Representation Bias

Explanation

Representation bias checks whether all relevant groups are adequately and evenly represented in a system's outputs — for instance whether a RAG or chat assistant's answers, examples, or retrieved content cover different customer segments, languages, or regions rather than over-serving some and under-serving others. It matters because uneven representation quietly disadvantages under-covered groups and signals skewed data or retrieval. It is measured by the Representation Coverage ratio, computed as Coverage Ratio = (group outputs) / (total outputs) for each group, with Disparity = max(abs(ratio − mean(ratios))) capturing the largest deviation from the average coverage across groups. To implement it, tag outputs (or retrieved sources) by group, aggregate coverage ratios over a representative evaluation set, compute the disparity, and surface both on a fairness dashboard with the underlying counts logged as evidence. Run it pre-deployment and continuously in monitoring. The thresholds require greater than 90% group coverage and less than 10% disparity; if coverage falls below 90% or disparity exceeds 10% it triggers a breach — investigate data or retrieval imbalance, rebalance corpora or prompting, and record remediation, supporting RBI Fairness and Equity and NIST MAP 2.1/4.1.