AI in hiring, and the way the EEOC actually thinks about it
Hiring algorithms are now treated like any other employment test. The agencies are not waiting for Congress to catch up.
A Fortune 500 company wrote a $2.275 million check in 2024 to settle a Fair Housing Act discrimination claim involving biased AI used in tenant screening. The EEOC’s first AI-related settlement, in 2023, was $365,000 against a company whose hiring software automatically rejected older applicants. With well over 80% of large employers now using some form of AI in employment decisions, the question is not whether algorithms are reshaping hiring. The question is whether your organization has thought through what the existing employment laws already say about them.
The first thing worth internalizing is that the EEOC has not been waiting for new AI legislation. The agency’s posture, articulated repeatedly by former Commissioner Keith Sonderling, is that AI tools used in hiring are simply selection procedures, and selection procedures have been regulated since 1978 under the Uniform Guidelines on Employee Selection Procedures. If an algorithm screens applicants and that screen disproportionately excludes members of a protected class, the analysis under Title VII is the same one courts have run on paper-and-pencil tests for fifty years. Whether the test is administered by a human or a model, the employer needs to show that the criteria are job-related and consistent with business necessity.
That framing matters because it shapes three patterns of risk that I see most often in practice.
The first is the proxy pattern. An algorithm that has been told not to consider age, race, or sex will frequently rediscover those characteristics by correlation through other features (zip code, length of employment history, name structure, even the type of email domain). The result is disparate impact at scale. A human hiring manager working through a stack of resumes can introduce bias by the dozen. An algorithm can introduce the same bias by the hundred thousand. That is what makes class exposure on AI hiring so much larger than class exposure on traditional employment decisions, and it is why disparate impact analysis on AI outputs is now table stakes.
The second is the vendor pattern. When you procure an AI hiring tool from a vendor, your company bears 100% of the liability for discriminatory decisions the tool makes about your applicants. The vendor’s marketing about “bias-free” algorithms is not a defense. What can shift some of the financial exposure is a properly written contract: indemnification for AI-driven discrimination claims, insurance requirements that match the size of potential class exposure, and contractual rights to the data and metrics you need to actually conduct a bias audit. Many vendor agreements give you neither. A bias analysis you cannot run because you do not own the underlying decision data is not a bias analysis.
The third is the shadow-AI pattern. Employees are using consumer AI tools to screen candidates, draft job descriptions, and summarize interviews, often without the employer knowing. Each of those uses is an employment decision the company is making, regardless of whether it was sanctioned. Each carries the same discrimination risk as a formally adopted tool, with none of the audit trail. A policy that says “do not use unauthorized AI tools in hiring” is necessary but not sufficient; what works is a short, named list of approved tools and a clear escalation path for adding new ones.
The state and local layer adds a second set of obligations on top of the federal floor. New York City’s Local Law 144 requires bias audits for automated employment decision tools. Illinois requires disclosure when AI analyzes video interviews. Maryland prohibits facial recognition in interviews without written consent. California has its own framework moving through implementation. If your company operates across state lines or hires remote workers, you are inside a patchwork of obligations that no single vendor has fully solved for you.
The practical takeaway for a Florida employer is to do three things in the next quarter. Inventory the AI tools actually in use, including the ones employees adopted on their own. Run a disparate impact analysis on the outputs of each tool that touches an employment decision. Pull your vendor contracts and read them for indemnification and data access. Most of the exposure I see does not come from a deliberate choice to use a discriminatory tool. It comes from companies that never asked which tools were running in their hiring funnel.