AI bias in hiring: the myths, the audits, and what to ask a vendor

Bias in hiring did not start with AI. People have always favoured names, schools and accents that feel familiar. What AI changes is scale and speed. A biased person affects the candidates they meet. A biased model affects every candidate it scores, every day, without anyone noticing.

That cuts both ways. A well-tested system applied consistently can be fairer than an overloaded recruiter skimming CVs at the end of the day. This article separates what is true from what is comfortable, explains what an audit can tell you, and gives you the questions to put to any vendor.

What the evidence actually shows

Two findings are worth knowing.

Models learn the past. In 2018 Reuters reported that Amazon had abandoned an experimental recruiting model after it taught itself to prefer male candidates. It had learned from a decade of mostly male CVs and penalised the word "women's". Engineers removed the specific terms but could not guarantee other biases would not appear.

Modern language models are not immune. In 2024 University of Washington researchers tested three language models on more than 550 real CVs, changing only the names. The models preferred white-associated names 85% of the time and female-associated names 11% of the time, and never preferred Black male-associated names over white male-associated names.

Neither study proves that every AI hiring system is biased. Both show that bias is the default unless someone tests for it and fixes it.

Five myths that mislead hiring teams

Myth 1: "AI is objective because it is maths"

A model is objective about the patterns in its training data. If the data reflects past hiring decisions, it reflects past preferences too. Maths does not remove bias; it makes it consistent.

Myth 2: "Remove names and photos and you are safe"

Helpful, but not enough. The University of Washington researchers noted that models can infer identity from education, location and word choice. Postcode, graduation year and gaps in work history can all act as proxies for protected characteristics. Illinois now explicitly bans using zip codes as a proxy in employment decisions.

Myth 3: "A person reviews the output, so we are covered"

This is the most dangerous myth. In a 2025 follow-up study with 528 participants, the same research group found that when AI recommendations were heavily biased, people followed them about 90% of the time. Review only protects candidates when the reviewer has the time, information and authority to disagree, and actually does.

Myth 4: "If the vendor's model is biased, that is the vendor's problem"

Not in law. California's regulations, in force since October 2025, make clear that employers can be liable for discrimination by an automated decision system even when a person makes the final decision. Vendors are exposed too: in *Mobley v. Workday*, a federal court in May 2025 allowed a nationwide age-discrimination collective action to proceed against an AI screening provider. Liability is shared, not transferred.

Myth 5: "We have a bias audit, so we are fair"

An audit is a snapshot, not a certificate. See below.

What a bias audit can and cannot tell you

Most audits calculate selection rates for different groups and compare them as an impact ratio. The common benchmark is the US federal four-fifths guideline: a group selected at less than 80% of the rate of the most-selected group is generally regarded as evidence of adverse impact. The same guideline notes that smaller differences can still count.

New York City's Local Law 144 requires these audits, by an independent auditor, for automated tools used on candidates in the city, and requires the summary to be published.

What an audit can tell you:

• whether outcomes differed between groups for the data tested;
• whether the gap crosses a commonly used threshold.

What it cannot tell you:

• whether the result holds for your roles, your candidates and your settings;
• anything about characteristics it did not measure. NYC audits cover sex and race or ethnicity, not age or disability;
• whether the audit itself was done well. In December 2025 the New York State Comptroller reviewed published audits and found at least 17 potential issues where the city regulator had found one.

The UK regulator's advice is practical: test regularly, and the ICO even suggests considering monthly bias reviews.

Twelve questions to ask any vendor

Ask in writing, and keep the answers.

1. What exactly does the system decide, rank or recommend, and at which stage?
2. What data was it trained on, and how do you check that data for historic bias?
3. Which characteristics do you test for adverse impact? Does that include age and disability?
4. Can we see your most recent independent audit, the method, and who performed it?
5. Can you run the analysis on our candidate data, for our roles?
6. Which inputs could act as proxies, such as postcode, school or gaps in employment, and how are they handled?
7. Can the hiring team see why a candidate was ranked where they were?
8. Who can change scoring criteria, and is every change logged with a reason?
9. How are candidates told that AI is involved, and can they ask for an alternative or a review by a person?
10. Does the system infer emotions or personality from faces or voices? (In the EU, emotion recognition in the workplace is prohibited.)
11. How long are inputs, outputs and logs kept, and can we export them?
12. What happens contractually if a bias problem is found?

A vendor who answers these clearly is a partner. A vendor who calls them "proprietary" is a risk.

How MetaDay handles this

At MetaDay, our AI recruiter runs the hiring work, including sourcing, outreach, screening calls and interviews, inside boundaries the company approves. Bias controls are built into how a run works, not added afterwards.

• Criteria changes need a reason. When a hiring manager changes the scoring criteria mid-run, the system asks why before applying it. That makes quiet drift visible, and leaves a record you can review.
• Candidates know it is AI. Every candidate is told they are speaking with an AI and can choose to speak with the team instead.
• People decide. The run ends with a ranked top three. The hiring team holds the final interview and makes the decision.

FAQ

Can AI hiring systems be biased?

Yes. Research shows language models can favour some groups by name alone. Bias is the default unless the system is tested on real outcomes and corrected.

What is the four-fifths rule in hiring?

A US federal guideline: if one group's selection rate is below 80% of the highest group's rate, that is generally treated as evidence of adverse impact.

Who is liable if an AI hiring system discriminates?

Usually the employer, and sometimes the vendor as well. Courts and regulators treat decisions made with AI as the employer's own decisions.

Fair hiring with AI is not a feature you buy. It is a habit of asking, testing and recording, every run.