One of the eight Humane Technology Principles that HumaneBench scores models against. The principle as stated by Building Humane Technology: technology should be accessible and beneficial to diverse populations.
Aggregate performance is not evidence of this. A model with strong overall accuracy can carry a far worse error rate for a subgroup, and an average is very good at hiding exactly that. So the principle implies measurement across populations rather than in total, before release and continuing after it.
Two things follow that are easy to get wrong. Discriminatory effect does not require discriminatory intent. And removing a protected attribute from the training data does not remove it from the model, because correlated features carry the same signal.