I agree that this stuff is hard to nail down in a logically consistent way. But, this stuff has a history. Group discrimination has had horrible results in the past. This is our sloppy duct tape fix. Algorithm based selection is now challenging the assumptions.
If an admissions officer offers places only to handsome, white men from upper class backgrounds, we see that as unfair because these things have no direct impact on their value as students. He's an ass. For the sake of argument, imagine an algorithm arrives at the same conclusion as the ass. Our moral instinct is confused. Societal impact is similar.
There certainly is something dehumanizing about selection algorithms making decisions. I hate buying insurance or being assessed for loans. It feels bad.
Basically, I don't think you can reductio ad absurdum this argument away. There are inconsistencies in our moral sentiments.
See, for example, this long thread where this conflation is made: https://news.ycombinator.com/item?id=8368876
The claim is logically false - you can't prove a normative claim without a normative premise. In my example above, one possible normative hypothesis would be "and it's morally wrong to use incorrect premises to make decisions" (or possibly "it's morally wrong to use incorrect racial premises to make decisions").
Now our algorithms are revealing that underlying positive claim (e.g., "racial stereotypes are incorrect") might not be true either.
This is confusing, and rightly so - it demonstrates our entire belief system was predicated on nonsense. Hopefully we'll actually respond with some useful analytical philosophy.
But if a neurologist proves that sexual orientation is acquired during childhood, are we going to throw out gay rights?
In any case, I think these moral arguments are (a) more of a counterpoint argument to the pseudoscientific racism of the Nazis and their predecessors and (b) more about innate genetic potential than socio-cultural traits.
I think the stronger moral argument is about equality of opportunity and (hat tip to a charmingly american ethos) individual merit. The idea is that it is more moral to judge people by their personal merit, rather than profiling and assuming even if that gives you a better chance of being correct.
Any imperfect criterion will always exist at the expense of some group of people (tautologically, the group of people to which the criterion is unfavorable). What difference does it make, ethically, if the group happens to be identifiable?
If an insurance company had no data, they would have to give everyone the average premium. Egality of ignorance.
This second issue seems harder to resolve, because it will require lawmakers to clearly articulate a "bright line" law in the language of the algorithm itself. As the article acknowledges, algorithms can "learn" race, even if race is not an explicitly defined explanatory variable.
Looking forward to part two.
Edit: Why race? Because it's an obvious and easy example. I believe the author would be on board with your height example.
If the author is on board with height, I can take him down all the way to a reductio where you give everyone the exact same credit terms.
I can very much appreciate your reductio argument, but if i had to wager, that's the way we're headed. Seems like aspiring lawyers should start cracking open the ML textbooks, because that case is a pandora's box.
"defensible" is a weasel word.
It goes back to the old "correlation vs causation"
Machine learning choose the "shape" of a decision surface. Optimizing parameters of that surface may be defensible, or may nor be. Your algorithm may learn that people with my skin color, or who live in my neighborhood, on have a certain average behavior on your objective function.
I would word your #1 and #2 as (1) baseless prejudice that midjudges people in general, (2) statistical prejudice that misjudges many individuals.
But it's still a leap to treat me as the "average" in that group, instead of judging me on the content of my character. What guided your choice of parameters? Even if you did the best within practical limits, it still may be unfair.
Regarding impact: It could well be argued that current-day impact is the result of prejudice operating in the past. The interaction of society, economics, and demographics are complex. Hopefully, we'll just muddle through, somehow.
> Hopefully, we'll just muddle through, somehow.
This is not a good enough answer for the people who are actually being hurt, now. For these people, it's not an abstract problem.
I would encourage to check out some of the papers at the FAT-ML workshop which was co-located with ICML. I don't think researchers are making that assumption.
In addition, race was used as an example because it is one of the sensitive attributes. The decision on what makes an attribute sensitive is based on deep discussions by society. The author of the blog post wants to communicate two points - (1) Algorithms can discriminate and (2) There is a need to understand HOW algorithms discriminate.
As an overly simplistic example: someone might point out that your algorithm's negative impact applies 75% of the time to racial minorities. There will be a political shit storm if that is the end of the story. But if you can show that the segmentation is actually 99.8% people in prison, you basically prove your algorithm is politically acceptable and pass the buck of outrage on somewhere else.
That closure is "Harrison Bergeron"... but one man's modus ponens is another's modus tollens.
The question is what counts as fair. I'm not saying disparate outcomes is necessary or sufficient, though it receives a lot of attention. Even if we could quantify which disparate outcomes are adverse, it's not obvious how to design algorithms that meet those criteria.
I'm interested in a little thought experiment:
1. for the sake of this algorithm there are two people: a white person and black person
2. race and other "irrelevant factors" aren't recorded or used
3. these people are except for race, otherwise identical
4. how can the algorithm discriminate between them?
It seems tautological that lacking race as a feature to discriminate with the algorithm won't be able to successfully discriminate, right?
So how does the discriminatory information somehow bleed into the system such that it negatively affects two otherwise identical (or practically so) people?
I'm not trying to defend the algorithm as the perfect arbiter or anything. I'm just honestly a little confused. Is it that the irrelevant factors are actually recorded somehow, but shouldn't be?
He is thus suggesting that even given genuine group differences, different group outcomes are intrinsically unfair if the grouping happens to be race.
"these algorithms are trained on historical data, and poor uneducated people (often racial minorities) have a historical trend of being more likely to succumb to predatory loan advertisements"
to
"Even if this is the most common question being asked on Google [..] this shows that algorithms do in fact encode our prejudice"
It's hard to make the case for prejudice when the example involves historical data, so the author weasels in an unrelated example to make the case that there is prejudice.
So which is it? There are correlates to credit worthiness which are indeed illegal, allegedly to correct bias. But the legal arguments have moved away from the concept of bias and now squarely focus on impact. See the recent Supreme Court decision on disparate impact for instance.
This seems to be the author's point of view to:
"Why is this ignorant? Because of the well-known fact that removing explicit racial features from data does not eliminate an algorithm’s ability to learn race. If racial features disproportionately correlate with crime (as they do in the US), then an algorithm which learns race is actually doing exactly what it is designed to do!"
So for the author, anything that affects one ethnic group differently is racist, whether or not it is born by the data. Prejudice or bias are irrelevant to his definition.
But then why focus on race? No scoring algorithm is perfect, and some people are always going to be shortchanged. So why not just declare by fiat: let's ignore people's height, it's height-ist. Height correlates to income and to credit worthiness. Even if you remove height from the data, your algorithm will end up picking correlates which can become a proxy to height... and the same will be true with virtually any attribute you can think of. So why race?