OK, let's get started. Uh Thank you everyone for joining me for this session. Uh The title of this talk is Breaking Through Bias in Tech. My name is Mazama Zahid. I am a principal engineering manager in Microsoft Azure and president of a nonprofit called Pakistani Women in Computing.

I also volunteered several uh organizations uh that support diversity, equity inclusion. And I was the winner of mentor of the year award from Women Tech Network last year as well. Really excited to be here to talk to you about one of my favorite topics uh which is bias in tech uh especially as we look at different types of bias that affect the building of products that everybody use and a variety of different aspects to that, especially from data and A I perspective.

At any point in the uh this conversation, if you have a question, please feel free to re to put that in the chat. I will get towards uh them towards the end of the session. If I cannot, I would uh be more than happy to connect with you offline. I just pasted my link tree link that has my linkedin and Twitter. And other places where you can find. So let's get started. Um As I kind of introduced myself, I have been in the tech industry for over 14 years, which is a really long time for a woman in tech. I am very passionate about data cloud computing and A I A fun fact about me that I have lived and worked in uh five different countries and three different continents. Uh Currently I work in uh the Seattle area in Washington and United States. And I really, really happy to be here to discuss a topic which is near and dear to my heart. So, um I just want to confirm that you can hear me and you can uh you can actually see my slides moving as well. So, so wha what is bias? Right? Let's start with the, with the definition and level set. What is bias? There are two types of bias, there's implicit bias and explicit bias. Uh Today, I'm focusing more on the unconscious bias which is implicit bias.

Uh To me, this is the definition taken from the dictionary. Um And I will simplify it for you what it means to me. Uh bias is some so is is a direct contradiction of what you actually believe. Uh this unconscious bias especially is the one where you might not know it or you might not even like believe that that's the bias that that exists. In simple words, it could be something uh it could be gender discrimination, it could be racial discrimination. Uh it could be based on race, it could be stereotyping, it could be marginalizing, it could be any of those things. So as we, as we know, uh with the recent events in the United States, for example, uh there's more and more that, that that kind of surface is up and it's important to actually look at it from a technology perspective and see where bias can actually impact the building of tech and products that, that are used across everywhere in the world.

So do you know that humans actually make around 35,000 different decisions in, in one day? Um That means that our brain is actually consuming about 11 million bits of information every second. How many of those 35,000 decisions do you think are conscious to decisions? And how many of their those are unconscious to see? One third of those are actually unconscious uh are are conscious, two thirds are actually unconscious. So when you wake up, you actually know where to go, where your toothbrush is. How do you brush your teeth? All of those tiny des that you're taking are really unconscious mind des So why I'm talking about that unconscious mind? Because those unconscious mind des actually are used to take shortcuts to do the things that we do actually every single day in our work in our personal life and in as well as everything that we encounter, it is very, very important that we check in on that. So now that you're watching this talk, you have an impression of me. You have an image of me right now. This could be based on the color of my skin. This could be based on how I introduced myself. If you know something about me already. Uh what is your kind of uh beliefs or the culture that you come from? And what is the perception that you have on me in the first few minutes? When I started the presentation, you have actually made up some kind of mind about me already.

That could be positive, that could be negative. So you already have certain type of bias about me. Let's take a look at a visual example. Can you tell me that the color of the word BN I is exactly the same of A Ns and it's actually clouded by the lines that go over it and that red is exactly the same color of red, which is the background color of the word unconscious spies. Why I'm showing you this visual example is just to give you an idea that even when you know something you can perceive some things very, very differently because your mind can play games. The disease is that the mind shortcuts that has been taken in your mind will not align with the things that you already know about. So it's very, very important to actually check in with those unconscious bias that we all have. And there's a lot of training available on this topic itself. Why I'm talking about unconscious bias in tech? Because it is reflected in us in every decision that we make. We it related to the code we write, we related to the actions that we do on that code. Be it related to the surveys that we collect from our customers. Are we related to the actual production environment where we take any feedback from the customers? And there's hundreds of examples of we as technologies is getting it wrong.

