The Myth of Neutrality: How AI is widening social divides

Speaker 1 [00:05]

I'm really, really happy to be here speaking in front of real people, not into the void that is the internet. I'm talking about the myth of neutrality or how AI is widening social divides. Few words about myself, even though I've already been introduced. Machine learning engineer at Innovex. There's some colleagues of mine out there, so come talk to us if you like. And I also enjoy baking, dancing, enjoying and laughing. Unfortunately, though, the talk that I'm giving today is not really a law thing matter, so you'll find out why. Imagine you're driving down the highway when you're suddenly being pulled over by the police. They check your driver's license in their system, specifically the picture, and they find that it matches the one of a person who is wanted for armed robbery. So what's going on? It might be something like this. So the cops checked your license, and they fed your picture into an algorithmic system which is backed by a database of mugshots by people that are wanted for certain crimes. And they find a match for your face. And so the officers will lock you up and you're going to jail. And if you think this is actually a drastic example, well, this has happened to at least three black men in the United States in just the recent few years. And so my talk wants to shed light on how things like this can actually happen, and where along the way of creating AI systems we have many pitfalls, basically, to uncover. And to begin with, I want to look at the overall AI landscape that sort of informs this development. And I want to start by looking at the sectors in which AI is being developed. and first of all let's look at big tech because they have unarguably the most money and the most power to develop these algorithms and deep neural networks and they're sort of you know pursuing certain agendas that for example community and government wouldn't pursue and then there's the military who are also have always been interested in the AI development and we'll find out a little bit more about this, and they sort of steer this into a direction that is more surveillance-driven. Then the next point is that we have a, yeah, let's say homogenous demographic pretty much in the IT sector. There's basically a lot of men in the sector, and this is not only going, I'm not talking about binary gender, and I'm also talking about other axes of identities such as race or ethnicity or religious background. And basically the others are the little unicorns that are painted here. So this also informs which solutions are pursued using AI technology. Third of all, there's also a geographic imbalance. So we have the US and China sort of leading this arms race to get to the forefront of innovation in this sector, whereas the others a little speck of dust as you can see on the slide. Now that we have this bird's eye view, we can start looking at the standard AI development process, and this will sort of lead you through my talk. I will start with research and funding, then go over to data collection and labelling, specifically looking at image data sets, large image data sets that are used for deep learning applications and then go on with training and testing, specifically the metrics that are used and how they can be problematic, and then finally look at the deployment and how the solution can lead to problems in the real world. So I think it makes sense to look at the birth of AI, the birth of this term. Artificial intelligence, the term was coined at the Dartmouth workshop in 1956, which was was organised in attendance by this amazingly diverse group of people that you can see on the slide. And these researchers believe that fully intelligent machines would be a possibility or a reality, rather, until the mid-70s. I'm not sure you've seen any terminators yet, I haven't, so, you know, this didn't happen yet. But I think it's interesting to look at their, you know, what they think about intelligence. So a quote from the workshop proposed was, every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it. Well, maybe you want to think about that quote. I actually don't believe that this is possible because we as humans haven't really established what exactly intelligence is. Really, we have some IQ tests and so on, but this doesn't check for everything that we think intelligence entails. So, yeah, Let's move on to the funding part, actually. Like I said earlier, who is funding the research? Well, first of all, military and intelligence agencies, especially DARPA, which is short for Defence Advanced Research Projects Agency, and they were and still are a major source of funding in AI's early days. They also fund other stuff, for example, they have been funding the Moderna vaccine, but they are still pumping money into this. We can see on the next slide that also the Department of Defense is actually a huge spender when it comes to AI. So this shows the U.S. contract spending on AI by government agency in the year 2020. And this is not money that is spent internally by these agencies on development, but rather given to third-party contractors, so for-profit firms that create AI solutions for the government. And the DOD, the spending by far outweighs any other agency in the U.S. One of the for-profit companies that a lot of money is being spent on specifically by law enforcement agencies is Clearview AI. Some of you might have heard from them in the news. They are a US-based company that sells access to its biometric identification software. And the way it works is basically you give them a photo of a person that you want information on and their software will give you all the other photos that they've got of this person and possibly more information like their name and address and so on. And to create this product, they actually went ahead and they scraped more than 10 billion photographs from Twitter, Facebook, and Instagram. These were public images, hence the CEO of Clearview says, well, all