Algorithmic bias is when an AI system treats groups of people unfairly, usually because it learned from biased past decisions or uses data that stands in for a protected trait, such as a zip code. In hiring, it shows up as qualified people from some groups being ranked lower or screened out more often.
AItopic
2 minread
3related terms
In one line
When an AI system treats groups of people unfairly, usually because it learned from biased past decisions or uses data that stands in for a protected trait, such as a zip code.
Bias usually enters in one of three ways. The training data reflects past hiring that favored some groups, so the model learns to copy it. The inputs include proxies, such as zip codes, graduation years, employment gaps or certain hobbies, that track race, age, sex or disability. Or the job criteria themselves are narrower than the work needs, and the tool applies them at scale without anyone questioning them.
Because AI works fast and on large volumes, a small bias in one decision can affect thousands of applicants. In the US, anti discrimination law covers decisions made with software just as it covers decisions made by people, and employers generally cannot shift that responsibility to the vendor. Testing selection rates by group, reviewing the inputs a tool uses and keeping people in the loop are the main defenses.
Why it matters to recruiters
Recruiters and HR teams own the hiring outcome, whatever software they use. Biased tools can shrink your talent pool, hurt your employer brand and lead to discrimination claims. Understanding where bias comes from helps you ask vendors better questions, set fair criteria and notice when results look skewed. Several laws, such as New York City's rules on automated hiring tools, now require formal bias testing.
An example from a recruiting desk
A company's AI screening tool ranks software engineers. HR compares selection rates and finds women pass the screen at a much lower rate than men. Digging in, they see the model rewards long unbroken tenure, which penalizes people who took career breaks. In this example they remove tenure as a factor, retest, and the gap closes to within the range they consider acceptable.
Tips that help
Ask vendors which inputs the model uses and remove any that act as proxies for protected traits.
Compare pass rates by group at every stage where a tool makes or shapes a decision.
Base screening criteria on skills and outcomes the job truly needs.
Keep a person reviewing results, especially rejections.
Mistakes to avoid
Assuming software is neutral because no one programmed it to discriminate.
Testing a tool once at purchase and never again after updates or new roles.
Frequently asked questions
How do you test a hiring tool for algorithmic bias?
Compare selection rates for different groups at each stage the tool affects. Under US federal guidelines, a rate for one group below four fifths of the highest group's rate is generally treated as evidence of adverse impact. Many teams also hire an independent auditor. Retest after any model or criteria change.
Who is liable if an AI hiring tool is biased?
In the US, the employer using the tool is generally responsible for discriminatory outcomes, even if a vendor built it. Vendors may share responsibility under some laws and contracts. Get bias testing details and contract protections from vendors, and keep your own records. This is not legal advice.
Can algorithmic bias be removed completely?
Not completely, but it can be reduced and monitored. Remove proxy inputs, use job related criteria, test outcomes by group regularly and keep humans reviewing decisions. The goal is a process you can explain and defend, with gaps measured, investigated and fixed when they appear.
