Large language models are moving into high-stakes domains such as legal information, hiring, and public services, where their outputs can affect people's rights and life opportunities. Yet most alignment benchmarks focus on general safety, helpfulness, or bias, leaving the question of adherence to internationally recognized human rights largely unexamined.
The new paper introduces UDHRBench, a benchmark built around the Universal Declaration of Human Rights. It is designed to assess whether model outputs respect the rights and freedoms enumerated in that declaration, providing a more principled yardstick for ethical evaluation than purely technical or safety-oriented tests.
The work highlights a gap between the legal and ethical norms that societies expect and the criteria currently used to evaluate AI systems. By grounding assessment in the UDHR, UDHRBench could help developers identify where models fall short and guide future alignment efforts toward internationally agreed standards.