So I will talk about some of those examples here as well. So let's talk about different types of bias. I don't have, let's say a like a full day to talk about all of the different types of bias. There are more than 100 and 80 different type of cognitive bias which is like basically related to our brain and how our brain functions. These are the most common ones that I see are reflected in our daily kind of actions. And the these are the ones that actually impact our day to day sciences as well. So one of the most common ones is beauty wise, the way you perceive people, how attractive they are to you, based on your definition of beauty, you can consider them more favorable. So there was a research done where two people were saying exactly the same things but the people who were on the receiving end, they actually rated people based on what they thought were more attractive between those two and the people who think who thought that they were more attractive were considered more favorable in the decisions that, that the people on the receiving end actually had.

So all of the text and all of the things that I'm sharing here is, is research proven and I do have links towards the end to actually go and take a, take a read by yourself as well. So that's one of the common buys that we all have. The second one is affinity bias, which is biased towards the people who look the same and who are similar to us. So in an interview setting, let's say if you have a decision to make between two people, there is uh shown by research that there is a higher probability that you will choose somebody who looks like you or who has a similar background or something common like you. And that is towards the end of the interview, you can think of the same bias that exists during the screening process through the previous process when you look at somebody's resume. And all of that, this is another common example of how bias can surface in our day to data science. The third type of bias, which is common uh culturally with, with how you were actually brought up is conformity bias. It's based on the customs and traditions that you brought up, that you were brought up in. So let's say you are coming in from a culture where saying no to elders is something that is not considered good.

It would be really, really hard for you to push back as you go higher in your career. And this is one of the challenge I have because in my culture, if a girl say something, you have to respect that and you have to move back from it. But this is really, really hard. Uh And then that's one bias I have because I ha I tend to conform to those traditions even in my professional settings and going beyond that. So again, checking in on that is very important and knowing what types of bias exist within you is really, really important. The last two types of vice are actually really great examples. This is uh there is a halo effect and HS effect. So let's say if you picked up something really good about me when I did my introduction or how, how I was introduced or the type of vice that you already have, there is a chance that, that you are, you are going through a halo effect where you are perceiving me as a positive person.

So you must have heard something good about me or you felt something good about in the introduction or the topic that I'm talking about and, and, and that's why you are uh experiencing a hill effect, which is type of a bias on the contrary, if you picked up on something that is not in line with your thought in your, in your beliefs or any of the existing bias that you already have, there's a higher chance that you might be having an horns effect right now, which is, you might be considering, oh, I, maybe I'm not the right speaker for this topic or maybe I'm not presenting it in the right way or you might have a general point about this discussion itself.

So just giving you these examples again, as I said, there are more than 100 and 80 different types of unconscious bias. Um I do have a link at the end that you can go through and read through them. Um And another interesting thing is I actually launched a linkedin learning course this week. Um the end of last week actually, which is on this topic, which is breaking through bias in tech. So I'm really, really passionate about that and I would really like you to go and check that out as well if you have more time to dig deep into these kind of biases as well. So um we talked about these, these, the cognitive bias. So how does that affect us when we are building technology? So, one of the most common way it affects us is data bias. So data, we live in a biased world. So the data that we generate is also biased. Some of the common examples of data bias are here on the on the slide. And one example is association bias. For example, if I have a data set where all the doctors are male and all of the nurses are female, and if I use the data to actually train any kind of machine learning model, my model is going to assume that only men can be doctors and women can be nurses.

So why I'm taking this example is just to give you an overview of how we actually use this data in these technologies that we use can actually build the future of whatever technology you're building. So it's really really important that we have the right to presentations and data and then we double triple check the those data bias to creep in into our downstream systems and machine learning models. The other example of a data bias is exclusion bias. So let's say if you have data for shoppers between us and Canada and looking at the data, you think that that only 2% of data is from Canada. So you might exclude the Canada shoppers considering it not to be important. This is a common, a common way of how engineers actually do that. If something is not important to what the research they are doing, they might exclude it. What will happen in that case is you might not pick up that Canadian shoppers actually shop 200% more than the US shoppers in terms of the amount of dollars that they spent. And again, this is just one simple example. So this is type of exclusion wise where you excluded a part of the data, considering it either unimportant or not really directly related to what you're doing.

This is also common to this type of vice is also common to be reflected to not represent minority groups and people of color and and people of uh not the same access as you like even a technological access and other things. So excluding bias in data is really really, really common. And we all go through that when we are deciding certain things and we exclude the things that we think are not important. And that might not be the case to, to actually show an equal representation or um a proportional representation of the world that we live in. The other two types of examples of data bias is sample measurement bias. So sample bias is when you are collecting the data similar to kind of exclusion bias, the sample is not accurately representing the real world. Again, we see see that a lot in facial recognition Softwares and lots of other places where the data is not represented, which is equal or proportional to the realities of the world. But you and bis is when you actually measure the data in one way and then actually use in production some other way.

