2 ms·
I have a problem with this whole algorithmic bias thing. If there's an algorithm that explicitly deals with race (e.g., If race == "black" put at end of line) t
by pmdulaney 7y ago
I have a problem with this whole algorithmic bias thing. If there's an algorithm that explicitly deals with race (e.g., If race == "black" put at end of line) then yes, that's a problem. But if an otherwise well-thought-out algorithm happens to treat blacks and whites differently, why is that a problem.
Suppose an algorithm designed to help people at risk of committing suicide ends up disproportionately helping middle-aged white males (who happen to commit suicide more than any other group). Is that a problem?
Or if an algorithm designed to deal with hypertension ends up giving more benefits to blacks? (who statistically are more likely to suffer from hypertension)
What you choose to optimize for is definitely an important matter, as Norbert Weiner's famous story about optimal defense against a nuclear attack illustrates.
- DanBC 7y ago> disproportionately We need to pay attention to high risk groups. We also need to pay attention to lower risk groups, especially if we do not understand why they're a lower risk. If they're a lower risk because they use different methods it's very dangerous to not pay enough attention to them, because method substitution can happen rapidly. And if this previously low risk group starts using a more lethal method their rates of death will suddenly increase, and we don't have any information about them because we didn't bother paying attention. > who happen to commit suicide more than any other group This is only true in some countries. In other places young women are the higher risk group. Imposing Western trained algorithms on a global audience will (and has) caused harm.