AI models now have a new way to show their bias: who they credit, and who they erase

Researchers built a new dataset to test how well large AI models attribute quotes. It turns out these models often fail to credit authors, especially for quotes from certain demographic groups, or they just omit the attribution entirely.

AI models are increasingly used to summarize and retrieve information. If they systematically misattribute or omit sources, especially for certain groups, it undermines trust and perpetuates bias. This new dataset provides a way to measure this specific problem.