---
title: "What Is Algorithmic Bias in Hiring? Causes & Testing"
description: "Algorithmic bias in hiring is when AI tools treat candidate groups unfairly. Learn where it comes from, where it shows up and how TA teams test for it."
url: "https://www.socialtalent.com/glossary/algorithmic-bias"
---

[Home](https://www.socialtalent.com/) » [Glossary](https://www.socialtalent.com/glossary/) » Algorithmic Bias

# What is Algorithmic Bias?

Algorithmic bias in hiring is when an AI or automated tool produces systematically different results for different groups of candidates for reasons unrelated to the job. It usually comes from training data, proxies or design choices, and it can affect resume screening, matching, ranking and AI interviews.

#### By Lee Flanagan

15th Oct. 2026 | Last Updated: 15th Oct. 2026

## Extended definition

Algorithmic bias is unfairness built into the output of an automated system. In hiring, it means a tool that screens, ranks, matches or scores candidates does so in a way that consistently favors or disadvantages a group, such as women, older applicants or people with disabilities, for reasons that have nothing to do with the job.

Nobody has to intend it, and that’s what makes it hard to argue with. Most hiring models learn from historical data, and historical hiring data is a record of past decisions, biased ones included. The famous example is Amazon’s experimental recruiting engine, which Reuters reported in 2018 had been scrapped after it taught itself to penalize resumes that included the word “women’s.” Its models had been trained on resumes submitted over a 10-year period, most of them from men. The tool wasn’t malicious. It was diligent.

So algorithmic bias isn’t separate from human bias. A model trained on biased human decisions repeats them faster, at greater scale, and with less visibility, because rejected candidates rarely see why.

## Where algorithmic bias comes from and how it shows up

It can enter a hiring tool in three ways:

- **Training data** — If the past hires a model learns from skew toward one group, the model treats that skew as a signal of quality.
- **Proxies** — Deleting protected characteristics doesn’t delete their influence. Zip code, college, graduation year, employment gaps and even hobbies can stand in for race, age, sex or disability.
- **Feature selection and design** — Which inputs the vendor uses, how the model defines a “successful” hire and where the score cutoffs sit all decide who gets through. A model trained to predict who was hired before is trained to repeat the past.

It then shows up wherever a tool makes or shapes a decision: resume screening and ranking, where a score decides who a recruiter ever sees; candidate matching, where systems that learn from past hires rank lookalikes first; and AI interviewers and assessments, where speech recognition, accents and interaction style can move scores. An [AI recruiting agent](https://www.socialtalent.com/glossary/ai-recruiting-agent) that chains several of these steps carries one skewed shortlist into everything that follows.

Testing starts with outcomes, not code. Compare selection rates by group at every stage the tool touches and look for [adverse impact](https://www.socialtalent.com/glossary/adverse-impact). The [four-fifths rule](https://www.socialtalent.com/glossary/four-fifths-rule) is a common first screen: if one group’s selection rate is below 80% of the highest group’s rate, investigate. A formal [AI bias audit](https://www.socialtalent.com/glossary/ai-bias-audit) goes further, with an independent auditor and, in some places, published results.

## Why algorithmic bias matters

Scale is the first reason. A biased interviewer affects the candidates in front of them. A biased screening model affects every applicant it touches, often before a human reads a single application. You can retrain an interviewer over coffee. A model needs someone to notice first.

Accountability is the second. Buying a tool doesn’t hand off responsibility for what it does. The employer still made the hiring decision, even if software did the sorting, and that’s how candidates, regulators and courts tend to see it too. It’s fair to say vendors carry some of this. It’s not safe to assume they carry all of it.

Regulation is the third. Some jurisdictions now require an independent bias audit of automated hiring tools, New York City among them, with public results and notice to candidates. Wherever you hire, expect more of this, not less.

## Common mistakes and misconceptions about algorithmic bias

- **Assuming software is neutral because it isn’t human** — Models learn from human decisions and inherit their patterns, minus the embarrassment that sometimes stops a person.
- **Thinking that removing protected fields fixes it** — Proxies carry the same information into the model by a side door.
- **Treating vendor assurances as the end of the job** — A vendor’s fairness testing reflects their data. Your applicant pool, criteria and cutoffs can produce different results, so track your own selection rates.
- **Testing once, at purchase** — Models, applicant pools and job requirements change. Re-run the numbers on a schedule and after any change to the tool or the role.
- **Letting the score become the decision** — Keep a named person accountable for each rejection, with a reason they could explain to the candidate.
- **Assuming human review cancels it out** — If reviewers only ever see the candidates the tool ranked highest, they inherit its bias and add a signature to it.

Related training

### AI Training for Talent Teams

Practical AI skills for recruiters and hiring managers, taught by working practitioners.

[Explore AI training →](https://www.socialtalent.com/ai-training)

## Frequently asked questions

How can AI agents introduce bias in hiring?

AI agents introduce bias the same way other hiring algorithms do, with more autonomy and fewer people watching. An agent that sources, screens or schedules candidates acts on patterns learned from past data, and proxies like college, location or career gaps can quietly stand in for protected traits. Because agents chain decisions together, one skewed step, such as a shortlist, shapes every step after it. Track selection rates at each stage the agent touches, not just at the end.

What is algorithmic bias in hiring?

Algorithmic bias in hiring means an automated tool treats groups of candidates differently in ways the job doesn't justify. It can come from biased historical data, proxy variables or design decisions, and it shows up in resume screening, ranking, matching and AI interviews. Because these tools run at scale, one biased model can affect thousands of applicants before anyone spots the pattern, which is the part that should worry a TA leader most.

Where does algorithmic bias in hiring tools come from?

There are three main sources. Training data: models learn from past hiring decisions, biased ones included. Proxies: inputs like zip code, graduation year or employment gaps can stand in for race, age, sex or disability. Design choices: which features the vendor includes, how "success" is defined and where the cutoffs sit. Removing protected characteristics from the inputs doesn't remove their influence on the output, which surprises a lot of people the first time they hear it.

How do you test an AI hiring tool for bias?

Start by comparing outcomes rather than reading code. Track selection rates by group at each stage the tool touches and check for adverse impact, using the four-fifths rule as a first screen. Ask the vendor what data trained the model, which inputs it uses and how it was tested. Then re-run the analysis on your own applicant data regularly, and again after any change to the tool or the role. A formal bias audit does the same thing with an independent auditor.

Who is liable when an AI hiring tool is biased?

It depends on where you hire and on the facts, and the law is still catching up with the technology. The safe working assumption is that buying a tool doesn't hand off responsibility for what it does to your candidates. The employer made the hiring decision, even if software did the sorting. So keep your own records of selection rates, ask vendors hard questions in procurement, and talk to employment counsel about your setup. This isn't legal advice.