the information we collect is collected legally, and it is all publicly available information. Well, now I want you to think about the last time you might have gone to a party that went really out of hand, and you got really embarrassingly drunk, and somebody took a picture in a really embarrassing moment of you and put it on Facebook without any privacy restrictions. Would you want that image to be in Clearview AI's database? I don't think I would. So I'm not sure it's just information that we're talking about. Now, let's look at big tech. Big tech unarguably has a lot of money, so they're buying everything they can get their hands on. And this slide shows what tech giants have been of the tech giants' acquisitions that cost them more than $1 billion, so everything else they bought isn't even on the slide, right? Between 2000 2000 and 2020. And as we can see, they have been buying more and more for more and more money over the time, and this, first of all, creates a problem because they're monopolising a lot of different markets, and also steering research directions, right? So this really has a bad influence on, you know, the economy as well. And secondly, they're leading to an AI brain drain at US universities or universities in North America generally. In this slide, we can see the number of AI faculty departures in North America between 2004 and 2019. As you can see, the numbers have been steadily rising over the years. All these faculty have left US universities to go and work for big tech specifically, right? This creates at at least two problems that I can see. First of all, these universities are left without their top faculty, meaning they will have less money because less funding will get secured through these people. And then secondly, what the people that made this study, Goffman and Jim, found out was that the students that are left behind at these universities founded less start-ups or less innovative start-ups, which is, again, bad economically. And, you know, some of you might be saying that, you know, but Google and all these other big tech companies, they're doing a lot of research, they're building these large language models and so on, so they do really have a positive impact on research as a whole. But I'd like to say to this that also the research within some of these companies at least does not seem entirely independent as we can see from the quote that I'm going to show you now. This is a quote by a senior manager at Google while they were reviewing a paper on recommendation algorithms written by Google scientists before publication. They said, take great care to strike a positive tone. This is generally not something that you would hear in a scientific review process, or you shouldn't at least, right? This tells you that Google has on top of the scientific review process, when you send your paper to a journal, and you get a double-blind review, they have an additional process which is kind of weird, and also kind of steering this research in certain directions again. One of the most prominent examples of this is actually what happened to these two women, Timnit Gebru and Margaret Mitchell. These women used to be the co-leads of the ethical AI team at Google until the end of of 2020, beginning of 2021, and they wanted to publish a paper that criticised, among other things, Google's large language models, and, well, like I said, they don't work there anymore, right? And if you want to know more about this story, because it's quite shocking, actually, then read this article that I've linked in there. You can find the slides on the conference website. Secondly, let's look into data collection and labelling, and we cannot really get around this ImageNet data set when we talk about this. So, ImageNet is the image classification data set. It's basically what started out the whole deep learning era that we're finding ourselves in right now, and it contains more than 14 million images in more than 20,000 categories, and basically when we talk about image classification, what we mean is we have images, and each image has at least one label associated with it, which we call ground truth. For an image of a cat, you would have cat as a label, for instance. The goal of the authors when they created ImageNet was to map out the entire world of objects, which is ambitious, to say the least. Where did these images come from? Well, they came from the internet. So, the authors scraped the images from search engines and photo sharing websites. And there's also images of people in there. So did they ask these people for consent? Nope. They didn't. And actually, they just looked at whether images had Creative Commons licences, and if they did, they were like, okay, well, that's a free-for-all buffet, let's just, you know, use these images. And they say, like many other people nowadays that are creating these large data sets, well, somebody doesn't want to be in our data set, they can just write us an email, right? But you don't know whether you're in the data set. You just have no clue. And how were these images labelled? You know, the labels need to come from somewhere. And this is actually quite interesting. So, they were based, the labels are based on another data set called WordNet, which was created in 1985 at Princeton University, and it's basically a hierarchical word database. And so, they used these as the basis for these images, for labels for these images, and put tasks up on Amazon Mechanical Turk, which is an online platform where gig workers or ghost workers, as they're called, will have simple tasks like, for instance, they get an image and a few labels, and they have to pick which label fits the best, or multiple labels. With this approach, they created ImageNet Dataset, which was finally published in 2009. Now, I want you to imagine you're one of these people that are labelling these images. You see an image of a person, and now you have to choose a label from among these. So, this is a word cloud, showing which are arguably