Definition · HR technology and AI in hiring
AI Bias in Hiring
Definition
AI bias in hiring is when an AI or algorithmic tool used in recruiting produces systematically worse outcomes for people based on a protected characteristic such as race, sex, age or disability. It usually comes from biased training data, proxy variables or flawed design rather than intent, and employers remain legally responsible for it.
Also called: Algorithmic bias, AI hiring discrimination, Bias in AI recruiting, Algorithmic discrimination
Formula
Impact ratio = Selection rate of a group ÷ Selection rate of the most-selected group
- Selection rate
- Number of candidates in a group who pass a step (advance, are scored above the cut, are hired) divided by the number in that group who went through the step.
- Most-selected group
- The group with the highest selection rate for that step, used as the comparison point.
- Four-fifths rule
- Under the federal Uniform Guidelines, an impact ratio below 0.80 is generally treated as evidence of adverse impact that needs a closer look.
Worked example
An AI screen reviews 1,000 applicants: 600 men and 400 women. It advances 300 men (selection rate 50%) and 140 women (35%). Impact ratio for women = 35% ÷ 50% = 0.70. That is below 0.80, so the screen shows potential adverse impact against women and needs investigation before further use.
How bias gets into AI hiring tools
AI tools rarely discriminate because someone told them to. Bias usually enters in one of these ways:
- Historical training data. A model trained on who was hired or rated highly in the past learns past preferences. If a company historically promoted mostly men into supervisor roles, a model predicting 'supervisor potential' learns features associated with men.
- Proxy variables. Removing race or sex from the data does not remove bias. Zip code, college attended, gaps in employment, certain hobbies, graduation year and even word choices correlate with protected traits.
- Measurement problems. Speech recognition that is less accurate for some accents, or video analysis that performs worse on darker skin, produces worse scores for reasons that have nothing to do with ability.
- Bad target definitions. Optimizing for 'similar to our top performers' copies whatever made those people look alike, including demographics.
- Disability effects. Timed games, voice analysis and video scoring can penalize candidates with disabilities unless accommodations are built in.
The practical lesson: you cannot judge a tool by its inputs alone. You have to test its outputs.
How to test for AI bias
The standard first check is adverse impact analysis using the four-fifths rule, shown in the formula above. Run it for every step where the tool makes or influences a decision, by sex, race/ethnicity and, where you have data, age band (40 and over vs under 40). NYC Local Law 144 also requires intersectional categories, such as Hispanic women.
A ratio under 0.80 is a signal, not a verdict. With small numbers, ratios swing wildly, so pair them with a statistical significance test. A ratio above 0.80 does not guarantee fairness either if a large sample shows a statistically significant gap.
Good practice also includes:
- An independent bias audit before launch and at least yearly.
- Checking whether the scores actually predict job outcomes. A tool that is fair but useless is still a bad tool.
- Reviewing what the model weights most, using explainable AI methods where available.
- Monitoring monthly after launch, because applicant pools and models change.
What to do when you find bias
First, stop using the affected step for decisions until you understand it. Then work through the causes: is one feature or question driving the gap? Is a cut score set too high? Is the tool scoring something unrelated to the job? Common fixes include removing or reweighting the problem feature, lowering or banding a cut score, replacing a weak measure (such as vocal tone) with a job-related one (such as answer content against a rubric), or adding human review at that step.
Under the Uniform Guidelines, a step with adverse impact can still be lawful if it is job-related and consistent with business necessity, but you also have to consider whether a less discriminatory alternative would serve the same purpose. Document what you found and what you changed.
Legal framework
In the US, AI tools are covered by the same laws as any other selection procedure:
- Title VII bans both intentional discrimination and disparate impact based on race, color, religion, sex and national origin.
- The ADEA protects applicants 40 and over; the ADA requires reasonable accommodation and limits screening out people with disabilities.
- NYC Local Law 144 requires a yearly independent bias audit of automated employment decision tools, a public summary of results and advance notice to candidates.
- Illinois amended its Human Rights Act, effective January 1, 2026, to prohibit employers from using AI that has a discriminatory effect and to require notice when AI is used.
- Colorado's AI Act was repealed and replaced in May 2026 by SB 26-189. From January 1, 2027, employers whose automated tools materially influence hiring decisions must give notice, explain adverse decisions within 30 days, and offer data correction and human review.
Outside the US, the EU AI Act treats recruitment AI as high-risk, and GDPR limits decisions based solely on automated processing. Vendors are increasingly being named in lawsuits alongside employers, but using a vendor's tool does not move responsibility away from the employer. Laws differ by jurisdiction; get legal advice for your situation.
Common mistakes
- Assuming a tool is fair because it does not 'see' race or sex.
- Accepting a vendor's audit of its whole customer base instead of testing on your own applicants.
- Testing only at launch.
- Letting AI reject candidates with no human review and no way to request an alternative process.
- Not recording demographic data (voluntarily provided and stored separately), which leaves you unable to test at all.
Frequently asked questions
What is algorithmic bias in hiring?
Algorithmic bias is when a software tool used in hiring produces unfair differences in outcomes between groups, often because it learned from biased historical data or relies on variables that act as proxies for protected traits.
Can AI recruiting tools be less biased than humans?
They can be more consistent, because they apply the same criteria every time, and they can be tested in ways human judgment rarely is. But they can also scale a bias across thousands of decisions. The answer depends on design, testing and monitoring.
Who is liable for AI hiring discrimination?
The employer is liable under US anti-discrimination law even when the tool comes from a vendor. Courts have also allowed some claims against vendors to proceed, so both can face risk.
How often should we audit AI hiring tools for bias?
At least once a year, which is what NYC Local Law 144 requires for covered tools, plus ongoing monitoring of selection rates. Re-test whenever the model, the job or the applicant pool changes significantly.
What is the four-fifths rule in AI hiring?
It is a rule of thumb from the federal Uniform Guidelines: if one group's selection rate is less than 80% of the highest group's rate, that is generally treated as evidence of adverse impact that needs investigation.