Fairwashing and Misleading AI Explanations

Fairwashing is a useful label for fairness language that promises more than an evaluation supports. A model explanation can help people understand an output, but explanation and fairness are distinct. A feature-importance chart may describe a prediction without revealing group differences in error rates, data gaps, or the effect of the surrounding decision process. A global summary may also say little about one person’s result.

NIST’s AI Risk Management Framework treats fairness, explainability, interpretability, validity, and transparency as separate trustworthiness characteristics. A claim that a system is unbiased based on one metric or a small test set can create false confidence. State the population, data, metric, threshold, subgroup performance, uncertainty, and known limitations. If the sample is too small to assess a group, report that gap instead of implying parity.

The EU AI Act contains defined transparency and information duties for providers of covered high-risk systems under Article 13. Article 86 gives a conditional explanation right for specified decisions and affected people. These provisions do not create a universal entitlement to a technical explanation for every AI output, nor does compliance prove a fair result. Explain legal scope accurately and distinguish user-facing explanation from internal fairness testing.

Fairness evaluation should continue after deployment. Changes in data, model version, population, thresholds, or human workflow can alter outcomes. Preserve versioned results and investigate material differences. A concise, bounded statement is more credible than a broad claim of neutrality.