some of the most offensive labels in the person category of ImageNet. We have words like pervert, bad person, loser, call girl, and some of which I'd rather not read out loud. And I'm asking myself, at least like, I have many questions about this, but how would you know what a bad person looks like from just looking at an image? That seems to be like opening the floodgates for lots of bad stereotyping, right? And then, also, second of all, why would an algorithm, what is the use case for an algorithm to know what a call girl looks like? Not sure. Also, wedding things don't look like this everywhere. And if you're confused now, let's look at this. So we have four images which all show people getting married, right? And these images were labelled by neural network which was trained on the open images data set. And for the first three pictures, you can see that the labels actually fit to what a human might describe this image as. So we have ceremony, wedding, bride, groom and so on. But the last image is labelled as person, people, even though it shows a couple getting married. What we can see here is that a lot of image datasets but also other datasets used in deep learning have a large skew geographically. Here, for the two datasets, open images and ImageNet, we can see that most of the pictures come from North America and Europe, whereas the rest of the world is just vastly underrepresented, which creates exactly those kind of problems like you saw, or imagine a Tesla suddenly driving in New Delhi when it's trained in Los Angeles. How is that going to work, right? Next, let's look at the training and testing, and specifically the metrics that we often use to check the performance of the algorithms. I want to start with talking about a landmark study which is called Gender Shades. In 2018, Joy Bolamwini and Timnit Gebru investigated bioseason commercial binary gender classification systems. Joy's work was inspired by her own experience of not being recognised by open source face detection software. You can see an image of her on the slide. As a black woman, she actually had to wear a white mask for the system to finally detect her face. And she thought, okay, how are these algorithms dealing with different skin colours? How does the performance differ for people with different skin colours? And so she decided to create a data set of lighter and darker skinned people to check this binary gender classification system, or multiple of those, three to be exact. And so she divided this data set into four groups, lighter-skinned female and male, and darker-skinned female and males, and she checked three commercial systems. And what you can see here are the mean accuracies for gender classification at first glance. And so you might think, okay, well, Microsoft has 93.1, Face++ 89.5, and IBM 86.5 per cent. That doesn't seem too bad, right? But then when you look at the subgroup accuracies, you can see vastly different results. So for the lighter male group, you get 100% accuracy for Microsoft, 99.2 for Face++ and 99.7 for IBM, whereas the darker female group has a lot less accuracy. And the largest gap is at 34.4% for the IBM system. And just to make clear how bad this result is, so 65.3% accuracy, If I throw a coin on any of these images, I will get roughly 50 per cent right. This is bad. This is a really bad accuracy. So, what this tells us is that a single success metric does not tell the whole story. Sometimes we really have to dig deeper. Well, Amwini's and Gebru's work then motivated many other researchers to assess biases and try to build fairer systems. And some people actually went ahead and said, okay, let's grab the problem at its roots and create fairer data sets. And one of these, you know, fairer data sets apparently is diversity in faces, which was created by IBM with the goal of advancing the study of accuracy and fairness in facial recognition. And what they did was they took images that were again scraped from Flickr, you know, without any consent and so on, and They annotated them with facial measurements such as facial width and height and how far your nose and mouth are apart. And then they let gig workers again assign perceived age, race, and gender. The reasoning behind this was that the measurements allow better assessment of accuracy and fairness and more fine-grained representation of facial diversity. But I keep asking myself whether diversity is really just represented by a variety of face shapes, or whether diversity means binary gender assigned by some gig worker somewhere in the world. AI creators are the ones that decide about the classification system, right? We create the boxes that other people are being put into, and some people don't fit these boxes, so they are labelled as other. What this means is that, you know, it's centralising power, and Craig Crawford has put this very nicely in her book Atlas of AI. The practice of classification is centralising power. The power to decide which differences make a difference. Now let's get to the final part, deployment, before I say a few words about fairness. I think hiring and firing is actually a very interesting topic when we look at what's going on in the world today, because when we look at traditional recruiting, the way we're used to applying for jobs is that we, you know, see a job somewhere, a job posting that we like, and we send our application to the company, and if they like it, we will get invited to an interview, hopefully speaking to a person, and if they like us, we get hired, if not, we won't. Today, it's a little bit different, because many of us are using online services like LinkedIn, for instance, and the first time we come into, you know, we get in touch with algorithms there, it's very early on when LinkedIn sends us alerts or emails regularly with open job postings, which are, you know, specifically catered for our location and our skill set, and so on, whatever information we offer LinkedIn. So, we might not actually get to see all the jobs that are out