You Ask, I Answer: Difference Between Fair and Unfair Bias?

In today's episode, John asks, what's the difference between fair and unfair bias? What's the fine line? So fair and unfair uh bias uh really comes down to two simple things, uh two very straightforward things, uh laws and values. So before you deploy any kind of models uh or build any kind of artificial intelligence systems, you need to understand uh what bias is. Bias is when your sample, your data, your whatever it is you're working with, in some way statistically deviates from the population you're sampling from.

And bias doesn't necessarily have to be good or bad, it's it just is. It's a mathematical concept, at least in the context that we're using it here. There's human bias, which is totally separate from statistical bias. Um for example, uh, you may want to sell your product to people who have higher incomes, right? That is allowed, certainly.

Uh it makes logical sense, and income is one of those things that's you know under somebody's control, uh, to a degree, whereas say they're their race, they're not under any control of it. They have an absolutely no choice in what race they are, they have no choice in uh what age they are, etc. So when you're talking about what's fair and unfair, uh, we're talking about are we building tools that disadvantage a population on in some way that is either against the law or against our values? And the two may not necessarily always agree. There are plenty of things that um you may decide as a company or as an individual are unacceptable to your values, even if they are technically legal, right?

You may decide. You don't want to say uh include, gosh, I don't know, firearms owners or coffee drinkers from uh you know doing business with you, and that's totally fine, um, because neither of those things are what's called a protected class. Protected classes are, and this depends on where you live, where you work, where you do business, uh the laws change from principality to principality. Um protected classes are attributes that are protected under law. Uh the in the United States, for example, age, uh, gender, um, sexual identity or gender identity, sexual orientation, uh, race, religion, um, yeah, religion and creed are pr are protected, uh, veteran status, disability.

These are all things that uh you may not discriminate on and have been held up in court um any number of times to say these are not permissible things to discriminate on. And so when we talk about fair and unfair bias, we are talking about uh does your software, which is what an AI model is, does your software in some way disadvantage one people on one of these attributes? If your software uh says that, you know, we're only going to show our product to uh white people, um, that would be illegal. That would be uh a an unfair bias uh along a protected class. And again, you may have values that add additional things that you don't want to discriminate on, that you don't want to um advantage or disadvantage somebody on.

Likewise, you uh bias is also something that uh where you're advantaging one group of people over everybody else. So it doesn't have to disadvantage one group, it can just it can disadvantage everybody except one group. So bias isn't inherently bad unless it crosses the lines of laws and values. Now, where this gets into trouble uh for AI and machine learning and and data science is not necessarily in the protected classes, not even necessarily in the ethics and values, but in how machines use data. And in particular, uh this is deals with correlates.

So a corla is when you have a variable or a group of variables that behave uh very similarly. So there is a strong correlate in at least in America, for example, uh, between income and race. There is you know the the more of a certain race you are, the more probable your income is higher or lower. And so if we're building a machine learning model, say to uh only be able to demonstrate to to show our ad or our uh system or our you know what we want us for customers or who gets in line, uh who gets privileged treatment, um and we're doing it on income level, we are potentially also discriminating on a protected class, right? Because again, there's that strong correlation between race and income.

So one of the biggest challenges that uh folks in the machine learning and AI space need to be addressing is doing things uh you know, doing matching, doing propensity scoring, doing uh regression analysis that demonstrates that the algorithm is discriminating on things that are permissible and it is not discriminating on things that are not permissible. So, for example, if I'm selling high-end coffee, right? Um, and I want to discriminate on income level, which is permissible, um, I also have to be able to demonstrate through statistical testing that uh say a black family uh with the same income level as a white family is just as likely to be shown our ad as the white family, right? Being able to say, like race in this case is not a discriminating factor. Like if you are black and you have uh above the certain income level, um, you are just as eligible just to see our ads.

If but what will happen in a lot of cases um is that people don't do this testing, people don't do this inspection of their models and the outcomes, and they end up discriminating. And uh whether or not the intent was to discriminate does not matter. It's whether you actually did, uh, because in a court of law, you'll be held accountable for your actions. It's like, you know, I didn't mean to hit that guy with a hammer, I just hit him with a hammer really hard. Well, no, it does it doesn't mean intent means nothing.

You hit the guy with a hammer. The same thing is true here, where you may not have intended to discriminate along racial lines or gender identity lines or veteran status, but you did, and you were liable for it. So a huge part of the work in AI and machine learning is to know to look for bias, to test for it, and then to test for correlates to things that are are not allowed, things that are out of alignment with the laws and values. Um and this requires money and time because these are expensive processes to go through and essentially deconstruct a machine learning model to understand how it's making its decisions, um, and it requires a willingness to do so. Now, if the company you're working for or the project you're working on, uh, the stakeholders say no, we don't want to do that, then uh at that point you are personally liable uh for outcomes because at this point you may have said, I I know there's a potential problem, I know that we may be doing something wrong.

If you don't take action to correct it, you're by definition an accomplice. So be aware of that. But it all a lot of what's happening in machine learning really comes down to those correlates, and you just have to test for them. You have to investigate, you have to know that there could be a problem. And that brings me to my last point.

You have to be asking these questions of your models. Do not assume that the software knows to look for bias, because most of the software out there, a few notable exceptions like IBM Watson Studio Studio, do know to look for it. But most software doesn't. Certainly, anything you code yourself does not inherently do that unless you put it in. So you need to be building that in as a process in your systems so that you are looking for fairness, you're looking for unfairness, you're looking for discrimination from the get-go.

And that can happen in the data, that can happen to the people you hire to work on the data, that can happen in the model, and it can happen in the model after deployment, uh, where you get something called model drift, where the model starts behaving in ways you didn't intend it to. So looking for unfair bias it isn't something you do one time. It is an ongoing process when you're working with machine learning tools. So really good question. Uh, there's a lot to unpack here.

There's a lot to talk about when it comes to bias uh in machine learning and AI and in general. Um, but these are things you must be aware of in order to reduce your risk, to reduce your liability, to reduce the likelihood that you get sued for thousands or millions of dollars. If you have follow-up questions, leave them in the comments box below. Subscribe to the YouTube channel and the newsletter. I'll talk to you soon.

Take care. Want help solving your company's data analytics and digital marketing problems? Visit trustinsights.ai today and let us know how we can help you.