Our methodology aims at using a Human Rights-Based Approach to investigate questions to be explored at each phase of the AI Lifecycle, focusing on fairness and debiasing AI.
The tools help people build and buy AI responsibly.
HumRights-Bench asks a prior question: can the models themselves reason about rights?
We put the leading systems to the test against international human rights law. They scored barely above chance, and stumbled worst at the most basic task, recognising when a right has been violated.
It is the first expert-validated benchmark of its kind.
