AI Candidate Scoring Looks Impartial, but That Doesn't Make It Fair

One hiring expert describes the situation this way: companies are "drinking from a fire hose" of applications. Popular openings receive up to 1000 applications in a single weekend. No human can read that volume, so an algorithm is increasingly the first reader of a résumé. A study reported by TechXplore raises an uncomfortable question: AI candidate scores can look impartial without being fair.

The available excerpt of the publication does not disclose the methodology, so what follows covers the logic of the problem and the context in which it arose.

Impartiality and fairness are measured differently

In engineering terms, impartiality almost always means statistics. The system scores résumés with different names the same way, and the shares of candidates who pass the screen don't diverge too much between groups. This can be tested, calculated and dropped into a report for auditors. Fairness is broader. It asks whether people were selected on the right criteria, whether those criteria relate to the actual job, and whether a rejected candidate can understand the reason and challenge the decision.

A model can easily pass the first test and fail the second. If it penalizes a gap in a career, the "wrong" university or a home district, even numbers by gender and age guarantee nothing. These features act as proxies: formally neutral, they reproduce old biases in practice. A telling precedent is Amazon's internal recruiting tool. According to Reuters, it downgraded résumés containing the word "women's," and the company abandoned it in 2018.

What is already measured and regulated

Regulators still mostly demand exactly this kind of statistics, which explains why an "impartial" verdict is so easy to obtain.

• Load: up to 1000 applications for a popular opening over a weekend.
• The classic test: the four-fifths rule (EEOC, 1978). The selection rate for any group must not fall below 80% of the rate for the best-performing group.
• New York: Local Law 144 has required an independent bias audit of automated hiring tools since 2023.
• EU: the AI Act classifies AI systems used for hiring as high-risk.
• Courts: the class action Mobley v. Workday, over possible age discrimination, was granted class status in 2025.

All of these mechanisms look at the output. They are worse at catching the case where the result is even but the path to it is questionable.

How to deal with it

Those who buy or build such systems should extend audits beyond the final percentages. Ask the vendor for the list of features the model uses to make decisions, and check each one against the requirements of the job. Counterfactual tests help: the same résumé with a changed name, graduation year or address. Rejections, especially mass automated ones, should pass through a live recruiter at least on a sample basis. And decision logs are needed so that a candidate can get a clear explanation.

For job seekers, the takeaway is pragmatic. A machine reads the résumé first, so the wording should match the text of the job posting, and the document's structure should be simple and readable for a parser. But responsibility lies with the employer. If even statistics are the only proof of fairness, the system is simply good at looking fair.