there, so they're pre-filtered. Secondly, you know, when you see a job that you like, you apply for it at the company, but they might now not be using a human to go through all the CVs, but they might be using an algorithmic solution for this. And, you know, your CV might be sorted out or not by the algorithm. Well, if you get through, and you're lucky, you, again, get to talk to a human during your interview. But if you're unlucky, then you might get to talk to a service like HireVue. This is a US-based company which created a software where you sit in front of your laptop at home, and you talk to a voice chat bot which is asking you questions that the company thinks are relevant for the job, and they record everything you say and do, they look at your emotions and your voice pitch and so on, and hopefully also about the content what you're saying. I'm not actually sure. Yeah, so the algorithm finally decides whether you're a good fit or not. And the hope is that this will get rid of all the biases that we humans inevitably have, but I just see layers of problems with this. Because first of all, if you talk to a chatbot, you cannot really tell it to clarify something if you you didn't get the question right. It will not understand what you're saying. You can ask that a human, and they might rephrase the question. Also, you will get no feedback after the rejection. If you're lucky, you might get something like a point score from 1 to 5, well, you scored 3.5 out of 5, so you didn't get the job, sorry, but how do you know how the score is calculated? You simply have no information. This also opens the door for discrimination that will, you know, you're not able to prove if you've been discriminated against, so you have no way to challenge this decision. And then third, which I find, you know, most, like, probably the most scary point of this is the scalability of it. Because in the good old days, you went to a company and there was an HR person. If they didn't like you, you know, you went to a different company with a different HR person. But now if you imagine there's a single vendor providing this system to all the companies that you're applying at, you're screwed. You're not going to get the job, right? Well, and also how do these algorithms actually determine whether someone is a good fit for a job? Amazon has actually tried to create a hiring tool, and this was being trained on resumes of applicants over a ten-year period. And as I've said, you know, the IT sector is a very homogenous place in society, so they actually realised that this tool discriminated against women. And they tried to fix it by making it blind to certain words that indicated gender in the CVs. But they soon realised that this didn't work, because the system kept finding ways to infer a person's gender from other seemingly unrelated factors, so proxy variables. So yeah, no way to fix this one. So they took the reasonable decision to trash the system. However, a few years later, so last year I read this article, Amazon now seems to think that firing people with algorithms is a great idea, specifically their drivers in the US. If you really want to get scared, then read the story that I've linked in here. But sometimes people say, okay, well, there needs to be some kind of solution to this. We need to build fairness into the process. We need to metrify fairness of what we think is fair into the algorithms in the systems that we built. I want you to imagine now that you had to build a fair hiring algorithm for IT specifically. Let's say the goal is to get more women into IT. You have this applicant pool, you have 20 applications and you have 10 open positions. We have two groups. female, which are the circles, and male, the triangles, and then we have some kind of extra feature like whether a person is qualified, for example, has a degree in an IT or tech-related sector. I'm sorry for the binary. This is just to make it less complex. I know the world is more complicated than that. So, the first fairness metric that we could use is called demographic parity. And this means that the probability to get hired should be approximately equal for both groups. And if we now draw a rectangle about the people that will get the job, we could, for instance, look at these candidates and give them the job. But some of you might say, well, you know, but now we don't really have the most qualified people in there, right? You know, there's one woman with a degree which didn't get the job, two without a degree which got the job. This seems kind of odd. Maybe we should take qualification into account when we think about fairness. So, we could use equal opportunity, which is similar, but it says the probability to get hired should be approximately equal for both groups, given that the individuals in these groups are qualified. Now, we have one woman that is qualified, and we have ten men that are qualified, and if we draw a line around the people that we could hire, it might look something like this. So we have ten slots filled, and so this means one out of one qualified women gets the job, which is 100%, and nine out of ten qualified men get the job, which is 90%, but this is as good as it gets where we just want to fill ten spaces, you know? But now, there's a problem with this as well because in the underrepresented group specifically, only the individuals that are the most privileged privileged anyway might get the job. So the women that come from, for example, high socioeconomic backgrounds, and this is especially in Germany, we know that this is actually the case, right? We know that people whose parents have an academic background are much more likely to themselves become academics and get a degree than people whose parents don't have an academic degree. So this also seems kind of unfair. And so some of you might say, well, maybe we shouldn't look into opportunities, but maybe we should make the outcomes equal. So we could look at equality of outcomes, also called affirmative action, and