Podcast: How Artificial Intelligence Creates Discrimination in #HR & #Recruiting
Future of Work Podcast, Episode 2
Artificial Intelligence (AI) technologies are increasingly becoming part of HR and recruiting technologies. In this interview, Jutta Treviranus, Director and Founder of the Inclusive Design Research Centre, discusses how AI pattern identification algorithms can eliminate candidates or employees who don’t fit the common pattern, and how these systems can be improved to better serve everyone.
This podcast is developed in partnership with Workology.com as part of PEAT's Future of Work series, which works to start conversations around how emerging workplace technology trends are impacting people with disabilities.
Welcome to the Workology podcast a podcast for the disruptive workplace leader. Join host Jessica Miller- Merrell founder of Workology.com as she sits down and gets to the bottom of trends tools and case studies for the business leader HR and recruiting professional who is tired of this status quo. Now here's Jessica with this episode of Workology. Jessica: [00:00:25] Welcome to a new series on the Workology podcast we're kicking off that focuses on the future of work. This series is in collaboration with the partnership on employment and accessible technology or PEAT. You can learn more about PEAT at peatworks.org. Jessica: [00:00:38] Everywhere I turn there seems to be conversations in the human resources and recruiting industry centered around artificial intelligence. It is truly permeating our landscape, however, myself, along with PEAT we wondered if A.I. was as inclusive as it could be. We’re talking digging deep looking at the inclusiveness of machine learning technology and exploring the implications for people with disabilities. I’m joined with Dr. Jutta Treviranus. She is the director of the Inclusive Design Research Centre and a professor at OCAD University. Jutta has been working towards more inclusive and accessible technology for much of her life. Her focus is now on improving artificial intelligence systems so they can better serve everyone, including people with disabilities, and organizes hackathons towards this goal. Jutta, welcome to the Workology Podcast. Can you talk a little bit about your background because I think it's really interesting and different than many of the guests on this podcast previously?
Jutta: [00:01:42] Ok. Sure. I've been interested in the liberating potential and worried about the exclusionary risks of computers and networks and digital systems since the emergence of personal computers in the late 70s. I was working with a very very diverse group of individuals at a university and I felt that computers and networks and digital systems had the potential to be wonderful translators. So I worked in that particular field getting to know the technology better and I started the inclusive Design Research Center in 1993 to proactively work to make sure that emerging socio technical practices new technologies are inclusive of everyone. The team that I created which continues and will be 25 years as of next June is a very very diverse distributed community with project partners all around the globe. We grew up with the web and have espoused open practices and communicate practices before I think even open was a common term. We moved to Oxford University in 2010 and I started a graduate program in inclusive design and the graduate program is intended to pioneer a more diversity supportive form of education. We're experimenting with an inclusive education to prepare graduates for all of the variety of things that they're encountering and the many changes that are happening in our community. So one of the things I'd love to tell you about is our perspective on Disability at the DRC we define disability not as a personal trait but as a mismatch between the needs of the individual and the service product or environment offered. So in that sense we are all experiencing disability when the system is not designed to match our needs. I'm hoping that we can see disability as a design issue that can be addressed not as a personal trait that creates an us them scenario. And so designing our artificial intelligence in our HR systems and the way that recruit in our our workplace ultimately to be more inclusive means that we're creating it to be better for all of us and to ensure that fewer and fewer of us experience disability.
Jessica: [00:04:19] This is fascinating to me because you have been working in this area this field for 25 years. And I think long before maybe anyone had certainly thought about artificial intelligence but really focused on inclusive technology and its use for a very long time.
