AFAIK it’s more “isn’t as good at detecting them as it is at detecting the ‘control group’ which is mostly white and adult”. That’s an important distinction because it’s probably much, much better than a human driver at detecting people as a general rule, it’s just slightly less good with certain groups that aren’t (but should be) in the training set.
For example, an article that claims in the headline that they’re “blind to dark-skinned people” actually explains “In fact, the system struggled to recognize individuals with darker skin tones almost eight per cent more frequently”. 8% isn’t nothing, and should be improved, but definitely doesn’t equate to blindness.
AFAIK it’s more “isn’t as good at detecting them as it is at detecting the ‘control group’ which is mostly white and adult”. That’s an important distinction because it’s probably much, much better than a human driver at detecting people as a general rule, it’s just slightly less good with certain groups that aren’t (but should be) in the training set.
For example, an article that claims in the headline that they’re “blind to dark-skinned people” actually explains “In fact, the system struggled to recognize individuals with darker skin tones almost eight per cent more frequently”. 8% isn’t nothing, and should be improved, but definitely doesn’t equate to blindness.