that could be saying there must be a 50-50 split between the groups after hiring. And when you think about it, you know, we have ten open spaces, but we can actually not fill them all because when we just look at this fairness metric, we can just hire eight individuals because we only have four women in the whole group. And then again, some of you might say, well, but now we have three unqualified women in favour of so many other qualified men, so how is this fair for them? Because they've, you know, even if their parents had an academic background, it still worked to get this degree, so how do we solve that? Secondly, even if you think this is a good way of solving this problem, so this is something that relates to quotas, having quotas in place, think about it this way. If you, as one of these women, get into a place where there's a male majority, an overwhelming majority, they might just take you for tokens, you know, that join this without doing anything for yourself, you know, to get there. So I'm wondering, I don't know whether you have an opinion on that, but I've been wondering which one of these is fair, right? So is the demographic parity equal opportunity or affirmative action or something else that I didn't specify here? Some of these are actually some of these metrics cannot be had at the same time. And what I'm trying to say or get at here is that fairness is inherently a political decision. So, the context really matters, right? We cannot just do this checklist approach to say, okay, this fairness metric will work in this context, but we really have to think about what we're dealing with with here. And also different individuals will have different opinions on what is fair. And really, we should not outsource these political decisions only to the select few developing AI systems. And I'm counting myself into that group as well. I'm also a very privileged person. Because if we do that, if we keep on doing that, we will end up with what I I call the AI feedback poop, right? To recap what I've been talking about. So, we have this research and funding on the top left, which is driven by military and big tech interests, for profit and surveillance, right? And then we have these image and other data set collection practices where people are not being asked for, you know, their content is just scraped off the internet, and they don't even know what's happening with it. And this then informs the algorithms that we build, and on top of that, we choose metrics that sort of favour the majority group and don't care about the minority groups that are in the data set. And then this leads to deployments or AI being used in the real world that really harms the wellbeing and life of real people, like in the example that I showed at the beginning of the three guys being arrested because their image matched some image in a database somewhere. And the problem with this last use case is that this might actually be used again to train further algorithms because it might be used as a data point in a predictive policing system. So this wrong arrest, you cannot guarantee that the wrongfully made arrest will then be taken out of this data set because someone will go ahead and say, ooh, three years down the line we actually realised this person is innocent. I don't believe that happens. And to finish on a bit of what we can do about these nodes, let's look at my advice for everyone, first of all. Really stay informed about what's going on. Don't only read the stories that say, you know, AI has superhuman capabilities and such and such area, but really think about what are the use cases that you want AI to be used in society, and which are the ones you really don't want AI to be used in. Secondly, join and organise collectives. There's many great collectives. I have some in my references later that are, for example, fighting against mass surveillance, and they're fighting for algorithmic accountability. You can donate to these organisations as well. Also, vote for politicians that really want to tackle this problem, and that see that, you know, AI can create problems, and that we also need to stop this monopolisation of big tech, because this will just get worse down the line. Then my advice for folk in machine learning, be critical. Really, whether it's about your own or someone else's data or metrics or algorithm, don't just look at one-dimensional accuracy, instance. And don't just be satisfied if someone says, okay, let's use this fairness metric, this will solve all of our problems. Then it should be moral first, math second. So it's really important that we identify harms and consequences before we formalise anything. We should really have actually baked this into the process of developing AI. And to do this, I think it's important that we also involve other human beings, because many of us just don't know about the lived realities of people that are already suffering the consequences and harms of these systems being deployed. Here, you should really try to talk to affected communities because they might already know what it's like to be surveilled and, you know, other bad stuff happening to them. Also, talk to social scientists because they have been doing a lot of research, especially also qualitative research, which the machine learning community has been ignoring for decades, and I think it's important to not only look at quantity but also quality, right? And listen and learn from these people. And with that, I say thank you very much for being here and listening to me rant about the state of AI, and I hope you have some questions for me. You can find me on Twitter in this handle, and also check out my references. I've got a lot of great books that you should read if you like reading books or if you're more into podcasts. I've got all the stuff, all the good stuff, series and so on. So check out my slides if you want to know more. And also, you know, down the rabbit hole, scientific literature for the ones who really want to get their teeth gritty.

