Policy
What Do You Owe a Candidate the AI Screened Out?
A candidate rejected by an AI screen is owed three things: notice the tool was used, an accurate account of what it evaluated, and a person who can look again and fix wrong data. No US rule in force requires a candidate-specific reason. Colorado's 2024 act wrote one, was repealed before it started, and returns narrower in January 2027: the tool's role in plain language, correction of wrong data, and meaningful human review [7]. The EU writes a near-equivalent [5]. A rejection you cannot explain is one you cannot defend.
The takeThe disclosure statutes are quietly a procurement rule, and I have yet to see anyone buy against them. Colorado's principal-reasons duty and the EU's Article 86 both assume the deployer holds something worth saying out loud, so a vendor that returns a band and a percentile leaves you owing an explanation the product cannot make. Colorado moved its date once, then repealed and reenacted the act in narrower form before it ever applied, which says how fast this ground turns over. That cost sits in no contract. It arrives one candidate at a time, and it is paid by whoever has to answer them.
Where Olive fits
Open a role and see what the work shows
An explanation is only as good as the record behind it, and a score has no record behind it. Olive returns six findings written by a human reviewer, each attached to a timestamped moment in the session, and grants the candidate the identical report, so the account you give and the document they already hold are the same thing.
Rank your shortlistWhat Do You Actually Owe Them?
Three things, and only one of them is genuinely contested. Tell them an automated tool was used. Tell them accurately what it evaluated: the criteria, not the arithmetic. Give them a person to reach and a way to fix data that was wrong about them. What is contested is whether you owe a reason specific to this candidate, and that turns on where they applied from.
Notice is the part most rules agree on, and it runs before the tool does. In New York City, candidates get at least ten business days' notice before an automated employment decision tool is used (on the careers page, in the posting, or by mail or email), and the notice must include instructions for requesting an alternative selection process or a reasonable accommodation 3. Read that clause closely: the rule requires the instructions, and says plainly that nothing in it requires you to actually provide an alternative process 3. Most employers find that gap the day a candidate uses the instructions.
Federal law adds no general duty to explain a hiring decision, but it does add a duty to keep one. Personnel and employment records, application forms and other records having to do with hiring expressly included, are preserved for a year from the record or the personnel action, whichever falls later, and once a charge is filed every relevant record is preserved until final disposition 4. What a screening tool produced about an applicant is a record having to do with hiring.
So the legal answer and the useful answer are different sizes. The legal answer is a notice and a file. The useful answer is a sentence the candidate can check, which is also the sentence you will need if they take this further. Write it into the policy instead of improvising one per candidate, since what belongs in an AI hiring policy is the same document approached from the other end.
Where Is the Floor Higher Than Notice?
Colorado, and the European Union. Colorado's 2024 AI Act, as signed, required a deployer whose high-risk system was a substantial factor in an adverse consequential decision, and employment is named as one, to give the person a statement of the principal reasons, including the degree and manner in which the system contributed, the type of data it processed, and the source of that data 1.
Two more duties travel with it. The person gets an opportunity to correct incorrect personal data the system processed, and an opportunity to appeal, with human review where technically feasible 1. The notice comes earlier still: before the decision is made, in plain language, in the languages you ordinarily use with consumers, and in a format accessible to people with disabilities 1.
Check the date before you build to it. The signed bill set compliance on and after February 1, 2026 1, a later act moved that to June 30, 2026 2, and before that date arrived the act was repealed and reenacted as a narrower automated-decision-making law taking effect January 1, 2027: a plain-language description of the tool's role within 30 days of an adverse outcome, correction of factually incorrect data, and a right to request meaningful human review 7. Build to the reenacted text, not to the 2024 bill.
The EU right is narrower in wording and wider in reach. A person subject to a decision taken on the basis of the output of a high-risk system, where that decision produces legal effects or similarly significantly affects them, may obtain from the deployer clear and meaningful explanations of the role of the system in the procedure and the main elements of the decision taken 5. Systems used to analyse and filter job applications and to evaluate candidates are listed as high-risk 6. The application timetable for those obligations is staged, so confirm which date attaches to you before planning around it.
None of this argues for three rejection letters. Nobody segments a rejection queue by state, and the strictest rule you touch is the one the process should meet.
