The AI Rule Belongs to Whoever Reads the Work

Why does committee ownership stall?

Because the committee is answering a procurement question and the rule is a reading question. A working group of HR, legal, IT and compliance can decide whether to buy an assessment vendor, and should. It cannot decide whether a particular submission still shows you what the stage was built to show, because that judgment belongs to whoever knows the work.

You can spot a committee-authored clause by what it leaves to the reader. "Candidates are expected to use AI responsibly and to disclose material assistance" clears every review it will ever face and answers nothing. Responsibly by whose standard. Material at what threshold. Disclose to whom, and then what. The reviewer holding the submission still has to invent all three, which is exactly the improvisation the policy existed to prevent.

There is a related trap in the artifacts that come out of these groups. In an online experiment with 410 German-based HR managers, adding explanation features to recruiting recommender dashboards improved how helpful and trustworthy the tool felt to managers with moderate or high AI literacy without improving their objective understanding of it, and the most complex explanations could reduce accurate understanding ¹. That is one online experiment on simulated dashboards, and its subject is a tool interface rather than a policy document, so it settles nothing about the second. The reason to carry it anyway: the feeling of clarity and actual clarity came apart under measurement, and a clause everyone nods at is a candidate for the same gap.

The committee still has real work. Give it the decisions it can actually close: which tools are approved, what the retention period is, what the standing notice says, and who signs off when the wording changes.

What does each of the three owners actually decide?

Three lists, and none of them overlaps. Legal decides what you may ask, what you may record, and what notices attach. HR decides that the rule is the same across requisitions, that candidates see it, and that exceptions get logged. The stage owner decides what is permitted at their stage and what an acceptable answer looks like.

Legal owns the boundaries. What you may ask a candidate about their own tools, what you may keep and for how long, which jurisdictions' notice duties attach, and whether an accommodation request changes the exercise. Legal also owns the answer to "the vendor said they handle this." In the Title VII guidance the EEOC issued in May 2023 and removed in January 2025, the agency said an employer administering a selection procedure may be responsible even when an outside vendor built it, and may be responsible for the acts of agents including software vendors ². That document is withdrawn and was never law, so treat it as the agency's reading rather than a rule, but the ownership lesson holds: buying a tool does not buy an owner.

HR owns consistency. One wording, published in one place, reaching every open requisition rather than the three where a manager happened to care. HR also owns the exception log, because exceptions are what a fairness complaint is made of.

The stage owner owns the rule for their stage. What an assistant may be used for in that exercise, what the submission has to contain, and what a thin answer looks like when it arrives. This is the half nobody can delegate upward, since it depends on what the work is. The clause-by-clause version of what they are authoring is in the five clauses a candidate AI-use policy needs.

Name the owner inside the policy

One line per stage, with a role in it. Take-home: owned by the hiring manager for the requisition. Live technical round: owned by the interviewer running it. A clause with no name attached is where improvised rejections come from, because the person holding the submission has to invent a rule and will never write down that they did.

Name roles rather than people. Individuals leave, change teams and go on holiday in the middle of a process, and a policy that has to be re-issued when someone moves will simply stop being re-issued. The role survives the person.

Give each named owner three things, or the name is decoration:

• A decision they are allowed to make alone. Usually: whether this submission still carries the signal the stage needs.
• A named escalation path with a response time. The person deciding at short notice needs someone reachable, not a mailbox.
• A place to record an exception in one line. Date, requisition, what was allowed, why.

That last one does more work than it looks. Exceptions are fine. Unrecorded ones are how a process loses its own memory: six months on, nobody can reconstruct whether the rule was applied evenly, and the candidate who asks deserves a better answer than a shrug.

The smallest case is worth naming too. In a company where the founder is the recruiter, the hiring manager and the person reading the work, the split collapses into one person, and the discipline that replaces it is writing the rule down before the first submission arrives rather than after. What the rest of that minimal process looks like is set out in the smallest hiring process a two-person company can defend.

How do you keep three owners from producing three rules?

Fix the wording centrally and let the stage owner set the bar. The rule that reaches candidates is one paragraph, identical everywhere, published by HR. What varies by stage is the standard the owner applies when reading, and that gets calibrated the way any interview standard does: agree in advance what an acceptable answer looks like.

The federal government's own practice guide is a usable template here. The Office of Personnel Management names three properties: every candidate gets the same questions in the same order, everyone is evaluated on a common rating scale, and the interviewers agree beforehand on what an acceptable answer looks like ³. The third property is the one teams skip, and it is the one that decides whether two owners reading the same AI disclosure reach the same place.

What replaces an agreed standard is freeform impression, and freeform impression has been measured. In one controlled test, predictions made after an unstructured interview correlated .31 with the outcome being predicted, while a single piece of prior evidence alone correlated .65; in a third study in the same paper, 96 of 169 participants chose to run an interview in which the answers were generated at random over running none at all ⁴. Undergraduates predicting a classmate's grades in a lab, not managers reading submissions, so the numbers do not transfer. The mechanism does: low-diagnostic information dilutes good information, and people make sense of anything put in front of them, including nothing.

So calibrate before the first submission, not after the first disagreement. One meeting, three sample answers, everyone writes down their read, then compare. The mechanics of running that with a panel are in getting an interview panel to judge AI use the same way, and the sign-off and review cadence around the document itself is in what belongs in an AI hiring policy.

Read the evidence >