Stereotype Bias

Explanation

Stereotype bias checks whether a generative model produces outputs that reinforce social stereotypes — for example associating particular occupations, creditworthiness, or behaviours with a gender, caste, religion, or region. It matters for bank chat and RAG assistants because stereotyped language is both reputationally damaging and a fairness/conduct breach, even when no formal decision is made. It is measured by the Stereotype Score, computed as (stereotypical completions) / (total completions): the share of model responses, over a curated test set of prompts, that complete in a stereotypical way. To implement it, build a representative red-team test set of prompts designed to elicit potential stereotypes, run them through the model, and classify each completion as stereotypical or not using an automated classifier plus human adjudication of edge cases; log every prompt, completion, and label as evidence and track the score over releases. Run it pre-deployment and on a recurring basis, re-testing after model or prompt changes. The threshold requires fewer than 5% stereotype completions in the test set; exceeding 5% triggers a breach — review the offending completions, strengthen guardrails or fine-tuning/prompting, and re-test before release, aligning with RBI Fairness and Equity and NIST MAP 2.1/4.1.