Why 'Scored 62' Is Not an Explanation
Because a number reports a position, not a reason. A candidate told they scored 62 learns that some other number was higher, and nothing else. They cannot check it, correct it or answer it, and neither can you when they write back. An explanation names something that happened: what the task asked, what they did, and which part of it fell short.
This is where the correction right stops being theoretical. Colorado's reenacted act gives the candidate a chance to correct factually incorrect data the system used 7, and a reason has to be specific enough to be wrong before anyone can correct it. A score cannot be wrong. "Your submission did not include the reconciliation the case asked for" can be, and occasionally is.
Legibility depends on the role, which is why one explanation template does not survive contact with a whole req list. Tell a product manager they scored 62 and you have said nothing about product management. Tell them the brief was deliberately ambiguous, that the first move was to ask the assistant for a spec rather than to name what was ambiguous, and that no criteria list existed before the deliverable did. Now they can agree, disagree, or point at the part you misread. That is the test worth applying to any tool before it removes people: could the candidate argue with what it produced?
A score survives that test only when the evidence underneath it travels with it. Most scores do not, because most were never designed to be read by the person they describe. Ask what a single candidate's record actually looks like, because the file a vendor owes you is the file you would be quoting from.
Ask What the Tool Produces for One Candidate
Before signing, ask the vendor to show one candidate's record, the artifact itself and not a dashboard view of it. What comes back is what you will have when someone writes to ask why. A score and a band means the rejection letter is already written and says nothing. Evidence with the criteria attached means you have a reason you can put your name to.
Three follow-ups belong in the same conversation.
- Can you export it, and does it outlive the contract? The preservation duty is yours, and it runs a year past the record and past final disposition of any charge 4. A dashboard you lose in renewal month is not a record.
- Is the tool a substantial factor in the decision? Colorado's 2024 act attached its duties at exactly that threshold 1, and the EU's attach to systems that filter applications and evaluate candidates 6. Something that ranks and cuts is deciding, whatever the org chart says.
- What does it produce for someone screened out in the first pass? Vendors demo the finalist view. The letters you have to write are mostly for the people the tool removed before anyone read them.
A bias audit answers none of this. The published New York City summary carries the audit date, the source and explanation of the data, the number of applicants, the selection or scoring rates and the impact ratios for all categories 3. That is a statement about a population, and the candidate is asking about themselves. Publish it because the rule says so, then answer the individual question in a separate sentence. The sorting between the two runs through the questions worth putting to a screening vendor.
What to Say When You Can't Explain It
Say exactly that, and say what happens next. A candidate told the truth (an automated screen was used, here is what it evaluated, here is who will read the application again) is in a better position than one handed a reason nobody can support. Do not reconstruct a reason after the fact. An invented explanation is the version that turns a rejection into a claim.
Four moves, in order.
1. Preserve everything about that candidate now. The tool's output, the configuration in force that day, the threshold it ran at, and the human step if there was one. Once a charge is filed the duty covers every record relevant to it until final disposition 4, and vendor deletion defaults are usually shorter than that. 2. Route the candidate to a person who can reverse it. An appeal that reaches a queue is not human review. Colorado's reenacted act names a right to request meaningful human review 7; the useful version is a named reviewer with authority to change the outcome. 3. Fix the notice for the next cohort, not the letter for this one. A rejection you could not explain means the notice you published described a process you were not running. 4. Move the tool off the removal step. A screen you cannot account for can still order a queue for a human to read. It should not be the thing that ends an application, because the ending is the act you have to justify.
One quieter option closes most of these conversations, and it is worth pricing before the other four: give the candidate the same record you used. That costs nothing when the record is legible and says only what you would defend in a room, and it carries its own exposure, which is why what candidate-facing reports change legally is worth reading before committing to it.
Common questions
Do you have to tell a candidate that an AI screened them?
Where a notice rule reaches you, yes, and before the tool runs rather than after. New York City requires at least ten business days' notice on the careers page, in the posting, or by mail or email, with instructions for requesting an alternative selection process or an accommodation 3. Colorado's reenacted act requires deployer notice at the point of interaction once it takes effect on January 1, 2027 7. There is no general federal notice duty, which is why one notice written to the strictest rule you touch beats four written to four.
