Category guideRead at source, 9 September 2026

AI Screening: What It Does, Where It Breaks, and Which Rules Already Bind It

"AI screening" is a label for at least five different things — ranking resumes, scoring an assessment, running a chatbot that asks knockout questions, analysing a recorded video interview, and deciding who sees your job ad. The law does not treat them as one. As of 9 September 2026, the rules that actually bind a US employer are: Title VII and the four-fifths rule in 29 CFR 1607.4, everywhere, since 1978; New York City Local Law 144, enforced since 5 July 2023, which requires an annual bias audit, a published summary and ten business days' notice; two separate Illinois statutes, one covering video interviews since 2020 and one covering all AI in employment since 1 January 2026; and California's FEHA regulations, in force since 1 October 2025, which make your vendor your agent and your screening data a four-year record. Colorado's much-discussed AI Act never took effect — it was repealed and reenacted in May 2026, and the replacement does not apply until 1 January 2027.

Five different products, one label

Vendors sell "AI screening" as one thing. California's regulators, drafting a definition they had to make stick, enumerated what these systems actually do — and the list is the clearest taxonomy anyone has published. An automated-decision system, under 2 CCR § 11008.1, performs tasks such as:

• Using computer-based assessments or tests — "questions, puzzles, games, or other challenges" — to make predictive assessments, measure skills or reaction time, measure "personality trait, aptitude, attitude, and/or cultural fit", or to screen, evaluate, categorize or recommend applicants;
• Directing job advertisements or other recruiting materials to targeted groups;
• Screening resumes for particular terms or patterns;
• Analyzing facial expression, word choice, and/or voice in online interviews;
• Analyzing employee or applicant data acquired from third parties.

Those five behave differently and are regulated differently. A resume ranker and a video analyser are not the same legal object, and a tool that only targets ads sits outside some of these rules entirely. The first useful thing you can do is stop saying "we use AI screening" and write down, in one sentence per tool, what each one actually decides.

The rule that was already there: the four-fifths rule

Before any of the AI statutes, there was Title VII and the Uniform Guidelines on Employee Selection Procedures, published in 1978 and still current. They apply to any "selection procedure", which includes a model that ranks resumes as surely as it includes a written test.

Two clauses in the same section are quoted far less often and matter more. Section 1607.4(C) sets out the "bottom line": if the total selection process shows adverse impact, the individual components should be evaluated one by one. And 1607.4(A) requires users to maintain records that disclose the impact of their selection procedures — with the sting in 1607.4(D), which says that where a user has not maintained that data, the enforcement agencies "may draw an inference of adverse impact" from the failure itself.

Put plainly: if your screening tool has a disparate pass rate and you have never measured it, the absence of the measurement is not neutral. It is the thing that lets an inference be drawn against you.

Where you actually stand, as of 9 September 2026

New York City: the audit, the summary and the ten days

Local Law 144 prohibits using an automated employment decision tool unless three things are true: a bias audit was done within one year of the use, information about that audit is publicly available, and notice was given. DCWP began enforcement on 5 July 2023.

The audit is not a certificate. Per DCWP's published FAQ, it must include "calculations of selection or scoring rates and the impact ratio across sex categories, race/ethnicity categories, and intersectional categories" — that last word is the one that catches vendors, because a tool can pass on sex and on race separately and fail for, say, Black women. The published summary must carry the audit date, the data source, the number of individuals assessed who fell in unknown categories, and the applicant counts, rates and impact ratios for all categories.

The notice goes to New York City resident candidates and employees at least ten business days before the tool is used, and it must state the job qualifications or characteristics the tool will assess and how to request a reasonable accommodation. And note the boundary DCWP draws: the law reaches candidate assessment and screening at any hiring or promotion stage, but not resume scanning for outreach or an invitation to apply. Sourcing is outside it. Screening is inside it.

Illinois: two laws, and people conflate them

The Artificial Intelligence Video Interview Act, 820 ILCS 42, has been in force since 1 January 2020 and is narrow: it applies when an employer asks applicants to record video interviews and uses an AI analysis of those videos for Illinois-based positions. Before asking for the video the employer must notify the applicant that AI may be used, "provide each applicant with information before the interview explaining how the artificial intelligence works and what general types of characteristics it uses to evaluate applicants", and obtain consent. An employer "may not use artificial intelligence to evaluate applicants who have not consented". Videos may be shared only with those whose expertise or technology is necessary to evaluate fitness, and on request must be deleted within 30 days — including every backup copy held by anyone the employer sent them to.

Section 20 adds a duty most employers have never read: an employer that relies solely on an AI analysis of a video interview to decide who gets an in-person interview must collect the race and ethnicity of applicants who were and were not advanced, and of those hired, and report it to the Department of Commerce and Economic Opportunity annually by 31 December, covering the twelve months ending 30 November.

The Human Rights Act amendment is the broad one. Since 1 January 2026, 775 ILCS 5/2-102(L) makes it a civil rights violation for an employer to use artificial intelligence that "has the effect of subjecting employees to discrimination" on a protected basis, or "to use zip codes as a proxy for protected classes", with respect to recruitment, hiring, promotion, discipline, discharge and the terms of employment — and separately, to fail to give notice that AI is being used for those purposes. The statute directs the Department to adopt rules on the circumstances requiring notice, the timing and the means, so check where those rules stand before you draft yours.

California: your vendor is you

California did not write an AI statute. It amended the FEHA regulations, effective 1 October 2025 by order of the Office of Administrative Law, and the effect on an employer is more immediate than a headline statute because it changes who is liable and what you must keep.

