Sofia is the candidate-experience lead at a 1,200-person fintech company with a recruiting team of 14. Last spring her team rolled out an AI resume-screening tool for high-volume roles, and for the first quarter nobody told candidates it existed. The tool worked well. The problem arrived as a LinkedIn post from a rejected applicant: "Found out the hard way that this company runs your resume through an algorithm before a human ever sees it. Would have been nice to know." The post collected 600 reactions in two days and three of Sofia's open requisitions saw application rates drop. Nothing about the tool was illegal in her state, and nothing about it was unfair. What broke was trust, and it broke because the disclosure was missing. Sofia rebuilt the entire program around a single principle: candidates should never discover the AI, they should be told about it before it ever touches their application.

The Question Is How to Disclose, Not Whether

Using AI in recruiting without telling candidates creates a trust problem the moment it is discovered. Candidates feel manipulated, they post negative reviews, and the legal exposure rises rather than falls, because a process nobody was told about is harder to defend than one that was described in advance. Sofia learned that the choice her leadership thought they were making, disclose or stay quiet, was never really available. The tool would be discovered eventually. The only genuine decision was whether candidates would hear about it from her or from a stranger on social media, and that decision is what determines the outcome.

Which means the useful question is how to disclose in a way that is honest, clear, and preserves candidate confidence in the fairness of the process. Done badly, disclosure alarms people without informing them, or makes claims the company cannot back. Done well, it does something most companies do not expect: it builds trust rather than eroding it, because a candidate who is told plainly what the process does and where a human sits in it has been treated as an adult. The rest of this lesson is the program Sofia built, starting with the places where disclosure stops being a choice at all.

When Disclosure Is Legally Required

Disclosure is not only a courtesy. In a growing number of jurisdictions it is the law, and the rules are specific enough that "we mention AI somewhere on our careers page" does not satisfy them.

New York City Local Law 144. If you use an automated employment decision tool, or AEDT, to screen a candidate for a job located in NYC or a candidate who is an NYC resident, the law requires you to notify each candidate at least 10 business days before the tool is used. The notice must tell the candidate that an AEDT will be used, identify the job qualifications and characteristics the tool will assess, and explain how the candidate can request an alternative selection process or accommodation. The law also requires that the tool has passed an independent bias audit within the past year and that a summary of the most recent audit be publicly available. Ten business days is roughly two calendar weeks, so a candidate who applies cannot be run through the tool the same afternoon.

GDPR, for EU candidates. If Sofia's company processes applications from the European Union, Article 22 gives candidates the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, with narrow exceptions. Where automated processing is used, Articles 13 and 14 require that candidates be informed of its existence and given meaningful information about the logic involved and the consequences. In practice this means a solely-automated rejection needs either a lawful basis and safeguards or a genuine human in the loop, plus clear notice up front.

EEOC and the ADA. Under federal anti-discrimination guidance, an AI tool that screens out candidates with disabilities, even unintentionally, can create liability. Candidates must be able to request a reasonable accommodation and an alternative process. A video-analysis tool that scores speech patterns, for example, may disadvantage a candidate with a speech disability, who is entitled to a different evaluation path on request.

Legal requirements vary by jurisdiction, so beyond the specific statutes above Sofia works from a set of general triggers that tell her when a disclosure is warranted regardless of where the role sits. Disclose when AI makes or significantly influences a hiring decision, which covers screening decisions, scoring, ranking, and interview assessment. Disclose when a tool analyzes personal characteristics, for instance a sentiment analysis tool that infers personality from a video interview. Disclose when the AI collects or stores personal data, since transparency and data privacy laws frequently require notice on their own terms. And disclose when the decision could carry legal consequences: if an applicant might later challenge the hiring decision, transparency about the process is not merely polite, it is legally important. Where the requirement is genuinely unclear, industry standards and company values close the gap. If you tell candidates you respect them, you disclose.

What Candidates Actually Need to Know

A disclosure that satisfies a lawyer can still fail a candidate. Sofia built her standard notice around eight things a candidate genuinely wants answered.

What AI is used: name the tool's function plainly. "We use an AI tool that screens resumes for required qualifications." Why you use it: "To evaluate every application against the same criteria, consistently and quickly." How it works: "The tool assesses whether your resume contains our required qualifications and produces a score that helps determine who advances to a phone screen." What it assesses: "It looks at technical qualifications stated on your resume. It does not assess personality, cultural fit, or potential." Its limitations: "It focuses on keywords and explicit qualifications and may miss relevant experience described in terms different from ours." Human review: "A recruiter reviews the tool's recommendations and can override them when they see qualifications the tool missed." Feedback: "If you are screened out, we tell you what would strengthen a future application." Fairness: "We monitor and audit the tool for fairness across candidate groups."

