AI ethics for HR: the five decisions you can't outsource

60-Second Summary

Ethical AI isn't a vendor checkbox. It's a series of named decisions that HR owns end-to-end.
Adverse impact testing is the floor, not the ceiling. EEOC's 4/5ths rule is necessary, not sufficient.
Human-in-the-loop only works if the human can override and is rewarded for doing so.
Document the model card, the data, the override log. If you can't, don't deploy.

Every vendor pitch ends with 'ethical AI'. Ethics isn't a feature — it's a decision log. When a candidate sues, a regulator inspects, or an employee asks 'why did the model say no', the question lands on HR's desk. These are the five places it lands.

The five decisions HR owns

Sourcing → Screening → Scoring → Surveillance → Separation

1
Sourcing
Are we using AI to find people? If yes — what training data, what exclusion patterns, can we audit who never gets surfaced?
2
Screening
Resume/CV scoring or video interview analysis. Disparate impact test required before go-live and every 6 months.
3
Scoring
Performance / promotion prediction. If the model affects pay or progression, the explanation has to fit on one page for the employee.
4
Surveillance
Productivity scoring, sentiment analysis, keystroke monitors. The bar isn't 'is it legal' — it's 'would we defend this on stage at all-hands?'
5
Separation
Layoff selection, PIP prediction. Never let the model make the decision; only inform it. Document who chose, why, and what the model said.

Tests before go-live

Adverse impact analysis — 4/5ths rule by sex, race, age, disability (where lawful to collect).
Counterfactual test — same candidate with name/gender swapped: does the score change?
Stability test — rerun 100 borderline cases; if outputs drift >10% with no input change, model is unstable.
Explainability — can a non-technical recruiter explain a single decision to a candidate in two sentences?

Human-in-the-loop that's real

HITL is theatre when the human is asked to confirm 200 decisions in an hour. Real HITL means: the human can override, the override is logged, override rates are reviewed monthly, and recruiters aren't penalised for overriding. If overrides are 0%, the human isn't really in the loop.

What regulators will ask for

EU AI Act (high-risk HR systems): conformity assessment, technical documentation, post-market monitoring.
NYC Local Law 144: annual bias audit, public summary, candidate notice.
Illinois AI Video Interview Act: consent + explanation + retention limits.
GDPR Art. 22: meaningful human review of any solely-automated decision with significant effect.

The one-page model card

For every model: purpose, data used, training cutoff, evaluation metrics, known limitations, owner, last audit date, planned re-audit. If a vendor won't give you one, treat that as the answer.

Written by Pawan Joshi.Sources cited inline.

First published 16 Jun 2026See site changelog →