• Agents are employers. An "agent" now expressly includes anyone acting for the employer to exercise a function traditionally exercised by the employer — applicant recruitment, screening, hiring — "including when such activities and decisions are conducted in whole or in part through the use of an automated decision system." An agent of an employer is also an employer under the Act. Outsourcing the screen does not outsource the exposure.
• Anti-bias testing is evidence. Section 11009(f) makes it unlawful to use an automated-decision system or selection criteria that discriminates, and then says the relevant evidence includes "evidence, or the lack of evidence, of anti-bias testing or similar proactive efforts — including the quality, efficacy, recency, and scope of such effort, the results of such testing, and the response to the results." Testing and then ignoring what you found is its own exhibit.
• Four years. Section 11013(c) extends record preservation from two years to four, and names "selection criteria, automated-decision system data" in the list of what must be preserved. Automated-decision system data includes data used in or resulting from the tool's application, and data used to develop or customise it for you.

Colorado: the law everyone prepared for, that never arrived

Colorado's Senate Bill 24-205 was the first comprehensive US AI statute and drove two years of compliance planning. Its effective date was pushed to 30 June 2026 by SB 25B-004. Then, on 14 May 2026, the Governor signed Senate Bill 26-189, which repeals and reenacts the whole of part 17 with a new framework built around "automated decision-making technology" and "consequential decisions" — including decisions about employment.

Section 5 of that act is what matters for planning: it takes effect 1 January 2027, and it "applies to consequential decisions made on or after January 1, 2027". Developers must hand deployers technical documentation from that date; deployers must give clear notice at the point of interaction, must provide a plain-language description of the technology's role within 30 days after an adverse outcome, and must keep compliance records for at least three years. Consumers get rights to their data, to correct it, and to request meaningful human review after an adverse outcome. The Attorney General enforces it, with a 60-day notice and cure period for actions brought before 1 January 2030, and there is no new private right of action.

Check what attaches to you

Pick where you hire and what your tools actually do. This applies the sources above as published; it is a reading aid, not legal advice, and it does not cover every state or every tool.

What actually goes wrong

Almost none of the failures we see are exotic model behavior. They are ordinary operational gaps that happen to now have statutes pointed at them.

• Nobody has the pass rates. The single most common answer to "what is your selection rate by group at this stage" is that the data lives in the vendor's system and has never been exported. Under 1607.4 that absence works against you.
• A knockout question is doing the screening, not the model. Availability, "must have own transport", a radius filter, an arbitrary years-of-experience floor. These are selection procedures with impact of their own, and they are usually configured by whoever set the tool up on a Tuesday afternoon.
• The notice is on the careers page, and the tool runs at application. New York City wants ten business days ahead of use. A banner published the same morning does not satisfy a ten-day rule.
• There is no route round the tool. If a candidate asks for an accommodation, someone has to be able to assess them another way. If no such path exists, the notice describing how to request one is describing something imaginary.
• The vendor will not give you the audit. Which is a commercial answer to a legal question. The employer using the tool carries the duty, not the company selling it.

Five questions to ask before you sign

• What exactly does this tool decide, in one sentence — and does it rank, filter, or reject?
• Can you show me selection or scoring rates and impact ratios by sex, by race and ethnicity, and intersectionally, from a bias audit within the last twelve months?
• Can I export the per-candidate pass and fail data, with group data where I collect it, without asking you?
• What is the documented alternative process for a candidate who requests an accommodation?
• Which inputs are location-derived — zip code, radius, commute time — and can I turn each of them off?

The part that is a recruiting problem

Screening tools are usually bought to solve a volume problem: too many applicants, not enough recruiter hours. That is a real problem. But the arithmetic underneath it belongs to the top of the funnel, not the middle.

Across the 891 Boostpoint-managed Meta campaigns in our 2026 Social Job Advertising Benchmark, the median cost per applicant was $13.88, and the volume-weighted blended figure across 1,334 campaign-months was $8.02. Every applicant a screen rejects was bought at something like that price. A tool that removes 60% of your applicants is not saving you 60% of anything — you already paid for them. It is buying back recruiter hours, which is a fair trade, but it is worth pricing honestly rather than calling it efficiency.

Which points at the cheaper fix, and it is not a better model. It is asking the disqualifying question in the ad and in the first exchange, so the people who cannot clear the bar never enter the funnel at all. Our guides to what makes a good job ad and pre-screening questions cover how to do that, and pre-employment assessments covers the scored-test end of the same problem.

Screening software: what the category actually contains

Applicant screening software is any tool that decides which applications a recruiter sees first. The category name covers four quite different things, and most buying mistakes come from confusing them.

• Knockout screening — disqualifier questions answered in the ad or the application form, scored instantly. Cheapest, and the only kind that works before you pay for the click.
• Resume parsing and keyword matching — ranks uploaded resumes against a job description. Largely irrelevant to frontline hiring, where most applicants have no resume to upload.
• Automated candidate screening — scores or ranks completed applications against your criteria and routes the top of the list to a recruiter, usually within minutes rather than days.
• Assessment platforms — add a test or a situational-judgment exercise. These carry validation obligations and adverse-impact exposure that the other three do not.

For hourly hiring the honest comparison is not between vendors, it is between screening software and a recruiter reading forms the next morning. Speed is most of the value: an application that sits unread for three days competes with every other employer the candidate applied to that week.

One buying rule worth holding to. Every applicant your screen rejects was already paid for — a median of $13.88 on social, and up to $74.62 for licensed roles. Software that rejects faster is worth having; software that rejects more is only worth having if the people it removes were genuinely unhireable. That is the difference between screening for disqualifiers and screening for preferences.

Frequently asked questions

This is one rule among about a dozen that attach at different points in a hire. Our hiring compliance hub sets out which obligation bites at which stage, and the four federal deadlines every US employer is on.