The line between honest and reassuring matters. Sofia's rule is that every sentence must be true and verifiable. If she cannot show an audit, she does not claim one. A disclosure that promises fairness the company has not actually tested is worse than no disclosure at all, because it converts a procedural gap into a credibility problem the moment a candidate asks for proof. The same rule cuts the other way on limitations: naming what the tool can miss reads as candor and costs nothing, because candidates already suspect it.

How and Where to Disclose: A Worked Example

Disclosure is not a single document. It is a sequence of touchpoints, each matched to where the candidate is in the process. Here is the exact sequence Sofia uses for a NYC-based analyst role that runs through her AEDT, built to satisfy Local Law 144's 10-business-day notice.

Step 1, the job posting. The notice lives in the posting itself, so the clock can start as early as possible:

"This role uses an automated employment decision tool to assist with screening. The tool assesses the technical qualifications and years of relevant experience stated on your resume. It does not assess personality or cultural fit. A recruiter reviews all results and can override them. You may request an alternative screening process or a reasonable accommodation by replying to your application confirmation email. A summary of our most recent bias audit is available at the link below."

Step 2, the application confirmation email, sent the day the candidate applies. This is the formal start of the 10-business-day notice window:

"Thank you for applying to the Analyst role. As noted in the posting, we use an AI screening tool that evaluates the technical qualifications on your resume. We will not run your application through this tool until at least 10 business days from today, which is [DATE]. If you would prefer an alternative screening process, or need an accommodation, reply to this email before that date and we will arrange it. You can review our bias audit summary here: [LINK]."

Sofia's applicant tracking system stamps the application date and calculates the [DATE] field automatically. If a candidate applies on a Monday, the screen cannot run until two Mondays later, accounting for weekends and any company holidays in between. Of the roughly 240 applicants the analyst role drew last cycle, 11 replied asking about the process and 4 requested an alternative review, which a recruiter handled manually. None of them learned about the AI from a stranger on social media.

Step 3, the interview confirmation. If the company also uses AI to assist interview scoring, the notice repeats at that stage: "We use a tool to help structure and score interviews for consistency. Every candidate is also interviewed by humans who assess additional dimensions like collaboration and communication." Step 4, pre-hire. Before extending an offer, the recruiter confirms the candidate understands how AI was used across the whole process, closing the loop so no surprise can surface after they join.

Getting the Wording Right: Clear, Not Vague, Not Overstated

Sofia keeps three sample sentences on the wall of the recruiting workspace, because most disclosure failures are wording failures rather than policy failures. The clear version reads: "We use AI to screen for basic requirements. A human reviews all candidates advancing to interviews." It names the function, bounds the scope, and locates the human, all in two sentences a candidate can act on. Everything Sofia's team writes is measured against that standard before it goes into a posting or an email template.

The two failure modes sit on either side of it. The vague version, "we use advanced analytics in our process," technically discloses something while telling the candidate nothing they could use, and it tends to be written by people who want the credit for transparency without the discomfort of specificity. The overstated version, "AI makes the final decision," is the opposite error: it describes a process more automated than the one actually running, which alarms candidates unnecessarily and, worse, misdescribes where accountability sits. Transparency builds trust. Vagueness creates suspicion. Overstatement creates a different kind of problem, since a candidate told the machine decided has been given a false picture of a process a human actually controls.

Addressing Candidate Concerns Without Defensiveness

Once you disclose, candidates ask questions. Sofia trained her 14 recruiters to treat each question as a chance to build confidence rather than a complaint to deflect.

"Will the AI be biased against me?" The honest answer points to evidence: the fairness testing, the human review step, and the feedback mechanism. "Is this legal?" "Yes. We comply with employment law, including the local notice and audit requirements, and we are transparent about exactly how the tool is used." "Can I opt out?" Answer truthfully. If an alternative exists, explain it and any implications, such as a slightly longer timeline or a different evaluation method. If it genuinely does not, say so and explain why, rather than implying flexibility that is not there. "How do I know it does not discriminate?" Share what you actually have: the audit summary, disparate-impact testing results, and the human oversight built into the process.

Sofia keeps a one-page internal FAQ so every recruiter answers consistently. The worst outcome is two recruiters giving a candidate two different stories about the same tool, which reads as either disorganization or evasion. She also treats the questions themselves as data. When four candidates in a cycle asked whether the tool reads cover letters, that was a sign her posting language was not specific enough about what the tool assesses, and she fixed the template rather than coaching recruiters to field the question better.

