AI bias in hiring
AI bias in hiring is systematic distortion in an AI-supported selection process that can unfairly favour or disadvantage people. It can arise from training data, design choices, organisational practices or how people use the output.
How unequal treatment enters a selection process
A model learns from the information and objectives its designers provide. If successful past hires are used as the target, the model may learn the effects of an earlier selection process as well as evidence of job performance. People who were never hired may be absent from the training outcomes altogether.
The problem extends beyond historical data. A chosen input can act as a proxy for another characteristic, and a reviewer can give an apparently precise score more weight than the evidence warrants. NIST SP1270, published on 15 March 2022, distinguishes systemic, statistical and computational, and human sources of bias. That framework helps explain why changing a dataset alone cannot address every failure.
What a practical example reveals
Suppose an employer values a particular university because several existing colleagues attended it. An AI tool trained to resemble those hires could favour the university even though the job requires a skill available through many education routes. This is an illustrative mechanism, not a finding about a named product.
Removing candidates’ names would leave that preference intact. Replacing the university signal with relevant work-sample evidence changes what the process measures, although the assessment itself still needs checking.
Evaluate errors as well as advancement rates
Record what the system is intended to predict, which candidates it is being used for and how its output affects decisions. Examine unsuitable recommendations and missed suitable candidates, rather than counting only the people who advanced. An overall accuracy figure can conceal different errors across groups or jobs.
Comparisons need enough relevant data and an appropriate legal basis, especially when sensitive personal data is involved. A difference in outcomes is a reason to investigate; it does not by itself explain the cause. Equally, similar overall advancement rates do not establish that every person was assessed fairly.
Ask a provider what was tested, on which population, with what limitations and which system version. Keep a record of changes and review the process when its role, inputs or model changes. Neither a general fairness claim nor a past audit guarantees the result of a new deployment. The EU AI Act sets legal obligations for relevant systems; it does not impose one universal quarterly hiring-bias audit schedule.