THE SHORT ANSWER

A business should define the decision and affected people, identify relevant harms, examine data and process, compare meaningful error and outcome patterns, test accessibility and edge cases, involve domain expertise and provide a way to challenge and correct consequential outcomes.

Bias can enter throughout the lifecycle

Potential sources
StageExample question
Problem framingWhose goal defines success?
DataWho is missing, mismeasured or represented by a proxy?
Model and thresholdWhich errors are optimized or accepted?
DeploymentHow do users rely on or override output?
FeedbackDo past decisions shape future training data?

Evidence & context: National Institute of Standards and Technology

Choose metrics from the harm and decision

Different fairness measures can conflict. A team should explain why a comparison is relevant, examine sample size and uncertainty, and combine quantitative testing with process review and stakeholder insight.

Use scenarios to reveal hidden assumptions

Illustrative example: a support-priority model appears accurate overall but misses urgent requests written in a less common language. The response is not only a new average score; it may require better data, routing rules, language access, human escalation and monitoring by segment.

Build fairness into the operating process

  1. Define affected people and plausible harm.
  2. Test relevant groups and edge cases.
  3. Document limitations and excluded uses.
  4. Provide human review and recourse.
  5. Monitor outcomes after deployment.

Evidence & context: National Institute of Standards and Technology · OECD.AI

Sources & further reading

  1. Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    National Institute of Standards and Technology. Voluntary, rights-preserving guidance organized around GOVERN, MAP, MEASURE and MANAGE. NIST was revising AI RMF 1.0 when checked on 28 September 2026, so organizations should verify the current version before formal adoption.

  2. Towards a Standard for Identifying and Managing Bias in Artificial Intelligence

    National Institute of Standards and Technology. Official NIST analysis of systemic, computational and human sources of bias across the AI lifecycle. It does not prescribe one universal fairness metric for every context.

  3. OECD AI Principles

    OECD.AI. Intergovernmental principles updated in May 2024 covering inclusive benefit, human rights and fairness, transparency, robustness and accountability. They are high-level guidance rather than a complete operational control set.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