Bias in, bias out
Models learn from data made by people, so they can reproduce stereotypes or perform worse for groups under-represented in training data. Fixing this needs better data, testing across groups and human oversight — not just a disclaimer.
MAKE IT CONCRETE
Early face-recognition systems were measurably less accurate for darker-skinned women, as shown by the 2018 Gender Shades study.
Ask who a system was tested on.
CHECK YOURSELF