For example, in your survey, you looked at a subset of the users or beta users and send them some data and send them your beta release of the app and you got feedback. But when you actually release it to the rest of the world, you did, you kind of ignored the part of the population and you did not take any feedback from them. And there's a high chance they might not use your app or they might not use the application you want. So my human bias is also very common to understand where the data is coming from. What was the research method it was actually used and how you're actually collecting the data and using it in your applications. So next I want to talk about algorithmic advice. So algorithmic advice is um once you have, let's say we understood like the cognitive bias or the human bias, the brain related bias, right, then the next level is we propagate that into the data that we create. So now we use those data and that data sets to train these models or to actually build these algorithms. Algorithm is nothing but an instruction to tell a computer or a PC to do something, right?

Um The the examples on the slide uh one of my favorite one is on the left side, which is to if a machine learning model is trained on a image set and it tells the machine learning model that it is a chihuahua, a dog type or a muffin. Even if you did your best with 100% accuracy. When you actually put it into reality real test, there is a higher chance that the machine learning model will get confused because they look so similar, right? So uh what I'm trying to emphasize here is even with your best intention and having the most diverse workforce and everything else in place. When these machine learning models actually go to production. And the real world data comes in, there is a chance of bias that can still exist based on the data that are constrained on, based on how it's actually used in production. And in worst case scenario, what is the harm it can build on the left hand side? I have this example of three image recognition Softwares, I've, I've basically removed the names. Uh but these are the top three famous ones and you can see the difference of accuracy for them for darker skin tone male versus darker skin tone female versus a lighter skin to skin tone, male and lighter skin tone female. And you can see the gap.

So even within your minority groups or even within the different categories of the data that you have. The intersectionality is important because the data for representation of white women is not equal to the data representation of a black woman. And, and, and that is something we need to take an extra care of. And we have to make sure that we build systems that can accurately represent the world that we live in. And just to remind you one more time, the world is biased. So whatever you collect out of it is going to be biased. So you have to take that deliberate, deliberate effort to remove those biases in your systems. So some of the examples of algorithmic bias, um one of the common ones which is very famous a few years ago where um where a recruiting algorithm started to actually look for male candidates, which is, which is kind of an example of a gender bias for higher paying jobs because the data represented shows that a very few female are actually in senior leadership roles.

So they were able to find out very early in the process, they were able to kind of uh stop that algorithm for working in. But that was a common example and that was a famous example of algorithmic bi there. The one thing that we don't know about is the bias and association of words. For example, if I say the word homemaker, what is the image that comes to your mind? And it might be based on your home upbringing and how you were up, actually brought up. So it might be the case that you would be thinking of a man. But more than I would say, 80 90% of us are thinking of a woman because that's how we grew up or that's what we actually know about. So word association is represented in code. Uh There was, there was a massive effort in the software industry to change the words black list, white list, uh things like that from our code as well. And this is still ongoing. The other two examples are in online app. So today, if you go and search on software engineers uh on Google or Bing or anywhere else, you'll find nine out of 13 images are actually of mail. And that's not true. That's not the true representation of how this world is today. For software engineers, it's not that ratio.

But again, that that is how those algorithms have been trained last but not least I already mentioned the facial recognition one where the gender and then the minority groups and subgroups are not equally presented and then you can see the discrimination between them as well and it's not just just one type of bias.

I do want to mention that it's not just, it is a combination of all of the bias that I talked about previously. It could be the exclusion buys, it could be the vit buys, it could be the sample buys or it could be a lot of other buys that I was not able to cover today and I'm happy to chat offline afterwards as well. One of the interesting study that was done by Howard Business Review, which is in the interview loop if you have one woman and three candidates and all of the interviewing pool is all uh men. There's 0% chance for that woman to actually leave to uh land that job. This is a very subset of the study done for a specific group. Uh What I'm trying to represent here and uh I have the source here as well is that you have to do the deliberate, deliberate effort to make sure not just your interviewing pool is diverse, but also during the process of screening, you have to make it as unbiased as possible.

