An AI-Disclosure Box Is Only Worth Adding If the Answer Changes Something

What Would You Do With a Yes?

Most teams, pushed on it, find the answer is nothing: they would still read the application, still run the screen, still ask the same questions. If a yes does not change a question you ask, a stage someone reaches, or a rule you apply, the field collects a liability and no signal, and it belongs off the form.

Run the three possible answers before the vendor toggle goes on:

• Yes. Now what. If the honest answer is nothing, the field is decoration. If the honest answer is that this application gets read more sceptically, write that down and look at it, because that is a penalty for candour applied to the people who were candid.
• No. Unverifiable, and the applicants most likely to write it are the ones who used a spellchecker and were not sure it counted.
• Blank. Indistinguishable from a form that timed out, a screen reader that skipped it, or a person who did not understand the question.

One version of the field does change something. Ask what the candidate used AI for on this application, in one or two lines, and the answer becomes an opening question for whoever runs the first conversation: which part, which tool, what came back wrong, what they changed. That is a field with a downstream use, which is the only kind worth adding.

Why an Attestation Cannot Be Checked

Because nothing behind it can be verified, in either direction. A yes is a self-report and so is a no, and the tools people reach for when a no looks doubtful fold under trivial edits. Holding every detector in one benchmark at a fixed 5% false-positive rate, swapping characters for lookalike homoglyphs dropped one tool's accuracy from 85.0 to 9.3, while five others lost an average of 40.6 points ¹.

Those are benchmark documents rather than applications, and one tool in the same test barely moved, so the finding is not that every detector collapses on contact. It is that the effort required to defeat one is seconds, which means the people a detector surfaces are the people who were not trying to hide anything. Pair that with the error direction and the picture gets worse: seven detectors run over 91 human-written TOEFL essays produced an average false-positive rate of 61.3%, against near-perfect accuracy on essays by US eighth-graders ². Different documents, an older generation of tools, and a false-positive rate rather than an accuracy rate. The direction is the durable part, and it points at people writing in a second language.

So an attestation is an honour-system question with no instrument behind it, which is a defensible thing to ask and an indefensible thing to enforce. The choice between trusting the answer, buying a tool, and watching someone work is the actual decision here: honour-system disclosure, detection, or observation is where that comparison gets made.

Write the Field and the Rule Together

Draft both sentences in one sitting or leave the field off. The field: what did you use AI for on this application. The rule, printed directly beneath it: what you write here is never itself a reason for rejection. The second sentence is what makes the first one answerable honestly, and a question nobody answers honestly is worse than a question never asked.

The rule has to be true in practice, which means somebody decides in advance what happens when the answer is uncomfortable. Settle whether an AI-written application is a reason to reject anyone before the field goes live, not in the moment when a hiring manager forwards one and asks.

What the answers are worth, in practice:

• "Drafted the summary, wrote the rest myself." Ordinary. Ask about the rest.
• "Used it to translate my answers from Portuguese." Says nothing about capability and everything about how the form was read. Do not treat it as a flag.
• "Used it to research your company and tighten my phrasing." Background reading and line editing. It says the applicant can aim a tool at the two things it is good at, and nothing about what happens when the tool is wrong.
• An answer describing a specific piece of judgment they kept. The interesting one, and the reason to have the field at all.

Keep the box small. Two lines of text, a stated limit, no follow-up sub-questions, and no separate consent language stapled to it. A long disclosure block reads as a legal trap and gets the answer a legal trap deserves.

What the Law Asks You to Disclose

The live disclosure duties in US hiring point at the employer. New York City bars using an automated employment decision tool on a city candidate unless a bias audit was done within the prior year, a summary of it is posted publicly, and the candidate had notice at least 10 business days before the tool is used ³. Nothing in that regime asks a candidate to declare anything.

That is one city, and it is a disclosure-and-audit regime rather than a ban, with the notice counted in business days. Other jurisdictions have added their own notice duties on different triggers and different timelines. This is public legal fact as of August 2026 rather than advice, and the specific question of what you may ask an applicant belongs with counsel: what you can lawfully ask candidates about their AI use is the version of this with the jurisdictions laid out.

An employer running an automated tool on candidates may owe them notice, an audit and a posted summary, while the candidate owes the employer nothing about the tools they used to write a document. A disclosure box that reverses that direction, collecting attestations from applicants while the employer's own use goes undescribed, is the version most likely to read badly if anyone ever prints both pages side by side.

In a vignette experiment with 921 working-age Austrians, a rejection from an AI with no explanation scored lowest on all four measures taken, and an AI rejection that came with an explanation drew the same fairness ratings as a human rejection without one ⁴. Hypothetical rejections rated by an online panel, on a scale where every condition sat below the midpoint. The finding still points at the same conclusion the field does: the explanation is the part that carries weight, and an unexplained automated decision is the worst available option.

Read the evidence >