Search the Atlas

Search risks, controls, and glossary terms

HighDiscriminatory OutputRealized

Discriminatory output

Fairness & Bias

Description

Generated guidance, explanations, or decisions vary by protected attribute, producing unfair or unlawful treatment.

Example scenario

Identical loan queries differing only by name/gender yield systematically less favourable language.

Real-world evidenceRealized

The New York Department of Financial Services opened a formal investigation into Apple Card / Goldman Sachs in 2019 after documented gender-disparate credit limit decisions. Amazon scrapped its AI recruiting tool in 2018 after internal audits confirmed it systematically downgraded resumes from women. Both are confirmed production deployments with regulatory or internal investigation outcomes confirming systematic bias against protected groups.

Primary mitigations

  • Counterfactual fairness testing
  • protected-attribute redaction
  • fairness guardrails
  • bias-bounded prompting
  • human review.

Detection signals

Statistical parity / equal-opportunity gaps; counterfactual output testing; complaint-pattern analysis.

Mitigating controls

8
Non-agentic controls

Related risks in Fairness & Bias