AI Does Not Just Inherit Hiring Bias, It Invents Its Own

The premise MIT Technology Review opens with is uncomfortable and fairly literal: the next time you apply for a job, a language model will probably read your résumé before any human does. What the article, published on 20 July, adds is not that observation but the nuance in the research it covers. We already knew models absorb human biases from their training data, and there are now signs that they also develop biases of their own, meaning systematic preferences that were not in the source data.

That distinction is not academic. It changes the mitigation strategy entirely.

Inheriting a bias and manufacturing one are different problems

If the bias comes from the corpus, the answer involves curating data, reweighting examples and auditing the hiring history the system was trained or tuned on. It is a hard problem but a known one, and there is a decade of literature on it in classical machine learning applied to human resources.

If the bias emerges from the model itself, those tools fall short. You cannot clean out of the data something that was never in the data. It forces you to evaluate the system output as a black box and to do it continuously, because a change of model version, of temperature, or even of prompt wording can shift behaviour without anyone touching the pipeline.

Résumé screening is the worst possible place for this

Candidate screening has every condition needed for a bias to go unnoticed. Volume is high, the decision is a rejection (nobody appeals a no that never arrives), there is no observable counterfactual (you do not know how the filtered out candidate would have performed) and the signal is weak and heavily correlated with protected variables: names, universities, postcodes, gaps in employment history.

On top of that there is a pattern we see in real projects: the model is almost never alone. It sits inside a flow with scores, thresholds and automations, and the number it returns ends up being treated as objective data because it comes from an API rather than from a committee.

The regulatory calendar is no longer far off

For any team operating in Europe, this arrives at a bad or a good moment depending on how you look at it. The obligations of the European AI Act for high risk systems, a category that explicitly covers the use of AI in employment and recruitment, start to apply on 2 August 2026, two weeks from now. That means risk management, technical documentation, logging, effective human oversight and data quality requirements.

In the United States, New York's Local Law 144 has required annual bias audits and candidate notification for automated employment decision tools since 2023. The direction of travel is the same on both sides: the burden of proof falls on whoever deploys the system, not on whoever is subject to it.

What we would do

Three concrete measures, none of them exotic. First, do not let the model issue the verdict: have it extract and structure verifiable information from the résumé and let an explicit, auditable rule apply the filter. Second, build a paired résumé evaluation, identical except for the variable you want to control, and run it on every model or prompt change as if it were a regression test. Third, keep the full trace of every decision, including the exact model version, because without that there is no possible audit and no way to defend yourself against a complaint.

Our measured reading is this: the underlying problem is not that AI is worse than people at hiring, a comparison that is hard to measure well anyway. The problem is that it scales the decision to thousands of applications with a consistency that turns a small bias into company policy. If you are going to automate screening, the serious work is not in picking the model, it is in building the evaluation that watches it.