AI Hiring Bias: What Changed in 2025, and What Didn't
RedHub AI Editorialupdated August 13, 20264 min read
In short
The EEOC removed its 2023 AI employment guidance in January 2025, and federal disparate-impact enforcement was later deprioritized by executive order. None of that changed the law: Title VII, the ADA and the ADEA still apply to algorithmic selection, private plaintiffs still bring disparate-impact claims, and city and state rules keep expanding. Enforcement posture and legal exposure are two different things.
Jump to a section6
Editor’s note. An earlier version of this post carried a fabricated customer testimonial and unsubstantiated performance claims about a named screening product, invented settlement figures, and 2025 EEOC requirements that did not exist. The body has been replaced. This post names no vendor’s product and endorses none.
This is general information for employers thinking about automated hiring tools, current as of August 2026. It is not legal advice, does not create an attorney-client relationship, and does not assess whether any particular screening practice is lawful. Employment law varies by federal, state and local jurisdiction and has changed materially more than once in the last two years. Confirm your obligations with qualified employment counsel before adopting or changing a screening process.
What actually changed in 2025
The EEOC published guidance in May 2023 on how Title VII applies to algorithmic selection procedures. In January 2025 it was removed from the agency's website. Executive Order 14281, signed that April, directed federal agencies to deprioritize disparate-impact enforcement, and reporting through 2025 indicated the Commission moved to close pending disparate-impact charges.
If you read only that, you would reasonably conclude the risk went away. It did not, and the reason is worth being precise about.
What did not change
Guidance is the agency's explanation of how it reads a statute. Withdrawing it does not amend the statute. Title VII, the Americans with Disabilities Act and the Age Discrimination in Employment Act all still apply to a hiring process that uses an algorithm, exactly as they applied before the guidance existed.
Three things follow from that. Private plaintiffs can still bring disparate-impact claims in federal court, and they do — Mobley v. Workday, an ADEA collective action over algorithmic screening, was still live into 2026. City and state rules continue to expand independently of federal posture: New York City's Local Law 144 still requires bias audits for automated employment decision tools, and Illinois and Colorado have moved in the same direction. And enforcement posture can reverse with an administration, while the exposure created by a hiring decision you make today runs for as long as the limitations period does.
Federal enforcement posture and legal exposure are two different things. Only the first one moved.
Disparate impact is the part that catches people
Disparate impact is the principle that a selection practice can be unlawful because of how its outcomes fall across protected groups — with no intent to discriminate, and no protected characteristic anywhere in the model.
This is why “we removed gender from the training data” is not a defense. A model that never sees a protected characteristic can still find a proxy for it on its own: a postcode, a gap in employment history, the name of a school, a hobby. The output looks neutral and the pattern is not. Reuters reported in 2018 that Amazon had scrapped an internal recruiting tool after it was found to downgrade résumés containing the word “women's” — the model was never told about gender. Amazon's stated position at the time was that the tool was never used to evaluate candidates.
The test is what your outcomes look like, not what your inputs were. Measuring inputs tells you what you intended. Measuring outcomes tells you what happened, and only the second one is what a claim is about.
Where bias-detection tooling actually fits
Screening tools are a reasonable control. They are not immunity, and any vendor telling you otherwise is describing something that does not exist.
What a detection tool produces is evidence that you looked. That genuinely matters — a documented, repeated audit is materially better than nothing when you have to explain your process. But liability turns on outcomes and on what you did about what the audit found. An audit that surfaced a pattern you then ignored is worse than no audit, because now the record shows you knew.
So the honest framing is narrow: detection is one control inside a compliance program, sitting alongside a documented process, human review of adverse decisions, and a retention policy that means you can still reconstruct a decision two years later. Treat any claim of guaranteed compliance as a reason to ask harder questions, not as reassurance.
Questions worth asking a vendor
Ask what task an accuracy figure describes, on what dataset, and what counts as an error — a percentage with no denominator is not a measurement. Ask for a named customer reference rather than an anonymized case study. Ask what the tool does when it finds something, because a flag with no workflow behind it is a report nobody reads. And ask what it will not do, since a vendor who cannot answer that has not thought carefully about where their product stops.
That last one is the tell. The useful tools in this category are specific about their limits.
A note on the international picture
Two things are commonly misstated. Canada's Artificial Intelligence and Data Act was a proposal in Bill C-27; it died on the order paper when Parliament was prorogued in January 2025 and has not been reintroduced, so it currently requires nothing. And while recruitment and employee-management systems are classed high-risk under Annex III of the EU AI Act, the Digital Omnibus moved those obligations to December 2027 — the transparency and AI-literacy duties were not delayed. Both dates move; check the current position before you plan around either.
Frequently Asked Questions
Can a bias-detection tool protect an employer from discrimination lawsuits?
No. No product can deliver immunity from discrimination claims, and a vendor offering it is describing something that does not exist. Screening tools produce evidence that you looked, which genuinely matters in a defense — but liability turns on outcomes and on what you did about what the audit found. An audit that surfaced a pattern you then ignored is worse than no audit, because the record now shows you knew. Treat detection as one control in a compliance program. This is general information, not legal advice; employment-law questions belong with your own counsel.
Is the EEOC still enforcing AI hiring bias?
Not in the way it was. The EEOC removed its May 2023 AI employment guidance in January 2025, and Executive Order 14281 directed agencies to deprioritize disparate-impact enforcement; reporting through 2025 indicated the Commission moved to close pending disparate-impact charges. Federal enforcement posture on this has genuinely shifted.
Does that mean AI hiring bias is no longer a legal risk?
No — this is the distinction that matters most. Withdrawing guidance did not amend the statutes. Title VII, the ADA and the ADEA still apply to algorithmic selection, private plaintiffs can still bring disparate-impact claims in federal court, and state and city rules continue to expand, including New York City’s bias-audit requirement for automated employment decision tools. Enforcement posture and legal exposure are two different things, and only the first one moved.
What is disparate impact, and why does it matter for AI hiring?
It is the principle that a selection practice can be unlawful because of how its outcomes fall across protected groups, even with no intent to discriminate and no protected characteristic in the model. That is why “we removed gender from the training data” is not a defense: a model can find a proxy on its own — a postcode, an employment gap, a school. Measuring outcomes, not inputs, is the actual test.