To be useful, ML systems have to have some kind of bias. However, the distinction here is that some of these biases are harmful biases. Kate Crawford talks about allocative harms (how resources are allocated) and representation harms (e.g. stereotypes).
Some of these harmful biases are really blatant. For example, labeling Blacks as "Gorillas" is offensive for many reasons.
Some of these biases "correct" but that's due to biases in the data set or society. The ProPublica investigation of recidivism prediction is a good example, where it was more likely to say that Blacks should not be released. However, police are also more likely to arrest Blacks, which naturally leads to this bias. Other examples here include Amazon's resume system that was biased against women (since they used Amazon's hiring practices as ground truth), and image search for "professional hairstyles" that showed White women but "unprofessional hairstyles" that showed Black women.
Other biases are also "correct" but greatly miss the underlying context. For example, a naive AI system might tell you don't go to a certain medical doctor that is a professor, since they have a higher rate of deaths. However, this doctor might also be a doctor of last resort, hence the high mortality rate.
What I'm trying to get to is that even the term "correct" has a lot of subtleties to it. In many cases, figuring out what is "correct" (or ground truth in ML terms) can be a clash of values and world view, and might have different results based on differences in race, gender, age, culture, context, and power.
Given enough data, and no doubt Amazon surveils their people more than most, they could determine the 'truth' along a more straightforward line.
"Does this hair style make more money for the company"
As hair can be a strong form of expression, there's probably a measurable delta here.
Going forward, smart companies will obfuscate the determination. I suppose that training an AI is not a bad way to pull this off.
(This is all further demonstration of just how complex a term like "correct" can be in this case, as you point out, but I think it's worth considering the whole spectrum and perhaps the "Devil's advocate" instances of potential correctness.)
This is one example recently given by the founder of a controversial AI-based insurance company (https://www.lemonade.com/blog/ai-can-vanquish-bias/), where he claims sufficiently fine-grained AI classification actually reduces group bias even if in some cases the output may, in aggregate, be statistically biased towards particular groups:
>Let’s say I am Jewish (I am), and that part of my tradition involves lighting a bunch of candles throughout the year (it does). In our home we light candles every Friday night, every holiday eve, and we’ll burn through about two hundred candles over the 8 nights of Hanukkah. It would not be surprising if I, and others like me, represented a higher risk of fire than the national average. So, if the AI charges Jews, on average, more than non-Jews for fire insurance, is that unfairly discriminatory?
>It depends.
>It would definitely be a problem if being Jewish, per se, resulted in higher premiums whether or not you’re the candle-lighting kind of Jew. Not all Jews are avid candle lighters, and an algorithm that treats all Jews like the ‘average Jew,’ would be despicable. That, though, is a Phase 2 problem.
>A Phase 3 algorithm that identifies people’s proclivity for candle lighting, and charges them more for the risk that this penchant actually represents, is entirely fair. The fact that such a fondness for candles is unevenly distributed in the population, and more highly concentrated among Jews, means that, on average, Jews will pay more. It does not mean that people are charged more for being Jewish.
>It’s hard to overstate the importance of this distinction. All cows have four legs, but not all things with four legs are cows.
>The upshot is that the mere fact that an algorithm charges Jews – or women, or black people – more on average does not render it unfairly discriminatory. Phase 3 doesn’t do averages. In common with Dr. Martin Luther King, we dream of living in a world where we are judged by the content of our character. We want to be assessed as individuals, not by reference to our racial, gender, or religious markers. If the AI is treating us all this way, as humans, then it is being fair. If I’m charged more for my candle-lighting habit, that’s as it should be, even if the behavior I’m being charged for is disproportionately common among Jews. The AI is responding to my fondness for candles (which is a real risk factor), not to my tribal affiliation (which is not).
One thing his post doesn't discuss is what might cause such a group correlation and how much agency is involved. In the case of candle-lighting, it's presumed that people (Jewish or otherwise) are doing it purely out of their own free will, or at least due to a belief/practice they largely have choice over.
If instead the root cause is hypothetically partly or wholly extrinsic (e.g. police being disproportionately more likely to arrest people among certain groups, with it remaining disproportionate after accounting for the true crime frequency/severity base rate for individuals in that group), then I think an analogue of the above example wouldn't hold up, because, as you say, the inputs would be inherently unjust, even if they're in some sense statistically predictive. So it'd be unfair to use such data.
