EEOC Guidance on AI in Employment Decisions, Explained

The EEOC's AI hiring guidance no longer exists on eeoc.gov. Type the old URL and you get a 404. That fact alone has convinced more than one HR team that algorithmic hiring tools moved into a legal gray zone sometime in early 2025 — that whatever rules used to apply have quietly lapsed along with the web page. They haven't. What disappeared was the agency's own explanation of how it reads Title VII and the Americans with Disabilities Act. Title VII and the ADA are statutes, passed by Congress, and neither one was touched.

Getting this right matters because the EEOC issued two separate technical-assistance documents on this topic, three years apart, aimed at two different statutes — and both were pulled from the agency's website on the same day in January 2025. Knowing what they said is still the clearest available roadmap to how the EEOC reads existing law, even with the hosting page gone.

The ADA guidance: three ways an algorithm can violate disability law

On May 12, 2022, the EEOC published "The Americans with Disabilities Act and the Use of Software, Algorithms, and Artificial Intelligence to Assess Job Applicants and Employees." It described three distinct ways an employment-screening algorithm can create ADA exposure, and they're worth separating because each one requires a different fix.

The first is a failure to accommodate: if an assessment measures something a disability affects — reaction time on a timed test, or verbal fluency on a voice-scored interview — and the employer doesn't offer an alternative way to be evaluated, that's a straightforward accommodation failure, not a novel AI problem dressed up in new language. The second is "screen out": a tool intentionally or unintentionally filters out applicants because of a trait connected to disability, even without ever asking about disability directly. The EEOC's own examples included resume tools that penalize employment gaps (which correlate with medical leave and disability-related absences) and gamified assessments that score reaction time, memory, or specific physical movements. The third is treating the tool itself as a disability-related inquiry or medical exam — something the ADA restricts before a conditional job offer is made. A "personality" or "cultural fit" assessment that ends up measuring traits tied to mental health conditions can cross this line even when nobody wrote it to do that on purpose.

The Title VII guidance: the four-fifths rule is a rule of thumb, not a safe harbor

A year later, on May 18, 2023, the EEOC published "Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII." Written as a Q&A, it walked through how the decades-old "four-fifths rule" from the 1978 Uniform Guidelines on Employee Selection Procedures applies when the selection tool is a piece of software rather than a paper test.

The rule itself is simple: if one group's selection rate is less than 80% of another group's, that ratio is treated as a rule-of-thumb signal of adverse impact. The part employers tend to skip is the EEOC's own caveat directly beneath it — a smaller gap than four-fifths can still count as adverse impact, particularly when a tool is screening a large applicant pool or when the employer's process discouraged some applicants from applying at all. In other words, clearing 80% doesn't clear you. It's a screening threshold for further analysis, not a finish line a hiring tool can be tuned to just barely pass.

Using a vendor's tool doesn't transfer the liability

The line in the 2023 guidance most compliance teams underweight is the one about vendors. The EEOC stated plainly that an employer can be held liable for a selection procedure's disparate impact even when a third-party vendor built and administered the tool. The guidance's suggested practice was to ask a vendor directly what adverse-impact testing it has run — and it went further, noting that if a vendor gets that assessment wrong, the liability doesn't move to the vendor. It stays with the employer that adopted the tool.

That framing tracks how employment discrimination law has always worked: an employer can't outsource its Title VII or ADA obligations to whichever company sold it the software, any more than it could outsource them to whichever consultant wrote its old paper-based aptitude test.

How this plays out: a resume tool that quietly filters out older applicants

Picture a 400-person logistics company, Harrow Freight Systems, that licenses a resume-ranking tool from an HR-tech vendor to triage applications for warehouse-operations roles. Nobody at Harrow configured the tool to consider age. But the model was trained on Harrow's own historical hiring data, and it learned to score down resumes with long, single-employer tenures and resumes with multi-year gaps — patterns that, in Harrow's applicant pool, correlate strongly with applicants over 50, some of whom have gaps tied to documented medical leave.

Run the two EEOC theories against those facts. Under the ADA "screen out" theory, a gap-penalizing model disproportionately screens out applicants whose gap reflects a disability-related leave, even though nobody told the tool to consider disability. Under the Title VII four-fifths analysis, if Harrow pulled its own selection-rate data and applicants over 50 were advanced to interview at, say, 55% of the rate of younger applicants, that's well below the 80% threshold and warrants a real adverse-impact investigation — not a call to the vendor for reassurance. This is not a hypothetical fact pattern dressed up for a compliance memo. It tracks closely with what the EEOC actually alleged against iTutorGroup: software that auto-rejected applicants past a specific age cutoff, discovered only because one rejected applicant reapplied with a different birthdate and was suddenly offered an interview.

The guidance pages are gone. Title VII and the ADA aren't.

Here's the part that actually changed. On January 23, 2025, President Trump signed Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," directing federal agencies to review and roll back AI-related policies issued under the prior administration. Four days later, on January 27, 2025, the EEOC removed both the 2022 ADA guidance and the 2023 Title VII guidance from its website. Neither page resolves today; both return a 404.

What didn't happen: Congress didn't amend Title VII or the ADA, and the EEOC didn't issue new guidance saying algorithmic discrimination is now acceptable. A technical-assistance document is the agency's stated interpretation of existing law for the public's benefit — pulling it down removes the agency's explanation, not the statute it was explaining. The EEOC's own Strategic Enforcement Plan for fiscal years 2024–2028, which was not withdrawn, explicitly names "technology-related employment discrimination" — including the use of algorithmic decision-making and machine learning tools in hiring — as a stated enforcement priority. And the iTutorGroup settlement, which predates the guidance withdrawal by roughly a year and a half, shows the agency was already willing to bring and settle a real case over exactly this kind of automated screening, using authority that has nothing to do with either withdrawn document.

What this means for your hiring pipeline

Treat the two withdrawn documents as a reliable statement of the EEOC's legal theories, even though you'll need to cite them by title and date rather than a working link — that's exactly how this article cites them. Before signing with any hiring-tool vendor, get its adverse-impact testing methodology and results in writing, not a verbal assurance that the model is "bias-free." Run your own periodic four-fifths check against real outcomes in your applicant pool; a vendor's aggregate testing across its whole customer base won't necessarily reflect how the tool performs against your specific candidate population. And if your organization has been treating "we don't have an AI-specific hiring policy yet" as a reason to defer this work, recognize that Title VII and the ADA don't distinguish between a biased algorithm and a biased human reviewer — the absence of an AI-specific policy has never been a defense to either one, guidance page or no guidance page.

For the broader federal picture beyond the EEOC — including the FTC's parallel approach of applying old law to new AI claims without waiting for AI-specific legislation — see our overview of US federal AI regulation and the deeper look at the FTC's own enforcement theory. The same guidance-comes-down-but-the-statute-doesn't pattern played out again in 2025 at the CFPB, which withdrew its own circulars on AI-driven credit decisions and adverse action notices while ECOA's underlying disclosure requirement stayed exactly as binding as before. And for the documentation practice that actually makes a vendor-testing defense credible if the EEOC ever asks for it, the NIST AI RMF's Govern function is the closest thing to a template for what that record should look like.