MIT Tech Review: AI develops recruitment biases faster than humans

In plain terms. MIT Technology Review (20/07/2026) reports research indicating that LLMs can develop their OWN recruitment biases - in addition to biases inherited from training data. In other words: cleaning the data is not enough.

Context

Since 2023, the debate on AI bias in HR has oscillated between two camps: "mirror of historical data" (inherited bias, correctable via debiasing) and "systemic amplifier" (bias also produced by the model, not correctable by data alone). The study relayed by MIT Tech Review supports the second camp without invalidating the first. The regulatory framework is complex: NYC Local Law 144 (AEDT audit) effective since 2023, California framework under discussion on AI-assisted HR decisions, EU AI Act - the milestones for "high-risk" HR systems are gradually deployed starting from 2026 (complete transition 2026-2027 depending on the nature of the system).

The Data

• Publication: MIT Technology Review, 20/07/2026 (underlying study to be verified upon review: academic vs industry nature, methodology).
• Reported result: LLMs can develop their own biased associations between demographic attributes and competence judgments, in addition to biases already present in the training data.
• Known precedents: Amazon 2018 (recruiter tool abandoned for gender bias), Meta 2026 complaint (algorithmic management filter).

Analysis

Two points. One: the argument "we train better, it will be fine" is insufficient with a constant structure - part of the bias formation is intrinsic to the maximization objective, not just the data source. Two: American (disparate impact) and European (AI Act art. 14) legal doctrines converge on the obligation of external audit, but neither prescribes a standardized method. Companies are testing with proprietary benchmarks, whose results are incomparable.

Scenarios

• Central (55%): Regulators (California, NYC, EU) set an audit standard within 12-18 months; HR-tech vendors (Workday, iCIMS, Eightfold) comply without the bias being resolved.
• Litigious (25%): A class action results in a settlement of nine or ten figures; deterrent effect on the adoption of LLMs in front-line HR.
• Human pivot (20%): Major recruiters reintegrate humans in the final line, following Meta's post-complaint model.

Implications

For a CHRO: documenting bias audits becomes a contractual obligation, not a checkbox. For an HR-tech investor: auditability replaces precision as a purchase criterion in 2027. For a founder: do not launch an AI scoring in HR without an integrated external audit component.

To Watch

• Publication of the primary study (ArXiv, journal, industry report).
• Follow-ups to the Meta lawsuit (algorithmic management filter).
• Release of the EU AI Act reference on high-risk HR bias tests.