Jutta: [00:04:35] Yes. Actually 25 years ago I started the center but I started in this when the apple two plus merged and Tandy Model 100 and all of those things that I'm sure very few people remember. Jessica: [00:04:49] Can you talk to our podcast listeners about your work in artificial intelligence and disability inclusion? Jutta: [00:04:55] Sure. I've been concerned about how diversity human variability and especially outliers fare in quantified data systems for some time this predates our use of data and big data analytics in artificial intelligence when we use or require big numbers of homogenous outcomes to draw any conclusions. People with disabilities are often the casualties. If you take say the needs and characteristics of any group of people and plot them on a scatterplot you get something like a star. First, it looks like an exploding star in the middle there will be a dense cluster and then other deeds and characteristics will spread out toward the periphery. People with disabilities are represented some distance from that dense core and if you look at the needs and characteristics at the periphery or where people with disabilities are in that starburst you will see that they are much further apart despite the fact that. And this demonstrates that people with disabilities are more diverse than people in the center but we tend to treat anyone beyond a certain invisible boundary away from the dense core the same that huge variability between people with disabilities becomes a problem because it means that you can never muster the numbers to be seen as significant and quite quantify data systems even though the people we relegate to the disability category are the world's largest minority
Jutta: [00:06:27] If not a majority as far as quantified data that requires statistical power and that's what we've required in all of our data systems is concerned people with disabilities are outliers and noise in the data set and what we do with noise and with outliers is that we lemonade it when we clean the data or when we norm the data. The difficulty with AI and why this becomes such a huge issue with AI is that AI bases its decisions on this data and these are hugely important decisions that we don't question because we think that artificial intelligence is not biased that is subjective that it's not subject to human foibles. So what happens is that AI AI does not recognize or understand people with disabilities people with disabilities don't appear in or fit the models. Machines used to make inferences and worse than that. I automates this bias against the outliers and amplifies it. If AI is a black box we can we can't challenge the ai ai decisions. And so what is happening is that people with disabilities are being impacted in quite a number of pernicious ways. One of your top you we're talking about H.R. and especially competitive HR decisions we use AI to filter the applications to determine who's going to be the most promising candidate who's going to get an interview. Jutta: [00:08:04] And if you're not part of the data set there's no data about your positive performance and so someone with a disability who is an outlier who isn't part of the model won't be selected. It however also permeates into other areas of our life loans and credit insurance. If your asset portfolio or profile is not something the machine is used to it won't determine that you are a good loan risk and if you have an anomalous medical history then insurance will not be something that will be decided to be a low risk if there is a security assessment or someone is determining sentencing predicting recidivism. Flagging security risks because you are unknown and something that is not part of what is part of the model or what is understood by the machine then you're more likely to get flagged. And so what you have is a vicious cycle. You are not part of the data set you're not understood. You're not therefore going to become part of the data set. So it's not a case of let's just add more data because the problem is that all of that data is eliminating you. And so you'll never get an opportunity to be part of the data.
Jessica: [00:09:29] This is fascinating because I feel like the sell in human resources and recruiting industry for artificial intelligence is that it's going to remove bias and help make better more consistent hiring decisions. But you're saying that it's removing eliminating an important group of people from even being considered for hire or promotion or any of these things related to the workplace. Jutta: [00:09:59] Right. Because it's based on past data. So it will perpetuate what has happened in the past. The evidence that someone has done well within a job will be based upon past data and the anomalous data. Most data brokerages will eliminate the anomalous state or they're trying to find the dominant patterns the strongest largest group so that it can be a certain decision that is being made. And so in that way it is eliminating what is seen as noise but in fact it's the individuals that are at the periphery of that scatterplot I was talking about or people who are in small minorities. Jessica: [00:10:47] So one of the areas that we've talked about throughout this Future of Work podcast series with PEAT is a look at the gig economy and some of these web based platforms. And there's a lot of different platforms in human resources and recruiting as well as the gig economy that that use AI and machine learning components. How can these technologies do better at incorporating all different types of individuals outside of that scatterplot so that they aren't impacting those who are different and were failing to bring those people into the community or giving them opportunities. Jutta: [00:11:26] So platforms and the flexible economy the sharing economy holds a lot of promise for people who previously face barriers to employment or who have difficulty participating in traditional employment and they hold a great deal of promise for people with disabilities when you are an outlier it is hard to find someone nearby with the same d that you have your subject to. As I was mentioning these vicious economic cycles products are not made for you if they are they cost more. Meaning you're scarce dollars worth less education. It's not optimized for you. Work opportunities don't recognize your potential. If you're lucky enough to get a position the tools you need to do your work won't be accessible meaning you can't demonstrate your optimal performance so that someone who is has a disability and faces barriers to employment in the current job market. There's many things that you're needing to battle. Platforms are a potential way to support people in recognizing their diverse needs and thereby diversifying the demand as well which can then help to trigger a diversification of production and supply. The more we push both the demand and the supply out to those edges the better it is for inclusion and for people that are outliers and it makes it it provides more choices for everybody and when there are more choices than you can find choices that fit you whether it's a product or whether it's a job or whether it's a service. So platforms can reduce fragmentation allowing lú the sharing and pooling of resources which makes it easier to address the requirements of individuals who can benefit from economies of scale. They also provide feedback loops to continuously refine available resources and not only can tools and resources be shared but also the building blocks are development tools needed to create things like inclusive product support training of people that face barriers to employment or to address gaps however there.