Avoiding the Three Disclosure Traps

Sofia watched three failure patterns sink other companies' transparency efforts, and she designed her program specifically to avoid them.

The vague disclosure. Saying "we use technology to enhance our recruiting process" technically mentions something while telling the candidate nothing. They cannot tell that AI makes screening decisions, so the disclosure satisfies no one. It happens because organizations want the appearance of transparency without alarming candidates, and the result is that candidates feel they learned nothing, which is exactly what happened. The fix is specificity: name what the AI does, how it works, and what it assesses.

Disclosure without fairness evidence. A company discloses AI use, a candidate asks whether the tool is fair, and the recruiter has nothing to point to because no testing was done. Transparency backfires and candidate confidence drops below where it started, because the disclosure raised a question the company cannot answer. The fix is sequencing: run the audit and the disparate-impact testing before you publish the disclosure, so every fairness claim has proof behind it.

The discover-it-yourself disclosure. This is the trap that caught Sofia's company the first time. The AI runs silently, a candidate finds out on their own and posts about it, and the non-disclosure now looks like deception rather than oversight. The reputational damage exceeds anything proactive disclosure would have cost, and it lands on a timeline you do not control. The fix is the entire program above: tell candidates first, every time, through channels they will actually see.

Practice

Each of these produces language you can put in front of a real candidate, which is the only test that matters.

• Write the job posting disclosure. Draft the AI-use language you would include in a live posting, then read it as a candidate would. Is it clear what the tool does, what it assesses, and where a human reviews? Would someone actually understand it, or does it only sound like it explains something?
• Build a candidate-facing FAQ. List the questions candidates would genuinely have about your use of AI, then write the answers you could defend, checking each one against evidence you actually hold.
• Answer a fairness email. Imagine a candidate writes asking whether your AI screening is fair to people like them. Write the reply. Notice what evidence you reach for and whether it exists.
• Design the pre-hire conversation. Script what you would say to a finalist about how AI was used across the process, before an offer goes out, so nothing surfaces as a surprise after they join.
• Handle the opt-out request. Decide in advance what you would say if a candidate asked to be excluded from AI screening, including what the alternative process is, what it costs in time, and what you would say if no alternative exists.

Reflection

These are worth answering from the candidate's side of the table first, which is where most disclosure decisions look different.

• How would you feel if you discovered, after being rejected, that an AI tool had screened you and nobody mentioned it?
• What specifically would make you feel confident that an AI recruiting tool was fair to you? Would your own current disclosure provide that?
• How would you explain to your leadership why transparency about AI matters, in terms they weigh: brand, legal exposure, and offer acceptance?
• What is your biggest concern about disclosing, and what would you do to address it rather than let it become a reason to stay quiet?
• Would transparency about AI use differentiate your company positively in your market, and what would you have to be able to prove for that to work?

Glossary

• Disclosure. Informing candidates about your use of AI and how it affects their path through the recruiting process.
• Bias. Systematic preference for certain candidates over others based on characteristics unrelated to job performance.
• Automated employment decision tool (AEDT). The category of tool covered by NYC Local Law 144, which attaches candidate notice, bias audit, and public posting obligations to its use.
• Alternative selection process. A different evaluation path offered to a candidate who requests it or who needs an accommodation, rather than the automated screen.

Related Lessons

Disclosure only works when the things you disclose are true, which is why it sits on top of several other lessons.

• Consent and Transparency: What Candidates Need to Know covers the consent side of the same conversation, including where notice alone is insufficient and an affirmative choice is expected.
• Legal and Compliance Partnerships: Ensuring AI Use Is Defensible is where the bias audit and legal sign-off behind your fairness claims get built, so your disclosure has something real standing behind it.
• Bias Audit Trails: Creating Evidence of Fairness Considerations supplies the evidence a recruiter reaches for when a candidate asks how you know the tool does not discriminate.
• Data Privacy Fundamentals: GDPR, CCPA, FCRA, and Regional Requirements handles the data-collection triggers for notice, which apply even when the AI is not making the decision itself.
• Understanding Candidate Experience: What Matters in Hiring puts disclosure in the wider context of how candidates judge a process, which is what turns a compliance step into a brand advantage.

Closing

The reframe that changed Sofia's leadership conversation is this: transparency is not a compliance tax, it is a differentiator. Most companies treat AI disclosure as something to minimize. A company that does it well stands out to candidates who are increasingly aware that algorithms screen their applications, and who have started asking about it directly. After Sofia rebuilt the program, her team began mentioning the fairness audit and the human-review guarantee in recruiter outreach, and candidates noticed. The same transparency that started as damage control became a talking point that signaled the company took fairness seriously.