Speaker 2 [34:23]

okay thanks for the talk we have some online questions but I think you can still submit some but I think we'll also have time to use some questions from the audience before the next talk the first one is where can we find the slides

Speaker 1 [34:43]

On the PyCon website, on my talk website, I guess. The Google slides are linked there. Great.

Speaker 2 [34:52]

And a bit more to the topic of the talk. Is there any change in the day-to-day work in machine learning after so many books and published research on ethics in machine learning?

Speaker 1 [35:07]

It's actually rather a young discipline, so I would say this gender shades paper that I showed came out in 2018, so that's only four years ago, and this was basically the breakthrough for this whole field, for people to actually start looking into it. People have been looking into it beforehand, but looking at how demographic subgroups perform or how the algorithms perform on them, that was sort of the first work that I saw in this regard, and now there's many, many more, and there's conferences on this, scientific conferences, and as we can see, there's also tracks at Python conferences and probably R conferences and so on, so this is a step in the right direction, but generally I would say in terms of day-to-day work, this is not really, this is mostly an afterthought when we're developing AI systems. And that's the biggest problem. How do we actually also tell our clients and our employers that this is a topic that really needs to be looked at? And this is where regulation comes in as well.

Speaker 2 [36:19]

Okay, those are all the questions we got online. So does anyone here in the room have a question for our speaker? I think I can just bring the mic to you. Thanks a lot for your nice talk. It was very interesting. I was wondering, what are the most important laws that we might need to enact in order to combat these problems that you have raised.

Speaker 1 [36:54]

Yeah. I'm not a lawyer, right? But there is this EU AI regulation draft that they've been working on for the past three years or so. And this might actually come into effect in 2024 or 2025 maybe. And this will be a little bit like the GDPR. So they will have massive fines for companies that go against this and this is in my opinion this is one step that needs to be you know tackled um yeah but in terms of i think it's it's diff it's a difficult question to answer because you have people in government deciding about these things that are being lobbied again by google and meta and so on right so they also want to you know have see their interests in these laws and a lot of the times politicians just don't know what they're dealing with they don't know you know what's what's happening with these with these laws and what the effects would be and i think many of us also can't foresee it because there will always be like you know gray areas that you can can be circumvented but i think regulation is important and also if if somebody goes against this that they have to pay a lot of money is an important thing but then i think um regulation can't only be the only answer we also have to have like a civil society who is really aware of this problem so not only people that are working in this algorithmic space but also people that you know like your grandma should care about this and your parents should care about this because surveillance will you know matter to all of us and so this is important to just spread the word and also show them these examples and on what could be happening in the near future and and it's already happening now.