Jutta: [00:13:45] So I've been a great proponent of platforms for those reasons. But some of the platforms are largely extractive platforms. The value comes from the workers but the workers don't govern the platform nor do they receive the profit. Because the focus is on short term competitive wealth production for the owners. These platforms are less likely to invest in diversification. Most of the AI products are used to have a quick win. Not only will people with disabilities fare badly in the predictive analytics. But as I was mentioning if you are not dissipating there won't be data that proves your successful performance and the AI that is frequently used in these systems is to find the quickest way to address an immediate demand. There is an alternative and that is there are there there's a platform co-op movements and they are governed by workers and the workers share the profit. They are motivated to create a thriving community. If you're interested in the greater social good it pays to be as inclusive as possible. And so there there is an evolution of some of these platforms and there are these emergent platforms that are not looking at data to support the immediate quick win but data to support a thriving community with a diversification of jobs and better benefits for the workers but also for the employers and for the consumers of the services at the inclusive Research Center. Jessica: [00:15:40] Let's talk a little bit of a reset. This is Jessica Miller-Merrell. You're listening to the Workology Podcast in partnership with PEAT. . Today we are talking about machine learning and inclusion with Jutta Treviranus. You can connect with her on Twitter @juttatrevira. Sponsor tag: [00:15:40] Future of Work series is supported by PEAT the partnership on employment and accessible technology. PEAT's initiative is to foster collaboration and action around accessible technology in the workplace. PEAT is funded by the U.S. Department of Labor's office of disability employment policy ODEP learn more about PEAT at PEATworks.org.
Jessica: [00:15:40] What suggestions do you have for podcast listeners maybe who are thinking about adding artificial intelligence technology to to their human resources are creating our workplace sort of technology stack so that maybe they select the right one or maybe select one that is more inclusive than the other.You focus on three dimensions of inclusive design. Can you walk us through these and maybe talk about how they apply to machine learning and AI technologies. Jutta: [00:15:50] Sure. So Universal Design has a set of principles and accessibility has a checklist. And one of the things that I was encouraged to do was to come up with a set of principles for inclusive design but inclusive design is intended to be relative to the individual. So it's not a one size fits all approach it's a one size fits one approach because we grew up in the digital domain and the digital can be adaptive. Unlike a building where you have to have the entrance work for everyone that might approach the building on a digital system can morphin adapt and present a different design to each one that that visits that digital says. So instead of a set of principles I devised the three dimensions and the three dimensions are the first dimension is that recognize that everybody unique and help people to understand their own uniqueness and create systems that fit that unique diversity of requirements or one size fits one the the second dimension is ensure that there is an inclusive process. This means designing the tables so that everyone can participate in the decision making and in the design because we all benefit from diverse perspectives. We have far better planning better prediction and much greater creativity if we have a diversity of perspectives participating in the design and what you want in inclusive design as co-designer so authentic code design where the individual that the design is intended to be used by is part of the design process and then the third dimension is recognize that we're in a complex adaptive system. No design decision is made in isolation. It reverberates out to all of the connected systems that are in the context of the design. So create a design that benefits everyone and create virtuous not vicious cycles.
Jessica: [00:18:17] I feel like all of these could be extremely helpful in many of these artificial intelligence machine learning companies that are building these platforms or anyone in really in technology to consider when when they are trying to to create a community or a technology that's inclusive to all. Jutta: [00:18:41] Yes definitely. Yeah. The Actually when we take it inclusive design approach to AI AI one of the things that that we frequently do with it in our inclusive design sessions or one activity that I often do is an activity called the grandparent grandchild conversation. I don't know if you have a toddler or if you don't have a toddler there's that continuously the unpacking of why are we doing this. And that is a little bit of a characterization of what happened when I started to look at the inclusive design of artificial intelligence because it prompted us to unpack not just what is happening with in a eye but also being part of an academic's situation. It prompted us to think about how are we dealing with evidence how do we determine truth. Are our research methods that we're using research methods that are supportive of diversity and understanding of diversity. And you can trace the the history of where we ended up with a I and with machine learning and how we're teaching our machines. It traces right back to statistical research and how we are determining what is true and what is statistically significant or what has statistical power within all of our research and academia the in terms of applying those three dimensions. Diversity supportive data and evidence is something that has benefit for all. We have been exploring something called small thick bottom up data that understands the outliers. Making ourselves smarter not just machines smarter as well. Small data is is also called and equals me data or and equals one Data think data is looking at data in context. The process that we try to encourage is transparent parent purchase the parent patrie AI code design inclusive process and in adding diverse perspectives.
Jutta: [00:21:16] And the third dimension understanding the complex adaptive system that we are in if we create machines that understand and recognize diversity and have inclusion embedded in their rules and models it will definitely benefit everyone in all of the experiments that we've done so far show that this is something that's critical not just for people with disabilities but for all of us. The goals that we have for artificial intelligence are better match when we don't ignore the outliers when in fact we are a diversity friendly in our knowledge and understanding. The really interesting and this is not part of your question but the really interesting thing that I found is that previously when we talked about stereotypes or bias or prejudice in conversations with hard scientists or with our fellow academia and it was seen as a soft science as something that was not easily verifiable and therefore in a sort of hierarchy of academic rigor it wasn't seen as something that was as well respected. And but now that we can actually manifest the issues within artificial intelligence we can show that look at if you don't include the state here or if you use only the dominant patterns and you ignore the full diversity if you base all of your knowledge and assumption on homogenous data or the same effect in one large data group then you're not going to do all of the things that you hope to achieve. Whether it's innovation or whether it's risk detection or whether it's better prediction or planning.
