Exploring Gender Bias in GenAI with Laila Sprejer

[Music] Hi everyone and welcome back to another episode of your career is a choice or chance. Jan AI in the workplace. I'm here at your host and for this amazing episode I'm beyond thrilled because we'll be dropping special episode in honor of International Women's Day. We will dive into the intriguing world of Jan AI and gender bias and guess what? We've got a fabulous guest in the house. She's deadled into this topic in the London School of Economics. A few years back and today she's leading the Data Science and AI themes in Emma's and Data One in Endocs. Hello Laila. Hello, thank you. It's great to be here. But before we begin let me introduce our AI co-host Alex is always Alex is here to learn to listen and to bring their AI unique perspectives to the final question of the episode. Thank you Yael and please to meet you Laila. I'm excited to join as your co-host in this fascinating discussion about gender bias in Gen AI. So Laila can you share some overview of Gen AI and the bias data and elaborate on the risks that associated with it? Yes of course. So I don't know about you but when I hear about Gen AI the first thing that I think about is judgey beauty and I think about how it helps me write code, how it helps me prepare to listen to you you know how it helps me plan my travel. But what we need to really think about is how Gen AI and AI in general is affecting all of our lives on a daily basis. The music that you hear, the moves that you see now it's algorithms that recommend you what to see and what to hear. When you buy something on the internet that's also based on that. When you apply to a job and you send your CV you have an automated algorithm that selects your CV based on some criteria. When you apply for a loan in a bank then you have a system that automatically can pre-select your peer reject you. So what you need to think about is what happens when those automated systems or semi automated systems are biased. In those cases it can have real impacts on our lives. For example representation bias and I'm pretty sure that you already saw Gen AI image generation that are biased to generating male CEOs and not women CEOs. Those reinforces stereotypes and generating our representation harms. It's not only about gender it's also about the egemonic representation of women and also about racial biases. So where does this bias come from? So I think it comes from a lot of places but the most important one is the fact that algorithms and Gen AI can only say what they been told to say and these algorithms are trained on data and this data if this data is biased then what the algorithms do is also going to be biased. And this data we need to understand that it's the consequence of hundreds of years of our social norms and social conceptions of women having certain role saying as particular things behaving in a certain way. If we go back to the example of a stable diffusion algorithm where some researchers ask the algorithm to produce images of CEOs and then most of the images that the algorithm produced were male CEOs. When we try to understand why it because most of the data that this algorithm was trained on for the CEOs were male it doesn't mean that the algorithm is lying it means that the algorithm is biased but that what's the other source of bias that it's important to mention the fact that you don't have enough women being out taking an active part on the process of developing these algorithms. So the misrepresentation of women in the data science and AI field is also a factor that contributes to the bias. So algorithms are being in a way developed by a biased or misrepresentative group of people that it's just a vicious circle if I understand correct. Exactly it's a vicious circle and we can discuss why you have that under a presentation of women in the field it comes not only from the the challenges that women have to enter the field but also the challenges that women have to stay in the field. Harassments in the workplace or the long hours of work in these particular fields the maternity leave and the need to do work at home and possibility to keep the pace that this can't work require so there's also all kinds of reasons why we have this misrepresentation. So you as a woman working in AI leadership and related to you know data science in AI. So I'm interested to know what challenges you have encountered and you know how do you recommend to overcome them. So I encounter all kinds of challenges in first person and in third person for example these things that I said about harassment about having to spend time with kids and and the household work this is a common problem that I see in on on new earth colleagues. The things that I mostly encounter myself around having always worked in things of males you know that's usually that's how we usually been in my case. Probably the most the most important one to me is around how men and women perceive differently their skillset and their confidence level on how much they know and their ability to speak up and stand in our further thoughts. There's a difference in how I think in how the workplace perceives a woman that is strong and picks up and decides what you know what we need to do and says this is what I believe is what we need to do and I think it's different when men men say that and when women say that. So something that I had to actually work throughout my life I think is to decide that I do know enough that I do believe in what I'm saying and I do need to speak up and I think that is something that I had to learn over the time. I understand now so what we're talking about that it is crucial to encourage women to take an active part to engage with Jenny. Let's start with that you know to engage to play with to ask the questions to interact with etc but also asks in gender or yourself in gender related positions I understand it's even more crucial right because you touching the heart of the thing itself in the data so do you have some may you know some may recommendations what organizations can do what women leaders can do crucial that women take part of these processes and also when I'm saying women I'm talking about all misrepresented groups like this is not only a problem for women but it's absolutely crucial that women take part of this process of developing algorithms they can help mitigate biases find biases because the first problem to stop them is to to find them right this is also why I think it's so important that when a government level and a company level went to enforce the practice of reporting metrics on biases of being transparent on on how algorithms are performing and how is our team composition and what percentage of women and men we have doing courses you know promoting employees to take courses on gender biases and management and by the way I think this is something that I'm not that's great and I'm taking the gender bias course and it's I really really like it so as we are celebrating or nearing international women's day or man I want to allow Alex to ask a final question here so Alex Lila how can we identify and mitigate gender bias in gen AI algorithms there's a lot of techniques that are you know nowadays are known on how to mitigate bias in algorithms the first step is data selection you know from the starting to be sure that your data is not biased and we need to monitor this continuously because the fact that it's not biased when we start the experiment doesn't mean that it's not going to get biases we go on and we know that also the world changes and the data changes with the world so this is a continuous effort that we need to do we also need to have big teams like we said before I'd need to have a representative team of people that can build these algorithms to ensure that that we're not building a biased product we need to work always towards explainable AI and transparency and measure and report metrics and biases and ideally we you know we can always have a human in the loop that checks the results before sending something outside right that's that's also another option that instead of fully trusting algorithms to market autonomous decisions also allow allow some monitoring from a human although human also have an ongoing so spy yes but that's another story right we can't fix it all right but it's important to to detect them and to acknowledge them and to know that all of us haven't got your spy yes and that's natural but if we can measure them and we can acknowledge they exist then we can do something about it and we can break the circle so Lila Alex I definitely have some pledges for the new you know the next year as to what to do in order to prevent some biases that are almost built into this and thank you very much for getting us thinking thank you thank you all for inviting me