Live intelligence

Latest AI updates

1 updates
Research

ADDIS separates true intersectional bias from additive effects in LLM audits

Researchers introduce ADDIS (Additive Intersectional Screen), an auditing method that fits an additive model over identity attributes and scores subgroup deviations by Additive Residual Distance against a Freedman-Lane null. On 150 synthetic datasets with no interactions, the standard criterion flagged 66.1 false subgroups per dataset on average, while ADDIS flagged 0.09 and recovered 84% of planted interactions versus 32% for a repaired standard criterion. The authors also build IndiPersona, 10,000 synthetic Indian personas with six identity attributes, and audit 13 open language models; the standard criterion flags 7.1% of 738 tested subgroups on average, compared with 3.0% for ADDIS. Gender identity crossed with sexual orientation is the largest pairing in every model, and counterfactual personas reveal a caste-by-class interaction in all but one model.

LLM fairnessintersectional biasmodel auditing