Policy
Asking Candidates to Disclose AI Puts Your Own Stack on the Table
Requiring candidates to disclose AI use creates no legal duty to disclose your own screening tools. Every US notice duty attaches to the automated tools an employer deploys, and none of them is triggered by asking a candidate what they used. Practically, treat it as one rule anyway. Publish, in a single paragraph, what AI touches an application on your side and what you are asking of them. A company that requires transparency in one direction only loses that argument in public every time.
The takeThe reciprocity argument is not a legal one and it should stop being argued as though it were. It is a credibility argument, and credibility is the only thing making a disclosure request work at all, since nothing behind it can be verified. A candidate who answers honestly is doing you a favour on trust. Asking for that favour while declining to say what reads their resume is the cheapest possible way to spend the trust you need.
Where Olive fits
Open a role and see what the work shows
Transparency is easier to publish when nothing automated needs describing. Olive makes no automated decision at all: a human reviewer writes each of the six findings, and a released report exports with its rubric, scorer and bank versions attached.
Rank your shortlistDo the two duties have anything to do with each other?
Legally, no. Every US notice duty attaches to the tool the employer deploys, not to what a candidate does with a chatbot at home. New York City's rule turns on an automated employment decision tool being used on a candidate. Illinois attaches to an employer using AI in recruitment or hiring. Nothing in either is triggered by asking a candidate a question.
The New York City version is the most specific. Local Law 144 bars using an automated employment decision tool on a candidate in the city unless a bias audit was completed within the prior year, a summary of the results is posted publicly, and the candidate got notice at least ten business days before the tool is used. It took effect January 1, 2023, and enforcement began July 5, 2023 1.
Illinois is broader and vaguer. Since January 1, 2026 it has been a civil rights violation for an employer to use AI that has the effect of discriminating in recruitment or hiring, to use zip codes as a proxy for a protected class, or to fail to notify an employee that AI is being used for those purposes. The statute leaves the timing, circumstances and means of that notice to agency rulemaking, so the duty exists while its shape does not 2.
California is the useful contrast. Since October 1, 2025 its amended employment regulations have treated an automated-decision system as covered by the state's discrimination rules, and they require no candidate notice at all 57. What they ask for instead is four years of retained records behind the decision, including automated-decision system data 6.
Read any of the three against a disclosure request and nothing connects. The duties run on the employer's tools, they are owed whether or not you ask candidates anything, and satisfying one has no bearing on the other. If you need the full jurisdiction-by-jurisdiction version, it is in what your AI notice must say and when each jurisdiction wants it.
Why does the asymmetry cost more than the disclosure does?
Because you are asking for exactly the thing you are withholding, and the candidate can see it. A request to describe how someone used a tool is reasonable on its own. The same request from a company that will not say what reads the resume lands as transparency running one direction, and that is what turns a fair process into a story told afterwards.
The trust you are drawing on is already thin. In a Pew Research Center survey of 11,004 US adults, 66% said they would not want to apply for a job with an employer that uses AI to help make hiring decisions, against 32% who would 3. That is stated intention rather than measured behaviour, and it is not an application drop-off rate, so do not price it as one. It does say what the average candidate brings to the question before you ask anything.
What they can actually see is close to nothing. A Cornell study sent 155 investigators to record what 391 employers publicly posted under Local Law 144 in a seventeen-day window in late 2023: 18 had posted a bias audit report and 13 a transparency notice 4. The authors are careful that these are compliance rates and not non-compliance rates, since the law lets an employer decide for itself whether a tool is in scope. Either way, the candidate's experience is the same. The employer's half of the process is invisible, and that is the baseline your request lands against.
The asymmetry also does something quieter to the answers you get. A disclosure question from an opaque process gets short, defensive answers, which is precisely the failure mode described in asking candidates how they used AI. Transparency on your side is not a courtesy here; it is what makes the answers worth reading.
Write the one paragraph that holds both halves
One paragraph, both directions, in the place the request is made. Say what AI touches an application on your side, say where a person takes over, say what you are asking of the candidate, and say what their answer is used for. Four sentences. If any of the four is hard to write, that is the half of the process worth fixing before publishing anything.
A working version, for a take-home:
> An assistant helps us sort and summarise applications; every rejection at that stage is reviewed by a person before it is sent. Nothing else in this process is automated, and no AI reads this assignment. Use an assistant on the assignment if that is how you work, and write us a few lines on what you asked it for and what you rejected. Those lines reach the reviewer with your work, and they are never on their own a reason to end the process.
Four moves, one paragraph, and it fits at the top of an assignment brief. Note what it does not do: it names no vendor, quotes no statute, and makes no claim about accuracy. It describes what happens, in the order it happens, to the person it happens to.
