4 ms·

The trend of decrying unbiased measurement standards as discriminatory continues. I guess you have to stoke the outrage machine somehow.

by dolni 4y ago

aerostable_slug 4y ago
The article is about measurement techniques like keystroke speed and ability to play games. From an ADA perspective, they're right that it discriminates against those with various disabilities. Otherwise, I agree with you on the ongoing trend to bias things until you get the result you want instead of the result that reflects reality.
JoeAltmaier 4y ago
Obviously the point is, arbitrary 'unbiased measurements' can be cherry-picked to get whatever discriminatory result a bigot wishes. People are awful, and will hide behind technobabble if it strokes their biases. It's good to be aware of this, because it's happened so many times in the past (literacy tests to vote etc). Or maybe AI is perfect and unperturbable, sure.
- dolni 4y ago
  > Obviously the point is, arbitrary 'unbiased measurements' can be cherry-picked to get whatever discriminatory result a bigot wishes. And just as easily, constructive means of measurement can be attacked as "biased" simply because the outcomes aren't what you want.
armchairhacker 4y ago
I would say AI right now (specifically machine learning) is very biased. Not opinionated. ML isn’t a raging bigot who hates women or black people for no reason. But ML really likes to generalize. Because ML itself is basically generalization, it literally works by recognizing patterns. If something looks like a duck and quacks like a duck, the AI will classify it as a duck. For most use cases this is good enough. Except this leads to over-generalization and stereotypes. For example, black people have statistically higher rates of certain negative qualities such as crime. Thus, an AI trained just on raw data is biased to predict that a black person is more likely to be a criminal than a white person with equal stats. People act like this is a big surprise but it shouldn’t be, it’s really just how ML works, and statistically the AI is right. But it’s still biased. What AI can’t do so far (and humans can do) is use reason. AI can only determine correlation, a human can recognize that some factors have more influence than others. Black people are more likely to commit crimes, but that doesn’t mean any one black person with equal stats as a white person is more likely to be a criminal. The data is skewed because a lot of black people also grow up in poverty and dangerous communities, with parents and friends who encourage bad behavior. AI can’t understand that an input parameter which correlates with the output doesn’t actually affect it and should be ignored. And this happens even if you remove the input, because the AI can re-discover it through other correlating inputs (e.g. the AI may implicitly determine someone is black by their location, income, job, hobbies). Hence why the judge AI may still predict black people are more likely to be criminals even if you don’t feed it skin color. Humans use their logical reasoning to “un-bias” themselves to an extent, AI can’t. I personally really support AI but I still think it shouldn’t yet determine things like sentencing and employment. It’s not just an issue with “politically correct” biases, every single “bias” (factor which correlates with the output but doesn’t actually influence it) will probably get picked up by the AI. You may get rejected from your job just because you tapped on the desk too much or looked off into space too much or any other seemingly random factor. AI is still really dumb: it’s good at tasks like image generation, and it’s also good at tasks when it can be helped with traditional algorithms and human supervision (like driving a car). But with interviewing, I just don’t see how AI can actually be useful and not just a buzzword / extra feature.

The trend of decrying unbiased measurement standards as discriminatory continues. I guess you have to stoke the outrage machine somehow.

by dolni 4y ago

The trend of decrying unbiased measurement standards as discriminatory continues. I guess you have to stoke the outrage machine somehow.

aerostable_slug 4y ago

The article is about measurement techniques like keystroke speed and ability to play games. From an ADA perspective, they're right that it discriminates against those with various disabilities. Otherwise, I agree with you on the ongoing trend to bias things until you get the result you want instead of the result that reflects reality.

JoeAltmaier 4y ago

Obviously the point is, arbitrary 'unbiased measurements' can be cherry-picked to get whatever discriminatory result a bigot wishes. People are awful, and will hide behind technobabble if it strokes their biases. It's good to be aware of this, because it's happened so many times in the past (literacy tests to vote etc). Or maybe AI is perfect and unperturbable, sure.

dolni 4y ago
> Obviously the point is, arbitrary 'unbiased measurements' can be cherry-picked to get whatever discriminatory result a bigot wishes. And just as easily, constructive means of measurement can be attacked as "biased" simply because the outcomes aren't what you want.

armchairhacker 4y ago

I would say AI right now (specifically machine learning) is very biased. Not opinionated. ML isn’t a raging bigot who hates women or black people for no reason. But ML really likes to generalize. Because ML itself is basically generalization, it literally works by recognizing patterns. If something looks like a duck and quacks like a duck, the AI will classify it as a duck. For most use cases this is good enough. Except this leads to over-generalization and stereotypes. For example, black people have statistically higher rates of certain negative qualities such as crime. Thus, an AI trained just on raw data is biased to predict that a black person is more likely to be a criminal than a white person with equal stats. People act like this is a big surprise but it shouldn’t be, it’s really just how ML works, and statistically the AI is right. But it’s still biased. What AI can’t do so far (and humans can do) is use reason. AI can only determine correlation, a human can recognize that some factors have more influence than others. Black people are more likely to commit crimes, but that doesn’t mean any one black person with equal stats as a white person is more likely to be a criminal. The data is skewed because a lot of black people also grow up in poverty and dangerous communities, with parents and friends who encourage bad behavior. AI can’t understand that an input parameter which correlates with the output doesn’t actually affect it and should be ignored. And this happens even if you remove the input, because the AI can re-discover it through other correlating inputs (e.g. the AI may implicitly determine someone is black by their location, income, job, hobbies). Hence why the judge AI may still predict black people are more likely to be criminals even if you don’t feed it skin color. Humans use their logical reasoning to “un-bias” themselves to an extent, AI can’t. I personally really support AI but I still think it shouldn’t yet determine things like sentencing and employment. It’s not just an issue with “politically correct” biases, every single “bias” (factor which correlates with the output but doesn’t actually influence it) will probably get picked up by the AI. You may get rejected from your job just because you tapped on the desk too much or looked off into space too much or any other seemingly random factor. AI is still really dumb: it’s good at tasks like image generation, and it’s also good at tasks when it can be helped with traditional algorithms and human supervision (like driving a car). But with interviewing, I just don’t see how AI can actually be useful and not just a buzzword / extra feature.