These new AI benchmarks could help make models less biased