Now we'll turn to an example of algorithmic bias that isn't directly tied to gender and race, just to help illustrate the topic and ground our discussions. This is a recent one, and it's actually tied to the global pandemic. So in the United Kingdom, and I believe specifically in England, Scotland and Wales. There is extremely well known system of examinations called the A levels, the students take at the end of their secondary school time if they wish to proceed to universities. These are exams that you take in specific subjects such as mathematics, which by the way, is called maths in this article. So due to the pandemic in 2020, students weren't able to sit for these exams in the normal fashion. So what happened is instead, it was decided that the teachers of the students in those subjects would assign to them what they thought their grades should have been for those A level exams, based on their performances in the classes on those subjects. The result was that the scores overall for the entire nation were considerably higher than they had been before. And so the government must have decided that it wasn't good to have lots of higher scores. So it decided to create an algorithm to examine this. And what it basically did is created an algorithm. This is not machine learning, much simpler. An algorithm that normalized the scores at the level of a school by school, so that the scores for this year would be somewhat similar to a particular school to the scores from previous years. Well, there is some level of reasonableness to that, but there are also some obvious issues with it. First of all, if you've got a school that has been putting in a special effort and the students are simply doing better than this year versus last year. That would be penalized, and they would not get credit for that effort. Secondly, at a smaller level of granularity, if you've got a student who is particularly good in a school that is low performing that student might get penalized, and actually the reverse can happen as well. And it actually turns out that even though this algorithm mainly reduced scores, there were also cases that it improved scores, and that turned out to be inequitable on a socioeconomic basis. Why is not explained? But it turns out that of the students who scores were increased to the A range or higher, that there were more than twice as many instances of that in schools that in our fee paying schools. What we would call private schools in the United States terminology versus students who are non fee paying schools that we would call public schools in the US terminology. And that seems quite inequitable. So the other thing to say is, dees this make a difference? It makes a huge difference. In the system in the United Kingdom these A levels have an enormous impact on what university students will get admitted to or whether they get admitted to universities. And in fact, what happened is that the university has made admission decisions based upon these teacher generated A levels. Then the government re-calibrated, and then many of those university admissions offers were rescinded by the universities. That of course, led to a big uproar in the nation that was affected in for Scotland and Wales and then in England as well. This is a classic instance of what will be one of the big themes of this module. And that is which is preferable, human based decision making in this case, the teachers making their assessment versus algorithmic based decision making? In this case based on essentially normalizing based on history. In this particular case, there was a sufficient public discussion that the algorithmic version was rescinded and the decisions based on the teachers grading was the one that was utilized. Next, we'll look quickly at two brief articles that discuss algorithmic bias in general terms as opposed to about some specific application. And that particularly look at this human versus algorithmic issue, which makes better or worse decisions that came up in the A level example. The first from the MIT Media Lab, is the article that's entitled Algorithms Aren't Biased we Are. This algorithm or this article focuses on machine learning algorithms and makes a couple of points. The first one we've heard already, and that is that the algorithms reflect the things that they're taught in the training data, which is based on decisions that humans have made. The second is a different point, and it's a point about feature selection. Feature selection is an aspect of machine learning algorithms that we won't really get into here. It's something that can be either done automatically by the algorithm or by the human who feeds the algorithm. But I want to at least motivate it a tiny bit. So think about you being in a position where you're writing an algorithm that decides whether to hire somebody or not for a job that mainly involves heavy lifting. And you've got physical data about that person, and that might include their height, their weight, how much they can lift, and how high they can jump. And you might decide that the first three of those are relevant to this particular job, but the fourth one, how high they can jump is not. That's an example of feature selection, and it's something that does play a role in determining the outcomes of algorithms and potentially their bias. The next article entitled Want Less Biased Decisions, Use Algorithms from the Harvard Business Review, gives another perspective on the human algorithmic decision making dichotomy. And in this case, rather than just focusing on the problems in algorithmic decision making, many of which are due to the human based issues in the training data. It focuses on the end result, which gives more, better, more good, I might say, or less bad decisions, an algorithm that is trained on human based data or the humans themselves. I think it's useful to quote the question that's asked in the article. Quote, how did the bias and performance of algorithms compare with the status quo? Rather than simply asking whether algorithms are flawed, we should be asking how these flaws compared to those of human beings. And so this article looks at cases where decisions are made that fundamentally impact human lives, such as the ones in criminal justice in financial decisions, in hiring. And it's an extensive article that looks at cases over 20 years, and makes the conclusions that algorithms are less biased than humans on the same tasks. It gives a number of categories of examples for those conclusions, including loan decisions, hiring, screening, setting a bail in criminal cases and the choice of board members for corporations. Now, some of what comes out in this article, reflects one of my favorite English expressions, damning with faint praise. Which means it's not much of a compliment at all that the algorithms may be bad, but the humans are even worse. I'm not claiming that's correct by the way, but I'm saying that that's an interesting perspective to consider. You can ask the question. If the algorithms are based on data from the humans, then how do the algorithms come out to be better than the humans? And the article actually answers that question and to quote it, it says, quote, human beings are remarkably bad decision makers. So as I said, I'm just trying to share perspectives. I'm not trying to back one or the other, but it's worth at least thinking about that. I think this does lead to a really important ethical decision that we do need to consider in this course. And that is, when does there need to be a human in the loop? Are we content with life impacting decisions being made solely by algorithms, being made partially by algorithms, or to think that they need to be made solely by human beings? And that's something that will come up in examples that we look at. I'll conclude this lesson with one example of the myriad instances where governments, whether their local, regional or national, are considering making laws and our policies related to algorithmic bias. If you just do a search online, you can find many, many examples of this. I selected one from New York City. That seemed to me to be particularly interesting in the diversity of issues that were involved, including the aspects of both policy and public participation. Let me just mention a few points that struck me as interesting. One, was that there was a fundamental problem of just finding all of the places where algorithmic bias was at work within New York City. Now New York City is a pretty complex and large place, but it was still interesting to see the examples that were being used. Some of them were considerably more mundane than we've talked about in this course so far. There are a lot to just sight to. One of them was deciding the routes of school buses, which streets buses should go down. Another one was deciding which buildings and publicly owned buildings, I suspect, but I'm not even sure should get the most attention for fire inspections. These can be life impacting decisions for people, and they're being made, at least in part by algorithms. It was interesting that the article talked about bringing together people from computer science, from civil liberties, from law and from public policy. And that comment is really at the heart of what we talked about a lot in this course. That we use data scientists, will have an obligation to participate in these sorts of civic conversations, both to inform and to discuss. So that wraps up this first introductory lesson about algorithmic bias. As I said in the next lesson, will concentrate on issues of algorithmic bias that are specifically related to gender and race.