Can you refuse to give a specific reason for the rejection?
In most US jurisdictions, yes. There is no general duty to explain a hiring decision. Colorado becomes the exception on January 1, 2027, when its reenacted act requires a plain-language description of the tool's role within 30 days of an adverse outcome, with correction of incorrect data and a route to meaningful human review 7. The practical answer differs from the legal one. A refusal is cheap once and expensive at volume, because it is also the answer you will give an investigator, and it reads the same either way.
Does a bias audit answer the candidate's question?
No. An audit is a statement about a population. The New York City summary carries the audit date, the source and explanation of the data, the number of applicants, the selection or scoring rates, and the impact ratios for all categories 3. None of that says anything about the person writing to you, and quoting it back reads as an evasion because it is one. Publish the summary because the rule requires it, and answer the individual question somewhere else.
What records do you have to keep after an automated rejection?
Personnel and employment records, application forms and other hiring records expressly included, for one year from the record or the personnel action, whichever falls later. Once a charge is filed, every record relevant to it is preserved until final disposition of the charge or any lawsuit based on it 4. That covers what the screening tool generated about the applicant. Vendor deletion defaults are often shorter, so a legal hold and an export belong in the contract rather than in a support ticket.
Should the candidate get the assessment output itself?
Often it is the cheapest option. When the record is legible and says nothing you would not defend, handing it over ends the conversation faster than a summary of it does. Read it first: output written for an internal audience, with shorthand or a bare number in it, lands very differently on the person it describes. Where New York City applies, a candidate can already ask in writing for the type of data collected, its source and the retention policy, and must get it within 30 days 3.
What if the vendor can't tell you why either?
Then you cannot explain the rejection, and that is a procurement finding rather than a support ticket. Ask for the single-candidate record in writing and keep the answer with the purchase file. A tool that produces only a score can still order a queue for a human to work through; it cannot be the step that removes people, because removal is the act you have to account for when someone asks.
References
- 1. Senate Bill 24-205: Concerning Consumer Protections in Interactions with Artificial Intelligence Systems content.leg.colorado.gov Section 6-1-1703(4): notice before a consequential decision, and for an adverse decision a statement of the principal reasons including the degree and manner the system contributed, the type of data processed and its source, an opportunity to correct incorrect personal data, and an appeal allowing human review where technically feasible. Employment is a named consequential decision, and the signed text set the date on and after February 1, 2026.
- 2. Senate Bill 25B-004: Artificial Intelligence Consumer Protections Delay leg.colorado.gov Extends the effective date of the SB 24-205 requirements to June 30, 2026.
- 3. Notice of Adoption of Final Rule: Automated Employment Decision Tools (6 RCNY 5-300 to 5-304) rules.cityofnewyork.us Section 5-304: at least ten business days' notice by website, posting, mail or email, including instructions for requesting an alternative selection process or accommodation, with the express statement that nothing in the subchapter requires providing one; data type, source and retention policy on written request within 30 days. Section 5-303: the published summary of results with applicant counts, selection or scoring rates and impact ratios.
- 4. 29 CFR 1602.14 - Preservation of records made or kept ecfr.gov One year from the making of the record or the personnel action, whichever is later, covering application forms and other records having to do with hiring; once a charge is filed, all relevant personnel records until final disposition of the charge or any lawsuit based on it.
- 5. Article 86: Right to Explanation of Individual Decision-Making artificialintelligenceact.eu An affected person subject to a decision taken by the deployer on the basis of output from an Annex III high-risk system, which produces legal effects or similarly significantly affects them adversely, may obtain from the deployer clear and meaningful explanations of the role of the system in the decision-making procedure and the main elements of the decision taken.
- 6. Annex III: High-Risk AI Systems Referred to in Article 6(2) artificialintelligenceact.eu Point 4(a) lists AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates.
- 7. SB26-189 Automated Decision-Making Technology - Bill Summary (as enacted) leg.colorado.gov Signed May 14, 2026: repeals and reenacts SB 24-205 as a narrower automated-decision-making regime taking effect January 1, 2027, with point-of-interaction notice, a plain-language description of a covered tool's role within 30 days of an adverse outcome, correction of factually incorrect data, and a right to request meaningful human review.
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.