Put it in three places, identically worded: the posting, the application flow, and the assignment email. Different wording in the three places is how a reasonable disclosure becomes an inconsistency someone quotes back to you. And keep it in the same register as the rest of the process. A paragraph written in compliance language next to an assignment written in plain English tells the candidate which one the company means.
What should you not publish?
A model inventory, a promise you do not keep, and somebody else's compliance claim. Those three are how a good paragraph turns into a liability. A list of vendors and versions goes stale within a quarter and invites a question nobody on the hiring team can answer. A human-review sentence that nobody actually performs is worse than silence, because now it exists in writing.
Describe capabilities and stages instead of products. "An assistant summarises applications before a recruiter reads them" survives a vendor switch. "We use Acme Screen v4" does not, and it will be wrong on the careers page for months after the contract ends.
Be equally careful with claims you inherited. A vendor telling you the tool is bias tested is a statement about their testing, and repeating it in your own candidate-facing paragraph makes it a statement about your process. The questions worth asking before you repeat anything are in what to ask an AI screening vendor that says it is bias tested.
Two things also belong somewhere other than this paragraph. The jurisdiction-specific legal notice, with its required contents and its ten-business-day clock, is a separate document with a separate audience. And what a rejected candidate is owed when they ask what happened is a harder conversation with duties of its own, worked through in what you owe a candidate who asks why the AI screened them out. Keep the reciprocity paragraph short enough that a candidate reads it. Everything else has its own home.
Common questions
Does any law actually require telling candidates that AI reads their application?
In some places, yes, and it has nothing to do with what you ask of them. New York City requires notice at least ten business days before an automated employment decision tool is used on a candidate, plus a posted bias audit summary. Illinois has required notice since January 1, 2026, with the timing and wording still delegated to agency rules. Its notice clause runs to an employee, which leaves an outside applicant on unsettled ground worth putting to counsel. California requires no notice and four years of records instead. The duties attach to your tools regardless of any disclosure request.
If we use no AI in screening at all, what should the paragraph say?
Say that, plainly, in one sentence: no automated tool screens applications, and a person reads every one. It is the strongest version of this paragraph available and almost nobody writes it. Check it first, though. An applicant tracking system with knockout questions, keyword ranking or a resume parser that drops candidates is doing something a candidate would call automated screening, whatever the vendor calls it.
Should the paragraph name the vendor?
No, and there is a real cost to it. Vendor names age out, get renegotiated, and prompt questions the recruiter fielding them cannot answer. Describe what happens at each stage instead: what the tool does, what it takes as input, and where a person takes over. That description stays true through a contract change, which means the version on your careers page is still accurate a year from now.
Does publishing this raise our legal exposure?
Publishing an accurate description does not create a duty you did not already have. Publishing an inaccurate one is the risk, in both directions: claiming a human review you do not perform, or describing a tool as merely informing a decision when it effectively makes it. Write what the process actually does today, have counsel read it against the jurisdictions you hire in, and re-read it whenever the stack changes.
What if candidates start opting out once we tell them?
Some will ask questions, and the ones who ask are worth answering. There is no measured drop-off rate to plan against here: survey evidence on stated willingness is discouraging in the abstract, but the abstract question is different from a specific process with a named human step in it. The higher-risk version is the reverse: a candidate who finds out afterwards that something automated decided their application, from a rejection they cannot get an explanation for.
References
- 1. Automated Employment Decision Tools: Frequently Asked Questions nyc.gov Supports the bias-audit and posting duties, the ten-business-day notice period, and the effective and enforcement dates.
- 2. HB3773 Enrolled (Public Act 103-0804), amending the Illinois Human Rights Act ilga.gov Supports the January 1, 2026 effective date, the effects standard, the zip-code proxy ban, the notice clause and its delegation to agency rulemaking.
- 3. Americans' views on use of AI in hiring (chapter of 'AI in Hiring and Evaluating Workers: What Americans Think') pewresearch.org Supports the stated-intention figures on applying to an employer that uses AI to help make hiring decisions.
- 4. Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability arxiv.org Supports how few of the observed employers had posted an audit report or a transparency notice, and the authors' framing of these as compliance rates.
- 5. Final Unmodified Text of Proposed Employment Regulations Regarding Automated-Decision Systems (Attachment B), 2 CCR sections 11008, 11008.1 calcivilrights.ca.gov Supports the October 1, 2025 effective date and the absence of any California candidate-notice duty.
- 6. Final Unmodified Text of Proposed Employment Regulations Regarding Automated-Decision Systems (Attachment B), 2 CCR sections 11009, 11013 calcivilrights.ca.gov Supports the four-year employment-records retention period, which expressly includes automated-decision system data.
- 7. Rulemaking Actions - Civil Rights Council ✓ calcivilrights.ca.gov The Council's own record of the automated-decision-system employment regulations: approved by OAL and filed with the Secretary of State, effective October 1, 2025.
7 sources, numbered by first appearance. How Olive sources claims
General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.