What's a Safe Answer When the EPLI Form Asks About AI in Hiring?

Why is 'no' almost always the wrong answer?

Because the question is not whether you bought an AI hiring product. It is whether a model's output touches an employment decision, and for most employers it already does. Resume ranking and match scores inside the applicant tracking system, sourcing recommendations, knockout scoring on screening questions, interview note-takers that produce summaries and suggested ratings: none of those arrived through a procurement decision anyone remembers making.

The person filling in the form usually cannot see all of it. Hiring sits in HR, the applicant tracking system sits in IT, the sourcing seat was bought by a recruiter on a corporate card, and the note-taker switched itself on inside the video conferencing account by default. Last year's honest no was written by someone with the same partial view.

Vendors also renamed the thing. A feature sold as smart matching, fit scoring, candidate insights or automated shortlisting is the same object the questionnaire is describing, and the marketing word is not the answer. Look at the output rather than the label: if it returns a number, a tier, an ordered list or a yes-or-no about a person, it is inside the question even when the contract never says AI. The ATS keyword filter that stopped working once every resume matched is often the same product one module later.

So the sequence is inventory first, answer second. Filling in the form from memory is how a true answer quietly becomes a false one, and the inventory takes an afternoon of asking four people what they bought.

What counts as AI in hiring on this form?

Answer the definition printed on the form. When the form gives none, borrow a written test instead of using your intuition. New York City's rule defines a tool that substantially assists or replaces discretionary decision making three ways: relying solely on a simplified output such as a score, tag, classification or ranking; weighting that output more than any other criterion in the set; or using it to overrule conclusions drawn from other factors, including human decision-making ².

That is a New York City standard for New York City hiring, enforced by the Department of Consumer and Worker Protection since 5 July 2023 ¹. It travels anyway, because it is written down, it turns on what you do with the output rather than on what the vendor calls it, and an underwriter reading your answer can follow it.

Run it per tool and per stage. Three buckets usually fall out:

• Usually inside. Resume ranking or scoring, knockout logic with a scored result, matching that reorders who gets read, an assessment returning a tier, a chatbot that ends the process on certain answers.
• Usually outside, and worth saying so on the form. Drafting a job post, a transcript with no rating attached, scheduling, a search whose ordering nobody treats as a decision.
• Argued both ways. Interview summaries a reviewer reads before scoring, and sourcing recommendations that decide who gets contacted at all.

Carriers write their own definitions and they are not always a statute's, so the printed wording governs your answer. When it is genuinely ambiguous, describe the tool rather than classifying it. A description can be incomplete. A checkbox can be false, and false is the word that matters later.

Write the inventory: one line per tool

One row per tool per stage, with the same columns every time. The shape matters because an underwriter, a plaintiff's lawyer and your own general counsel all want the same handful of facts, so writing them once means never assembling them under deadline. Keep it as a document you can date, version and send, not as prose squeezed into the questionnaire's text box.

Two rows commonly get left off, and both belong in. The first is anything a staffing agency, an RPO or an executive search firm runs on your behalf: the form asks about your employment decisions, not about who holds the software licence. The second is the tool one manager brought in alone, which you find by asking people rather than by reading a procurement report.

Then write the two sentences the whole inventory exists for. What the tool produces, and what a human does with it before anyone is rejected. If nobody can name that human, that is the finding, and it is cheaper to find now than in a deposition.

What happens if the answer turns out to be wrong?

Not a fine. What you risk is the argument about coverage, at the moment you need the coverage. California's Insurance Code states the rule plainly: a representation false in a material point, whether affirmative or promissory, entitles the injured party to rescind the contract from the time the representation becomes false ⁵. Materiality turns on the probable and reasonable influence of the facts on the party the disclosure was owed to, not on the event ⁶.

Those two sections are California law, enacted in 1935 and still the operative text. Your policy is governed by the law named in it, and your state's analogous provision may set a different bar, so the question to send your broker and your counsel is which provision governs misrepresentation on this application and what it takes to trigger. That is a question for counsel, not an answer a page can give you.

The clause worth reading twice is from the time the representation becomes false ⁵. A no that was true in January stops being true the afternoon someone switches on a scoring module, and nothing about that afternoon feels like an insurance event to the person doing it. Because materiality is measured by what the underwriter would reasonably have wanted to know, the fact that the new module had nothing to do with the eventual claim does not settle the question ⁶.

So put the trigger where tools actually get approved. Any new tool, module or stage that returns a score, an order or a filtered list about an applicant goes to whoever signed the last application, before it goes live. That is one line inside what belongs in the AI hiring policy and who signs it, and it is the part of this that has to survive staff turnover.

Which documents sit behind each line?

Five per tool, and you can name them today. The audit, with its date and the auditor's name. Your own selection rates by stage. The notice candidates were given, in the version actually sent. The written human-review step. The retention rule saying how long each of those is kept and by whom.

The second one is the document employers most often do not have, and federal law has asked for it since 1978. The Uniform Guidelines say users should maintain records disclosing the impact their tests and other selection procedures have on employment opportunities by race, sex and ethnic group, and say a selection rate under four-fifths of the highest group's rate will generally be regarded by federal enforcement agencies as evidence of adverse impact ⁴. The same section adds that federal enforcement agencies may draw an inference of adverse impact from a failure to keep that data where a group is underutilized ⁴. An empty file is not a neutral file.

Retention has a floor under it. Personnel and employment records must be preserved for one year from the making of the record or the personnel action, whichever is later, and once a discrimination charge is filed the employer must preserve all personnel records relevant to it until final disposition ⁷. Screening logs and tool outputs are personnel records when they decide who moves forward.

Three requirements are specific and dated under the New York City rule: a bias audit within one year of use, a published summary of results, and notice to candidates at least ten business days before the tool is used ¹. There is also a trap inside the vendor's audit. Once you have used the tool, you may rely on an audit built from other employers' historical data only if you supplied your own data to the independent auditor ². A vendor PDF covering somebody else's applicants is not the document it looks like. Raise that first when a screening vendor says the tool is bias tested, and expect running the adverse impact numbers on your own pipeline to end up as your job rather than theirs.

If those documents are missing, you have company, which is not the same as being covered. In a 2024 study of Local Law 144, 155 student investigators recorded 391 employers: 18 had posted an audit report and 13 had posted a transparency notice ³. The authors call the gap null compliance, since employers hold wide discretion over whether a tool is in scope, so a missing document does not prove a breach. An underwriter reading your answer faces the same ambiguity, which is why the inventory travels better than the yes. Candidates read the notice and the audit summary too, and writing the inventory so it survives being read by the people it describes is the cheapest test of whether the answer on the form is true.

Read the evidence >