Check What They Decided, Not Whose Project It Was

Is describing a team's AI project in the first person a lie?

Rarely, and the format is part of why. An interview asks what a candidate did, in the singular, and a scorecard has one column with one name at the top. Almost no AI work inside a company happens alone: someone picks the tool, someone writes the evaluation set, someone signs off on what ships. A person answering the question as asked has to speak in the first person or refuse it.

But a lie is available here, and it looks different: a role that never existed, a title never held, a system the person never opened.

The pronoun is not where the line sits. Saying I about a project with four other people in it is a claim about scope. Scope claims stretch in both directions: some candidates round up, some round down, and the modest ones lose interviews for it.

Saying I built the retrieval pipeline when somebody else wrote every line is a claim about identity, and identity claims can be checked against a record.

Read your own notes before you read the candidate. If the debrief line says led the AI rollout and nobody wrote down which decision that involved, the sentence belongs to the round.

Notes round up too, and they do it silently, weeks before an offer exists.

A rollout, like the one in that debrief line, needs its own reading. Leading a rollout is adoption work rather than proof of craft, so a finalist who ran one honestly has a story about persuasion, procurement and training, and little to say about building anything.

Nothing in that account is dishonest. It answers a question about scale, and the offer was probably priced on craft.

Ask which decisions were theirs, in one call

Twenty minutes and three questions, before anything else moves. Which decisions inside that project were theirs. What they would do differently if they started it again on Monday. Who else would describe the work the same way they just did. A candidate who can name a call they made, and the thing it cost, is describing collaborative work honestly, whatever pronoun they used in the loop.

Ask about a decision, and let them pick which one.

A strong answer arrives with coordinates. A date, an option that lost, a person who disagreed at the time, a consequence somebody paid for.

Candidates who did the work can usually say what almost went wrong, because the near miss is the part they still think about.

A thin answer keeps climbing a level. Asked which call was theirs, it returns the project's outcome. Asked again, it returns the team's process. Nothing lands on a person, and two follow-ups are enough to see that; a third is unkind and adds nothing.

The question is not new to the loop, only late. The portfolio version asks who did the work on an AI-built portfolio, and the engineering version asks what an AI feature really was. Both belong in the round, where a thin answer costs a conversation rather than an offer.

When does the difference become worth acting on?

When the claim was false in a checkable way and the decision rested on it. A credential nobody conferred, a system the person never opened, a certification with an issuing body that answers the phone. Scope inflation is a different thing, and an interview that rewards a confident story selects for it.

If a third-party screening company produced the report you are about to act on, the order of operations is fixed.

Federal law, restated in the FTC and EEOC's joint guidance for employers, requires a copy of the report and a written summary of rights before the adverse action, so the person can read it and explain it ¹.

A phone call you made yourself to a former manager falls outside that sequence.

Whether the offer can be withdrawn at all is a question for employment counsel in your jurisdiction. Asking costs less before the conversation than after. The neighbouring case, rescinding an offer after the take-home broke your AI rule, at least has a stated rule behind it.

No rule was stated here, which makes the ground thinner.

Ask one procedural question before anything else: has the person resigned. An offer pulled from someone still employed and an offer pulled from someone who gave notice on Friday are different acts, whatever the law in your state turns out to say about either.

Call someone who was in the room, and weigh what comes back

A corroboration call is worth making and worth discounting. The US Office of Personnel Management's guidance on reference checking, written for federal hiring, describes the version it recommends: a structured phone conversation, the same questions for every contact, about 20 minutes each, with a minimum of three contacts ². Written requests, the same page notes, typically draw low response rates and less useful information.

Do not read the finalist's account aloud first. A reference asked to confirm a story confirms it. Ask each contact to describe the project and say who made which call, then compare the versions afterwards, which is the only part of the exercise that produces information.

Discount what comes back.

In 2022, Sackett, Zhang, Berry and Lievens rebuilt the meta-analytic estimates for selection methods. Reference checks were one of eight procedures they set aside for insufficient information, leaving no revised figure at all ³. The number still quoted for them, .26, is Schmidt and Hunter's 1998 estimate, carried forward unrevised.

None of that means a reference call tells you nothing, only that it cannot carry the decision by itself.

A senior offer usually buys judgment under load, and that is the argument for acting anyway. Someone hired for calls they have not made before will be asked to make them anyway, and the team carries the gap.

Write the next round so the decision is what gets asked for

Change the question, because the question produced the answer. A round that asks candidates to describe a project gets a project described. A round that asks for one decision, its date, what it cost and who disagreed at the time gets something a second person can check. The second version is harder to inflate and much harder to borrow from a team.

In the 2016-2017 application cycle, one orthopaedic surgery residency program verified every publication its applicants had listed. Thirteen of 1,100 citations were misrepresented, 1.18%. The same program had found 18.0% when it ran the check in 1999 ⁴.

The authors credit the PubMed identifier the application form had started requiring, and offer that as a hypothesis rather than a tested cause.

One program, one specialty, and a claim that happens to sit in a public index.

AI work has no such index. A decision does leave traces: the pull request, the ticket, the version of the deck, the evaluation spreadsheet, the colleague who argued the other side. Ask for one of them and the claim stops being a story about a project.

The same test works earlier, on the application itself. Three things on an application you can check resolve against a record. A description of how somebody thought while working does not.

Asking for one decision is the closest an interview gets to something checkable.

What the interviewer writes down also carries information.

Among 7,650 people hired at one Chinese internet technology company between 2016 and 2018, the number of job-related capabilities an interviewer named in the post-interview notes tracked later performance and promotions ⁵.

The effect was small, roughly 2% of a performance measure per standard deviation. Only candidates who passed the interview could be observed at all.

The debrief line is still sitting in your applicant tracking system, and it still says led the AI rollout. Rewrite it as the decision the person named, the date it was made, and who argued the other way. Then it belongs to somebody.

See a sample report >