly the gap my research is about. So here is the longer version.

This site was rebuilt with Claude, in Claude Code, and Codex, over a few working sessions in August 2026, when I was procrastinating over some other work... The model wrote most of the code you are looking at: the three-column layout, the hover reveals in the research list, the figure lightbox, the photo easter egg behind my headshot, the typography setup, and the scripts used to verify all of it. It deleted the previous design’s components once this one replaced them. It also ran an accessibility audit that caught four text-contrast failures sitting just under the WCAG AA threshold (that I just learnt) — in code it had written for me earlier the same day.

The research, and every claim this site makes about it. Each project summary began as my own notes — the conversation with Jessica that became the disclosure work, the hours of thematic analysis behind the qualitative tooling, the coding agent that informed me a file I was actively editing was “apparently not in use.” The model tightened that prose; I sent it back whenever it stopped sounding like me.

What the site should feel like, which parts to cut, when something was finished, and which of the model’s suggestions were quietly wrong: those were judgments, and judgments delegate poorly.

Worth recording, because this is the part disclosures usually leave out. Partway through, the model decided the hand-drawn underline beneath my headline was broken. It was not. Its evidence came from screenshots taken in a headless browser that renders that particular element badly, and on the strength of that evidence it rewrote working code three times, each attempt more confident than the last.

I opened the page on my own screen, saw the underline sitting exactly where it belonged, and told it to put everything back. Nothing in the model’s output announced that it was wrong. The thing that caught it was me looking.

This statement took longer to write than several of the features it describes. That is not a complaint; it is the finding. Honest attribution asks you to reconstruct decisions you made quickly and half-consciously, and to write down the moments you did not come out looking sharp. A checkbox costs nothing. This costs something, and the cost lands on the person being most forthcoming.

Whether that cost is fairly distributed — between students and instructors, between the people who write disclosure policies and the people who have to satisfy them — is the question my collaborators and I are working on, most recently in a CRAFT session at FAccT 2026.

If you are here because you are drafting a policy, or because you are a student trying to work out what is fair to write down: the question I would ask is not “did you use AI?” It is what you decided, and how you would know if you were wrong.