Speaker 2 [38:45]

We have another question from the chat. I'll sprinkle that in. Do you think that we should actively push these thoughts when working for clients? How do you deal with this?

Speaker 1 [38:59]

Yeah. Yes, I think this is important. We also had this discussion very recently at my workplace, and this is sometimes also difficult to deal with, I think, because companies have an economic incentive to do certain things, right? And so I think it's sometimes clever to bring this economic incentive together with these ethical concerns, right? To say, okay, but look, two years or three is down the line. All good.

Speaker 2 [39:33]

Now your time is.

Speaker 1 [39:36]

A bit of, yeah, like two or three years down the line, this regulation will come into place, and then you need to be ready, because otherwise you will pay a fine of €10 million, and then it will be like, oh, OK, maybe we should look into this.

Speaker 2 [39:55]

Okay, we got another question about whether the slides are really uploaded. I will look into this and ask the organizers. I checked earlier.

Speaker 1 [40:04]

I checked earlier, but...

Speaker 2 [40:06]

I'll check with the organizers, but I think some time around today they should be there if they are not already or otherwise

Speaker 1 [40:14]

Otherwise, I can post them on Twitter after the talk.

Speaker 2 [40:17]

Sorry.

Speaker 1 [40:18]

Perfect.

Speaker 2 [40:18]

Perfect.

Speaker 1 [40:25]

Hi, so my question is about the problem that many of these problems that you described actually tie back to Issues that are just present in our society, right? So we get this data that is inherently racist sexist whatever and then the algorithms oftentimes just reflect that And so my question is do you think that we as a community just need to make sure that we're not even worse than this? Or do you think that we should actively try to improve on this standard and shape the underlying society in a way? and if so aren't we then almost reinforcing the same problem where we as the select few actually then start shaping the society of course in a way that we think is right and just and good but maybe not everyone disagrees with this yes uh yes we should so to answer the first question i think it's really important that people who actually know Thank you.

Speaker 2 [41:16]

Because, do...

Speaker 1 [41:17]

inform others right this is this is really important and yes i get um i get what you're saying you know we we will also have certain agendas right um but if we don't share this information this will actually be worse right so this is this is not an option right um and with regards to shaping society it's not actually that easy i mean like many people have tried for a very long time to make, you know, changes be seen in the world around them, and I think if just a few of us try to change something for the better, it will not really have such a big impact on society, but we might get the ball rolling, you know, and this is important. And also keep in mind that, like, advocacy work or, you know, it's tiring, right? And will get a lot of people fighting back because they don't really understand the problem or they're in a very privileged position like big tech companies and they don't want things to change in that way.

Speaker 2 [42:24]

Think we have time for one final question. I think that one was first

Speaker 1 [42:49]

There is a very interesting book that's called Data Feminism that I will just suggest everybody to read. It's super interesting on how to really go to the furnace of the data. I don't know if it's there, probably. The authors are there, the last one for sure. No, it's about how to... It's true that we need to convince politicians to do something about it, but the truth is that if you don't convince society, politicians will not do anything about it. And the question I'm always asking myself is people are annoyed by GDPR right now and now you convince them that uploading their pictures on Facebook is actually putting them inside all these databases and this is, it's dangerous and it's bad and usually what I get is, I don't care, I don't have anything to hide and I was just wondering if you have any strategy or anything when you discuss, when you try to inform people about this topic. Yeah, it is difficult. some people will not be convinced. That's just the way life is, unfortunately. I think for GDPR, what really sucked about GDPR are the cookie banners. I mean, come on. Who likes cookie banners, right? But I think having a lot of examples that might relate to their lived reality is also something that might help convince them, right? And also, you know, It's a slow process, right? We have the saying in German, so if you keep the drops coming onto the stone, at some point it will make a cave, and yeah, maybe also talk to people around them that you can convince that then can talk to them as well, so this is something, if you get the ball rolling at some point, maybe change will happen, but I do not have a recipe, I do not have a checklist, because I think that's actually the wrong approach. I think we all have to find our own way of dealing with this.

Speaker 2 [44:50]

Okay, thanks again for your talk.