AI bias in hiring
Also called: algorithmic bias, bias in AI recruiting
Where the bias comes from
Three sources, in approximate order of frequency:
- Training data: the model learns from historical hiring decisions. If historical decisions favored one demographic, the model treats that as the right answer and reproduces it.
- Feature selection: which signals the model is given access to. Excluding “name” doesn’t help if “graduation year” is a proxy for age, or “neighborhood” a proxy for socioeconomic status.
- Feedback loops: a model trained on past hires, where past hires were filtered by past biased decisions, learns the bias compound — every retraining widens the gap.
The Amazon hiring-AI case (2018) is the textbook example: a model trained on a decade of mostly-male hires learned to penalize “women’s” — as in “women’s chess club captain” — and the model was killed.
What audits actually catch
Bias audits are required for high-risk AI systems under the EU AI Act. They typically check:
- Disparate impact: do candidates from protected groups advance at the same rate?
- Equalized odds: is the false-rejection rate the same across demographics?
- Counterfactual fairness: would the same candidate get a different score with a different name or photo?
No audit catches everything. The honest read: audits reduce known biases; emergent biases require ongoing monitoring.
Mitigation that works
- Human review of every advance and every rejection, especially around the threshold.
- Diverse training data, where the dataset reflects the candidate pool you want, not just the one you have.
- Regular re-audits: every quarter, not once at launch.
- Vendor transparency: ask what data trained the model, what fairness metrics were tested, when the last audit happened.
Where Join fits
Join’s AI features are audited per EU AI Act requirements; bias-audit results are publicly summarized at trust.join.com. See the features page.