The cost of the program was modest: an audit, a few templated notices wired into the applicant tracking system, and a one-page FAQ. The return was a candidate experience that no longer generated viral complaints and that recruiters could describe with confidence instead of hedging. Transparency about AI use is an opportunity to build trust and demonstrate that you are thoughtful about fairness, and it is one of the cheapest such moves a recruiting team can make. Transparency is a strategic advantage rather than a liability. Use it.

Key Takeaways

• Disclosure is often a legal requirement, not a courtesy. NYC Local Law 144 requires at least 10 business days notice before an automated employment decision tool is used, identification of what it assesses, an available bias-audit summary from an independent audit within the past year, and a route to request an alternative process or accommodation. GDPR Articles 13, 14, and 22 govern automated decisions for EU candidates, and EEOC and ADA rules require an accommodation and an alternative process on request.
• Disclose whenever AI influences the decision, reads personal characteristics, stores personal data, or the decision could be challenged. These general triggers apply even where a specific statute does not, and where the legal requirement is unclear, company values close the gap.
• Tell candidates what they actually need to know. What the AI is, why you use it, how it works, what it assesses and does not, its limitations, the human review step, the feedback path, and your fairness checks. Every sentence must be true and verifiable.
• Disclose in a sequence, not a single document. Job posting, application confirmation email, interview confirmation, and a pre-hire check. For an AEDT under Local Law 144, the confirmation email starts the 10-business-day clock and the screen cannot run until that window closes.
• Be clear, not vague and not overstated. "We use AI to screen for basic requirements. A human reviews all candidates advancing to interviews" beats "we use advanced analytics," and it also beats "AI makes the final decision," which misdescribes where accountability actually sits.
• Run the fairness audit before you disclose. A transparency claim you cannot back with evidence converts a procedural gap into a credibility problem the moment a candidate asks for proof.
• Answer candidate questions consistently and honestly. A shared internal FAQ keeps 14 recruiters telling the same true story. If opting out is not possible, say so and explain why rather than implying flexibility that is not there.
• Avoid the three traps. Vague disclosure that explains nothing, disclosure without fairness evidence, and silent use that candidates discover on their own. The discovered-deception version does the most reputational damage of all.

Frequently Asked Questions

What exactly has to be in the notice under NYC Local Law 144? Three things, delivered at least 10 business days before the automated employment decision tool is used on the candidate: that an AEDT will be used, the job qualifications and characteristics the tool will assess, and how the candidate can request an alternative selection process or a reasonable accommodation. Separately, the tool must have passed an independent bias audit within the past year and a summary of the most recent audit has to be publicly available. Sofia's posting and confirmation email together cover all of it, with the email timestamped so the notice window is provable.

Can we run the screen sooner if the candidate says they do not mind waiting? Sofia does not build her process around exceptions to the notice window, because the 10 business days is the statutory floor, not a service-level target she negotiated with candidates. Her applicant tracking system calculates the date automatically from the application timestamp and holds the screen until it passes, which removes the judgment call from individual recruiters entirely. If a candidate's specific circumstances raise a question about timing, that is a question for legal rather than for the recruiter under pressure to fill the role.

Does disclosure mean we have to let candidates opt out? It depends on the regime and on what you are running. Under Local Law 144 the notice has to explain how a candidate can request an alternative selection process or accommodation, and under EEOC and ADA rules a candidate must be able to request a reasonable accommodation and an alternative process, particularly where a tool might screen out candidates with disabilities. Beyond those, the honest answer to a candidate is whatever is true in your process. If an alternative exists, explain it and its implications, such as a longer timeline. If it does not, say so and explain why rather than implying a flexibility you cannot deliver.

Will disclosing scare good candidates away? This is the fear that keeps most teams quiet, and Sofia's experience ran the other way. Of roughly 240 applicants for her analyst role, 11 asked questions and 4 requested an alternative review, all of which a recruiter handled without difficulty. What actually cost her applications was the undisclosed version, where a single post from a rejected candidate reached hundreds of people and coincided with a drop in application rates on three open requisitions. Candidates are not alarmed by the existence of a screening tool. They are alarmed by finding out about it from someone else.

What do we say when a candidate asks how they can be sure the tool is not discriminating? Share what you actually have: the audit summary, the disparate-impact testing results, and the human oversight built into the process, including the recruiter's ability to override the tool. What Sofia will not do is offer reassurance in place of evidence. If the testing has not been done, the correct response internally is to pause the disclosure and run the audit first, because a fairness promise you cannot substantiate does more damage than the silence it replaced.