← ENGINEERING NOTES
RESPONSIBLE AI · ADVISORY NOTE · 5 MIN READ

Fairness is a test, not a policy

If an automated screening decision moves when only the candidate’s name changes, you have a regulatory exposure, whatever your policy says. The matched-pair tests, abstain rules and audit trail we recommend before any such system goes live.

For: HR and talent leaders, compliance, CTOsRelated engagement: AI Reliability Audit, Transform

The exposure

Draft — to be completed

  • AI screening (resumes, applications, claims) risks bias, opacity and false confidence; regulators and candidates increasingly ask for evidence, not policy.

Where standard controls fall short

Draft — to be completed

PRACTICEWHAT IT HIDESCONSEQUENCE
A written fairness policySays nothing about behaviourNo evidence
Removing obvious sensitive fieldsProxies (address, school, gaps) carry the signalBias persists
Single-pass scoringUnstable scores presented as decisionsInconsistent outcomes
No injection check on documentsText hidden in a resume influences the scoreManipulation
No audit logDecisions cannot be explained after the factRegulatory risk

What we recommend

Draft — to be completed

  1. Injection scrubbing and PII redaction before scoring.
  2. Scores built from weighted, explained components, not a single opaque number.
  3. Repeated scoring with abstain-on-uncertainty when runs disagree.
  4. Matched-pair probes: name swaps and proxy flips must not move the decision.
  5. Group fairness snapshot across the applicant pool.
  6. Compact per-decision audit log with the explanation.

OUR RECOMMENDATION

Run the name-swap test before launch and again every time the model or prompt changes. It takes an afternoon and it is the question a regulator will ask first.

Findings are from EonAI’s reference systems, built and tested on synthetic data. They describe patterns we see across clients’ systems, not a client engagement.