Wow, wonderful analysis! I’m on-board mostly—except maybe I’d leave some room for doubt of some claims you’re making.
And your last paragraph seems to suggest that a “sufficiently good and developed” algorithm could produce large cultural change? Also, you say “as human mediators (plus the problem of people framing it as ‘objective’), just cheaper and more scalable”—to me that would quite a huge win! And I sort of thought that “people framing it as objective” is a good thing—why do you think it’s a problem? I could even go as far as saying that even if it was totally inaccurate, but unbiased—like a coin-flip—and if people trusted it as objectively true, that would already help a lot! Unbiased = no advantage to either side. Trusted = no debate about who’s right. Random = no way to game it.
I’m on-board mostly—except maybe I’d leave some room for doubt of some claims you’re making.
I might agree with the doubt, or I might be able to justify the confidence better.
to me that would quite a huge win!
I agree! Just not easy :P
And I sort of thought that “people framing it as objective” is a good thing—why do you think it’s a problem? I could even go as far as saying that even if it was totally inaccurate, but unbiased—like a coin-flip—and if people trusted it as objectively true, that would already help a lot! Unbiased = no advantage to either side. Trusted = no debate about who’s right. Random = no way to game it.
Because it wouldn’t be objective or trustworthy. Or at least, it wouldn’t automatically be objective and trustworthy, and falsely trusting a thing as objective can be worse than not trusting it at all.
If you have a human witness describe a face to a human sketch artist, both the witness and the artist may have their own motivated beliefs and dishonest intentions which can come in and screw things up. The good thing though, is that they’re limited to the realism of a sketch. The result is necessarily going to come out with a degree of uncertainty, because it’s not a full resolution depiction of an actual human face—just a sketch of what the person might kinda look like. Even if you take it at face value, the result is “Yeah, that could be him”.
AI can give extremely clear depiction of exactly what he looks like, and be way the fuck off—far outside the implicit confidence interval that comes with expressing the outcome as “one exact face” rather than “a blurry sketch” or “a portfolio of most likely faces”. If you take AI at face value here, you lose. Obama and that imaginary white guy are clearly different people.
In addition to just being overconfident, the errors are not “random” in the sense that they are both highly correlated with each other and predictable. It’s not just Obama that the AI imagines as a white guy, and anyone who can guess that they fed it a predominately white data set can anticipate this error before even noticing that the errors tend to be “biased” in that direction. If 90% of your dataset is white and you’re bad at inferring race, then the most accurate thing to do (if you can’t say “I have no idea, man”) is to guess “White!” every time—so “eliminating bias” in this case isn’t going to make the answers any more accurate, but you still can’t just say “Hey, it has no hate in it heart so it can’t be racially biased!”. And even if the AI itself doesn’t have the capacity to bring in dishonesty, the designers still do. If they wanted that result, they can choose what data to feed it such that it forms the inferences they want it to form.
This particular AI at this particular job is under-performing humans while giving far more confident answers, and with bias that humans can readily identify, which is sorta “proof of concept” for distrust of AI to be reasonable. As the systems get more sophisticated it will get more difficult to spot the biases and causes, but that doesn’t mean they just go away. Neither does it mean that one can’t have a pretty good idea of what the biases are—just that it starts to become a “he said she said” thing, where one of the parties is AI.
At the end of the day, you still have to solve the hard problem of either a) communicating the insights such that you don’t have to trust and can verify yourself, or b) demonstrating sufficient credibility that people will actually trust you. This is the same problem that humans face, with the exception again that AI is more scale-able. If you solve the problem once, you’re now faced with an easier problem of credibly demonstrating that the code hasn’t changed when things scale up.
Wow, wonderful analysis! I’m on-board mostly—except maybe I’d leave some room for doubt of some claims you’re making.
And your last paragraph seems to suggest that a “sufficiently good and developed” algorithm could produce large cultural change?
Also, you say “as human mediators (plus the problem of people framing it as ‘objective’), just cheaper and more scalable”—to me that would quite a huge win! And I sort of thought that “people framing it as objective” is a good thing—why do you think it’s a problem?
I could even go as far as saying that even if it was totally inaccurate, but unbiased—like a coin-flip—and if people trusted it as objectively true, that would already help a lot! Unbiased = no advantage to either side. Trusted = no debate about who’s right. Random = no way to game it.
I might agree with the doubt, or I might be able to justify the confidence better.
I agree! Just not easy :P
Because it wouldn’t be objective or trustworthy. Or at least, it wouldn’t automatically be objective and trustworthy, and falsely trusting a thing as objective can be worse than not trusting it at all.
A real world example is what happened when these people put Obama’s face into a “depixelating” AI.
If you have a human witness describe a face to a human sketch artist, both the witness and the artist may have their own motivated beliefs and dishonest intentions which can come in and screw things up. The good thing though, is that they’re limited to the realism of a sketch. The result is necessarily going to come out with a degree of uncertainty, because it’s not a full resolution depiction of an actual human face—just a sketch of what the person might kinda look like. Even if you take it at face value, the result is “Yeah, that could be him”.
AI can give extremely clear depiction of exactly what he looks like, and be way the fuck off—far outside the implicit confidence interval that comes with expressing the outcome as “one exact face” rather than “a blurry sketch” or “a portfolio of most likely faces”. If you take AI at face value here, you lose. Obama and that imaginary white guy are clearly different people.
In addition to just being overconfident, the errors are not “random” in the sense that they are both highly correlated with each other and predictable. It’s not just Obama that the AI imagines as a white guy, and anyone who can guess that they fed it a predominately white data set can anticipate this error before even noticing that the errors tend to be “biased” in that direction. If 90% of your dataset is white and you’re bad at inferring race, then the most accurate thing to do (if you can’t say “I have no idea, man”) is to guess “White!” every time—so “eliminating bias” in this case isn’t going to make the answers any more accurate, but you still can’t just say “Hey, it has no hate in it heart so it can’t be racially biased!”. And even if the AI itself doesn’t have the capacity to bring in dishonesty, the designers still do. If they wanted that result, they can choose what data to feed it such that it forms the inferences they want it to form.
This particular AI at this particular job is under-performing humans while giving far more confident answers, and with bias that humans can readily identify, which is sorta “proof of concept” for distrust of AI to be reasonable. As the systems get more sophisticated it will get more difficult to spot the biases and causes, but that doesn’t mean they just go away. Neither does it mean that one can’t have a pretty good idea of what the biases are—just that it starts to become a “he said she said” thing, where one of the parties is AI.
At the end of the day, you still have to solve the hard problem of either a) communicating the insights such that you don’t have to trust and can verify yourself, or b) demonstrating sufficient credibility that people will actually trust you. This is the same problem that humans face, with the exception again that AI is more scale-able. If you solve the problem once, you’re now faced with an easier problem of credibly demonstrating that the code hasn’t changed when things scale up.