AI interviews are biased in ways human interviewers never were.
Vendors sold AI screening as the cure for human bias. The 2026 audits show it introduced four new biases humans don't share — and amplified two existing ones.
The pitch was beautiful: remove the human, remove the bias. Three years and a wave of NYC Local Law 144 audits later, the picture is less flattering. AI screening tools don't eliminate bias — they redistribute it, and in some cases manufacture entirely new forms of it that no human interviewer would ever exhibit.
- Accent bias at the acoustic feature layer — humans adjust, models penalize.
- Background visual bias — candidates in poorly-lit home environments score lower on 'professionalism' regardless of answer content.
- Engagement bias against neurodivergent candidates whose eye contact and facial affect don't match neurotypical baselines the model was trained on.
- Run-to-run instability — the same candidate, same answers, scored 0.4 standard deviations apart on two consecutive runs in 38% of cases.
- 'Bias-free, structured, repeatable interviewing.'
- 'Eliminates the halo effect.'
- 'Audited annually for fairness.'
- 'Demographic-blind.'
- Bias measurable across accent, age, neurotype, and lighting.
- Halo effect replaced by 'video aesthetic' effect.
- Audits often run by the vendor on a synthetic dataset, not real production data.
- Models inferred protected characteristics from voice and video with 70%+ accuracy.
- Demand the actual 4/5ths rule disparate impact data from the last 12 months of production use — not the vendor's synthetic audit.
- Run a same-candidate stability test: have 20 employees take the assessment twice. If scores move more than 10%, the instrument isn't measuring what you think.
- Always pair AI screening with a structured human interview before any rejection decision. Never let the model auto-reject.
- Disclose to candidates that AI is in the loop, what it scores, and how to request a human review. (Required in NYC, IL, CO and the EU in 2026.)
Disparate impact theory (Griggs v. Duke Power, 1971) holds that a hiring practice is illegal if it disproportionately excludes a protected class — even with no discriminatory intent. AI interview tools that optimize for 'matches successful past hires' will, mechanically, recreate whatever demographic skew existed in the historical data. The tool's vendor calls this 'pattern matching.' EEOC calls it disparate impact. The naming doesn't change the legal exposure.
Add Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. AI vendors optimize for measurable proxies (interview score, retention) and silently encode all the bias baked into those proxies. The math layer makes it look objective. It isn't. It's the historical bias, laundered.
HireVue removed facial expression analysis from their product in 2021 after academic pressure. The math layer was deemed indefensible. iTutorGroup settled for $365K in 2023 for AI-driven age discrimination. Companies using these tools without bias audits assumed the vendor had handled it. The vendors hadn't. The companies signed the contracts. Under disparate impact law, the buyer is on the hook.
- Demand the vendor's bias-testing methodology and most recent disparate-impact audit results in writing.
- If you're hiring in NYC, confirm NYC LL 144 compliance and public audit summary before signing.
- Run your own 90-day disparate-impact analysis on every AEDT used in your funnel.
- Never let an AI tool make the final hire/reject decision. Always a human in the decision loop.
- Require model documentation (training data, accuracy benchmarks, intended use) in the contract.
- Brief your legal team on every AEDT in use. They will not know unless you tell them.
- Re-audit annually. Bias drifts as the model retrains.