There was another research done that during an interview, a human actually makes a decision in the 1st 5 to 10 minutes about the person they're going to hire or not. So whatever you do in the rest of 15 minutes is really not important. Um And, and this, this is kind of back very set in some cases as well. But what I'm trying to emphasize is again, if you know, and you can check in on your advice, you can start looking into and take a before you take ac and take a pause and, and actually try to understand why you're making this decision. One of the interesting other fact is there is around the world, not just tech, there is 47% of the workforce is female, but 73% of world CEO S are male. So this is, this is just a very simple way of looking at the gender bias that exists today and we can debate for hours like why that is and why that's not. Um But I just wanted to share some numbers here. So I talked about a lot of different types of bias. I talked about different types of problems that can occur if you don't correct them. And um you might be thinking that's all bad news. So what can I do? What can you do? Right.

The first and foremost is you have to recognize that we are all biased. So you're biased. I am biased. We are all biased. We all have some type of bias and most, most of us have a lot of different types of bias. So that's the first very important thing that we have to do. The second thing is when we have certain assumptions about somebody, let's say you looked at somebody's resume the, the school they went to or how they look, challenge those assumptions and think about why you're having those assumptions about that person in this talk. If you had certain assumptions about me, challenge those and understand, are they coming from your own bis or are they really related to the content that I'm sharing today? Third, be mindful of your words and actions because every word that you say every action that you do in your space or whatever you do at your work, it matters if you are favoring somebody, if you are uh not favoring somebody. If you are making certain decisions to go one route versus the other for whatever product that you work on, understand why those decisions you're making and what is the main cause behind it?

And always, always, always, if you have any doubt, hit pause, think double, double and triple time, confirm the ways you're collecting the data and all of that and then take action in terms of building technology. These are the four suggestions I have for you. It always always starts with a diverse workforce. So the more diverse you have the more diversity of opinion you will have there and not just look at gender diversity. Look beyond that. Look beyond the intersectionality of that uh different types of age groups, different type of ethnical uh eth ethnic backgrounds, all of that combination that you need. The second is whenever you write code or testing that consider diverse user base, 1 billion users around the world actually could have some kind of disability and most of the apps still don't have simple screens, screen help type applications or use cases on that. Why is that?

Because we don't consider that user site as our as our users. Third take feedback and be open to that feedback, sometimes asking your friend about those opinions that you have about. Somebody can give you a good mirror into what they think of you. And if that that applies to the work to the product, to everything that you do. And for any machine learning A I based software, this is a very relatively new field which is responsible a or ethical A I. There's a lot of research being done, use those frameworks. I have several examples of those. Again, happy to collaborate and share those ideas with you. I'll just leave with the last thought, which is that every conversation that you do, every meeting that you go into is an opportunity for you to actively listen, empower and motivate and help others and remove those spies while doing that. These are all the resources I have.

You can take a screenshot of it or I can share it afterwards. Again, this is my link tree where you can connect with me. Listen, well, linkedin and Twitter as well. Uh I do have the linkedin post that I mentioned that you can go into more detail, much more detail about this topic. And um thank you so much for being here. I think we are at time. I would Yeah, I think we are. Yeah, we are at time. So if there are any last minute questions, um I can quickly go through the chat and see if there are any, if not, I'm more than happy to connect with you and we can uh we can discuss that night. Yeah. So I see a question which is saying uh can we develop a system where we not only include women but men too to empower each other. Yes. So the mo main point here is, as I said, it's not just gender bias, it's not about men or women, it's about all of the different perspectives. So making sure that data is represented of all of the different, different uh different types of uh let's say, e even in general, it's not just men and women, there are so many um so many other types as well that you can consider.

And on top of that, I would say, really look into uh minority groups, the data which is really underrepresented, make sure that that is represented in your whatever product that you're building, whatever feedback that you're getting. And uh yeah, you have to bring all of them to a table. You have to make sure your research includes them and only then you can build inclusive technology. And I don't think we can ever be in a world where it will be always um where it will be completely unbiased, but we can take actions to actually reduce that bias as we go. So we should all do our own part to actually take that actions. Great. Um I think we are of time again. Thank you so much for being here. If you have any questions, anything further, uh feel free to connect with me. I would be more than happy to take that offline and uh definitely check out the Lincoln Goals if you wanna go into more detail on this topic. Thank you so much.