Jessica: [00:23:14] I felt like that last piece is I think something that employers or businesses can can really take advantage and think about right. What happens when you aren't including all people from different backgrounds education levels interests. And how does that impact creativity the success of your organization moving forward with a project now. You have some data maybe some research to support the case for diversity work itself is changing quite significantly. Jutta: [00:23:48] We still have systems and understandings that come from the industrial age when we need a direct replicate able workers. And that's really not the reality at the moment. If a organization wishes to survive in the current reality then what you need is you need an agile diverse collaborative team and we've set up our systems and the data that we're teaching those systems information and rules and algorithms from another era the best way to effect the cultural change that we need within work is to create AI systems that are diversity supportive that doesn’t rid us of those that outlying information and that don't ignore people who are at the periphery and who offer an alternative or greater variability of skills and understandings and perspectives.
Jutta: [00:25:54] So the question was what advice to give. In choosing the AI and machine learning programs like if somebody was purchasing it is that. Jessica: [00:26:03] Yes so let me give you an example. There isn't an artificial intelligence technology it's a video technology. And they use AI to measure truthfulness and your body language to determine if you're happy angry sad mad if you are what type of personality are so HR and recruiting leaders are seeing a lot of different types of technology that whether it's video or sourcing technology that will pull the best candidates to the front lines and score them and present them to an employer. What kind of things should they be thinking about when they're looking at these these technologies. Because all of them are not created equal and they're certainly not all likely thinking about inclusive design.
Jutta: [00:26:55] The first thing that I would recommend is that you choose a system where it's apparent what the rules are and you know what set of data the system is trained on. Is the system trained on a full diversity of the types of employees that you hoped to choose. Or is it trained on a homogenous data set that has been cleaned of edge scenarios. The other thing that I would encourage you to do is to find a system where there is the opportunity for a feedback loop where new data can be added where you can add additional filters or at least query the filters and remove filters that cause that may cause bias. The other thing that I would do is to choose a system that is reactive or that is responsive to the scenario that you that you are working in. So not something that only gives you generic data but that allows you to continuously adapt the system to fine tune the working environment and the requirements that you have. And then lastly I would encourage you to get a system that allows you to do some predictive modeling so scenarios that may not have been found within the current data set because the the one thing as I was describing people with disabilities or people who have been traditionally left out of a work environment will not have data that can be used for decisions. But there are ways to model current situations that were not there in the past to see well what with this particular skill what with this new strategy what would this add to the team that I have. That's the the last piece that I would say is that you want any system that understands a team not in a system that is trying to produce a set of repeatable similar homogenous workers. So the an AI system that can orchestrate a variety of diverse perspectives and diverse skills to make a good team within the workplace. Jessica: [00:29:18] Awesome. Well thank you so much for taking the time to talk with us today. Where can people go to learn more about you and you what you do.
Jutta: [00:29:27] They can go to the inclusive Design Research Center Web site. Actually my name is a unique identifier so if you search from my name many things will come up related to inclusive design.You can learn more about the challenges that we design challenges that we've been engaging in to stretch the design of artificial intelligence to be more inclusive. And there we have a design challenge called start your machine learning engines and race to the edge in it. We have many different schools that are putting together edge scenarios that current artificial intelligence systems may not understand or may not have heard of or recognize and we are using those edge scenarios to challenge the ai ai developers to see which one is better at encompassing the full range of human diversity. Jessica: [00:30:36] Well we'll go ahead and put links to the Bigidea.one and then kad your program and then as well as some other research that I have found that was pretty interesting in your work. I'm focused on inclusive design so thank you so much for taking the time to talk with us. Thank you. This by far has been a really eye opening discussion for me and hopefully for you on the subject of artificial intelligence and human resources and recruitment. I’m looking at some ways that this tech might identify patterns that a know and only eliminate candidates or employees who don't fit traditional patterns or models. While I love technology I absolutely do. It is such an important part of what I do. We do need to push back a bit on these AI technologies in our industry and educate ourselves outside of their demos and slick marketing to really understand the things that we need to be asking questions about. Thank you for joining the work I'll you podcast a podcast for the disruptive workplace leader who is tired of the status quo. This is Jessica Miller-Merrell. Until next time can visit workology.com to listen to all our previous podcast episodes.