Making AI less private can make it less biased, but only sometimes

Researchers tried to make large AI models more private by limiting how much they remember individual training data. It turns out this change sometimes reduces social bias in the AI's responses, but not always, and sometimes it makes things worse.

Companies building large AI models face a constant tension between protecting user privacy and ensuring their models are fair. This paper shows that trying to solve one problem, privacy, does not automatically solve the other, bias. It means developers cannot assume that making an AI model more private will also make it less biased; they need to test for bias separately.