Then there's the grayer area. What if a group is hypothetically disproportionately represented among a certain data set or proxy but the representation is commensurate with the true base rate among individuals of that group?
In some sense, it's not unfair, because you're getting actual data based on what people are actually choosing to do or not do.
But it opens the door into larger arguments of culpability, free will, being dealt a bad hand, etc. It's inherently unfair to be born into a very poor family or a crime-ridden area or a house with lead paint or as the ancestor of generations of people who were oppressed, abused, shut out of society, and otherwise treated very unfairly, let alone potentially abducted, enslaved, and/or subject to genocide. Even if the true rate hypothetically lines up with the proxy, there still might linger impactful and lasting trickle-down effects from generations of very unfair and incorrect proxies. So it could potentially be correct inputs, correct outputs, but still unfair in a deep sense. However, is it unfair to the point of it violating discrimination laws? I don't actually know. And I could see many different arguments about the ethics of such outputs.
And then there are of course the epistemological problems / meta-problems, here, which might be the trickiest of all: how do or can you know the data is accurate, how do or can you know the true base rate, etc. So it's very difficult to tell in practice how fair any particular metric is.
Bias is clearly a major issue for AI, but I think it's a pretty nuanced subject. It's easy (but of course deeply necessary) to list all the actual and theoretical failure modes, but it's hard to always truly determine how fair something is and exactly what ethical and philosophical principles to use when judging fairness.
I know that's basically just a reiteration of your point, but I always see this framed from the perspective of how easy it is to get things wrong, without examples of cases where one could potentially "steelman" the wrongness; or earnestly steelman it yet still ultimately conclude it doesn't conform to a particular society's values, even if it might conform to laws. (Or at least a subset of a society's values - given some of the seeming fundamental value divides in the US. Two people could agree about most of the above but come to very different conclusions if one of them is socially left-leaning and the other is socially right-leaning.)
Such a thought provoking post. Thank you. So much to learn from this. I would expect no less from CMU.
Recall all the times in stats where an estimator can be an unbiased estimator of a correlation while being a biased estimator of a causal effect.
So you get some people saying it (the correlation) is correct and other people saying it (the causal effect) is incorrect. Both are right! To stop talking past each other, they need to talk about bias with respect to the correlation or bias with respect to the causal effect in this particular direction.
But what frustrates me is when the correlation side uses the (true) correlation to argue against a system being biased with regards to something else (w.r.t. a definition or w.r.t. a causal effect or w.r.t. a literal translation or w.r.t. some more complicated aspect of the system), and that harms are okay because the bias is a correct bias.
We need to work on our terminology so that we can stop talking past each other. It doesn't help that our models have weird biases in absurdly complex function spaces, but we have to progress beyond a first-stats-course one-size-fits-all definition of bias.
Also, who is (oft-)suppressing the "elephant in the room"?
If you see in a data set that Danes are tall, and that Kenyans are fast, and that Ashkenazi’s are smart, then it is a valid hypothesis that should not be thrown out outright, that the reason that’s the case is due to actual differences inherent to the population groups and not any other confounding factors.
As for your second question: mostly progressives, leftists and liberals.
Example: men are known to commit the vast majority of violent crimes. But using that statistic to convict someone, deny them a job etc. would be inappropriate.
1. Someone from Utah is more likely to be a member of the Church of Jesus Christ of Latter Day Saints than someone from Pennsylvania.
2. Someone from an Arab speaking country is more likely to be Muslim than someone from a non Arab speaking country.
3. Someone who says "eh" at the end of every sentence is more likely to be Canadian.
4. Someone who says y'all is more likely to be from the south.
5. If someone asks me to "Please do the needful" they are likely from India.
I've purposely chosen non extreme examples because there are many basis all over the place. Bais ≠ prejudice.
Ultimately if we artificially restrain AI from being "baised" in any form we are really shooting ourselves and those most disadvantaged in the foot because instead of being able to use AI to discover the basis and then work on fixing it we instead just to pretend it doesn't exist.
Finally a more provocative example. People who get pay day loans are less likely to pay back loans, black people are more likely to use pay day loans, ergo black people are more likely to default on loans. If we try and just force an AI to ignore this then we paper over the problem. If instead we start to examine causality we can start to figure out the root of the issue and how to address.
I mean it took so long for Kahneman and Tversky ideas on bias to disperse that we can even be talking about bias in this context.
Bias isn't even the real problem with ML, noise is obviously.
