Can you use AI in hiring? What employers need to know

The first hiring job owners want to hand an AI is sorting résumés, because it’s the most tedious hour of the week. It’s also the step with the most legal exposure, because it decides who a person ever gets to see. The rest of hiring has plenty of safer work for an agent, and it’s worth knowing which is which before you buy anything.

Is it legal to use AI to screen job applicants?

Yes. No federal law bans it. The EEOC’s position is that the laws against discrimination based on race, color, religion, sex, national origin, age 40 or over, disability and genetic information apply to AI in employment “just as they apply to other employment practices.” Using software does not change who made the hiring decision.

The EEOC’s one-page summary gives two ways it goes wrong. Discrimination can be intentional, like a résumé screener programmed to reject certain people. It is also illegal when a practice that looks neutral has an unjustifiable disparate impact. One of its examples is video interviewing software that scores an applicant low because a disability changes how they speak.

One thing has changed on the federal side. The EEOC’s longer 2023 technical assistance on adverse impact in AI selection tools no longer loads on eeoc.gov; it returned a 404 when we checked in September 2026. That document explained the law. It was never the law, and the rules it described are still in force.

What is the four-fifths rule, and why does it matter for AI screening?

It is the long-standing federal test for adverse impact. Under the Uniform Guidelines on Employee Selection Procedures, a selection rate for any race, sex or ethnic group below four-fifths (80%) of the rate for the group selected most often is generally treated as evidence of adverse impact. A screening tool’s pass rate is a selection rate.

A worked illustration, with made-up round numbers: if a tool passes 50 of 100 applicants from one group and 30 of 100 from another, the second group’s rate is 60% of the first. That is under 80%, and it is the kind of result that gets a tool pulled. The Guidelines also say smaller gaps can count when they are statistically and practically significant, and larger ones may not when the numbers are small.

What does New York City’s Local Law 144 require?

If you hire or promote in New York City with an automated employment decision tool, three things must be true before you use it: the tool has had a bias audit within one year, a summary of the results is public, and candidates or employees got notice at least 10 business days before use. Enforcement began July 5, 2023.

In practice, ask the vendor for its most recent bias audit summary before you sign. If it doesn’t have one, that answers the question for New York City.

What changed in Illinois in 2026?

From January 1, 2026, the Illinois Human Rights Act makes it a civil rights violation for an employer to use AI that has the effect of discriminating against protected classes in hiring, promotion, discipline, discharge or the terms of employment, or to use zip codes as a proxy for a protected class. Employers must also give notice when they use AI this way.

The law says “has the effect of,” so intent is not required. The Illinois Department of Human Rights is directed to write the rules on when notice is required, how far ahead, and how it is given, so check its current rules before you roll anything out there.

Where does AI help in hiring with the least risk?

In the work around the decision: drafting job posts from your notes, scheduling interviews, answering candidates’ questions about the process, sending status updates, and organizing references for a person to read. None of that decides who advances. The risk rises sharply the moment a tool scores, ranks or rejects people.

This is the same split that runs through every agent decision: an agent drafting and organizing is cheap to get wrong, and an agent deciding about a person is not. If you are weighing an agent against another hire for the recruiting admin itself, the comparison of agent, automation and new hire covers that choice.

How do you use AI in hiring without creating a discrimination problem?

1. Start with the admin work above, where the tool never decides who advances.
2. If a tool scores or ranks applicants, get the vendor’s adverse-impact results by group in writing, plus any bias audit summary.
3. Run the four-fifths check on your own pass-through numbers every quarter. Your applicant pool is not the vendor’s test set.
4. For the first months, have a person review every rejection made at the screening stage.
5. Offer another route for applicants who ask for one, including anyone who needs an accommodation for a disability.
6. Write down which tool and version you used, what it was told to look for, and who reviewed its output.

The last step feels like paperwork until someone asks why they were rejected. Then it is the only thing you have. For the general pattern of where a person has to sign off, see where approval belongs in an agent workflow.

Who should hold off

If you hire in New York City and your vendor cannot produce a bias audit summary, do not use its scoring features there. If you cannot get pass rates by group out of the tool, you cannot run the four-fifths check, so you cannot know whether it is safe. And if your hiring volume is a handful of people a year, the tedium you would save is small next to the exposure.