Emerging Technologies and AI Bias Against Persons with Disabilities

uh and if you're you're accounting you might have just heard that the the webinar is being recorded so that's also an excellent thing um you're probably going to if you do have some uh questions there are folks who can actually answer those for you if you encounter any technical or accessibility issues please please reach out to our event support team for assistance um and for attendees who joined us via the enable ottawa virtual event platform please feel free to type your questions or your comments uh into the chat that's located in the virtual event lobby and for attendees who are here through a zoom webinar please feel free to type your questions in the q a box that's at the bottom of your screen so i can see that recording has already started that's fantastic and we have a very full agenda for today full of a lot of great stuff so i'm very excited about that as well what i would really love to do is pass this along as sort of for a good formal opening to suzanne blanchard who is the vice president of students and enrollment at carleton university and she's also the chair of the canadian accessibility network uh she brings a ton of experience uh into the you know her work in these areas and we i believe we said earlier suzanne that suzanne has been opening this session since the beginning of enable ottawa so i'm very excited that she's going to be doing that and as a final little piece just before i hand it off to suzanne um i'll give you a little bit of visual description for any of those of you who could you know benefit from that um so i'm a cisgendered white woman who is wearing some fancy uh blue glasses today it tends to match my hair and i have short hair and i've duded it up in a blazer today so for any of you who appreciate a bit of a visual description that's what i offer to you today so i'm going to pass it on to suzanne so that she can formally open up the day and we can get started with some fantastic presentations thank you so much tara so welcome everyone to enable ottawa this is our third time hosting this event online and we're really so thankful you could be here to join us at our seventh annual enable ottawa conference on innovations in assistive and adaptive technologies we're proud of the legacy this event has created within the carlton and ottawa communities and are pleased to be able to continue to bring together individuals and organizations from all sectors with an interest in advancing accessibility from across the ottawa region across canada and thanks to online platforms across the world enable ottawa plays a key role in how the carlton community celebrate celebrates national accessibility week this year's theme inclusive from the start ties in perfectly with the goal of enable ottawa and with carleton's commitment to continue building on our reputation as canada's most accessible university carlton has an established history in creating an accessible and inclusive environment and we remain committed to continue pushing accessibility forward and ensure it's embedded into everything we do at carleton so that we're inclusive from the start and no one is left out in addition to enable ottawa carlton is also proud to take time throughout national accessibility week to highlight carlton-based initiatives and projects that are contributing to furthering a culture of inclusivity not just on our campus but across canada as well i'd like to make note of the significant process and actions taken towards implementing carleton's coordinated accessibility strategy the first strategy of its kind within the canadian post-secondary environment over the last year dedicated groups from across our campus have worked together to take action on 30 of the 40 recommendations made to move the needle of accessibility forward in addition the canadian accessibility network led by the read initiative at carleton continues to grow with more than a hundred individuals from over 65 organizations representing public private non-profit organizations government department and businesses which have joined us in our mission to create a national platform to empower persons with disabilities improve socioeconomic and health outcomes and change mindsets we're also pleased to share that the read initiative at carleton our event hosts have been awarded funding to continue the david c only initiative project where they are taking the learnings and work from our first phase of this project which ended in 2020 and will now be working towards building capacity of post-secondary service providers across ontario to enhance the employment supports they provide for students with disabilities now before i finish i would like to take a moment to thank the enable ottawa planning committee which includes the reid initiative here at carleton who worked in partnership with the city of ottawa the ottawa chapter of the tetra society of north america and independent living canada to plan this year's event i would also like to thank ai media world-class leaders in translation and captioning who are providing captioning and asl interpreter services throughout our two-day conference in addition to our partners i would also like to acknowledge and thank our many speakers presenters and panelists who have taken time from their busy schedule to participate in this forum and share their perspectives their research their designs and perspectives with us we're so grateful for your support and we recognize that assistive and adaptive technology will play an ever increasing important role in creating an inclusive world we're excited to watch it unfold and get involved thank you everyone enjoy the wonderful program we've put together for you we're really looking forward to your desire to learn your engagement your curiosity and your questions so now tara i'll pass it back to you to introduce our first speaker thank you so much susan and thank you for that opening so i get the pleasure to introduce our first speaker who is going to be providing a a brief and thought-provoking talk on um the emerging technologies and ai or assistive artificial intelligent bias against persons with disabilities it is my extreme pleasure to introduce to you if you don't already know this fine person yuda travanus who is the director of the inclusive design research center but more so uh what i would like to say is this is someone out there in life who causes shift uh and i say that in the best way possible some great positive shift that yuda has been engaged in causing for a long time so yuda take it away thank you tara and do i need to oh sorry do i need to swap my displays again can you see just one slide or two slides we can see the one slide okay perfect thank you okay so um as tara said i'm going to talk about emerging technologies and ai bias against disability and i'm it's a great pleasure to be here uh today with you and to open this up because uh i think as part of and the enabling we need to also have a commitment to do no no harm and um i would like to also acknowledge that i'm in the ancestral and traditional territories of the mississauga of the credit the hood nashoni the nishnabe and the huron-wendap so to begin we know that all new tools come with new opportunities and risks if you've been in this field at all you know that new innovations technologies for people with disabilities are a double-edged sword and if you have a disability the opportunities and the risks are usually at the very extremes and given that we live in the age of acceleration when um innovation and progress is a matter of survival we have a business acceleration that imperative it's either adapt or die with respect to the systems that we're using and it seems that we're on jeffrey west's ever accelerating treadmills where once we one innovation is taken off we have to go into a faster treadmill so what happens then is because of this innovation imperative it has meant that we focus more on the opportunities and not the risks and um there are certainly huge opportunities at the moment we know that machines are automating accuracy consistency and efficiency and they follow a formula but the other piece of the double-edged sword is that this is also the uh the source of benefits but the source of harms and we have devised a whole set of new power tools for making decisions with the promise of greater accuracy consistency and fish efficiency in terms of our decisions and these power tools are wonderful at finding matching sorting labeling measuring optimizing calculating analyzing at scale using big data and with that come extreme opportunities for people with disabilities to recognize speech gestures patterns to find a target object or pattern if you can't see to match and label objects to remember forever and remind on time to sort possible paths to find the optimum to detect common mistakes and correct them and we have a whole cornucopia of wonderful life-changing technologies that are emerging from um systems that will tell you what is in your environment if you can't see it that will interpret your gestures um so that they can be used for a whole set of things that will act as your personal service workers that will automate your um your prosthesis that will replace your vision that will even read your mind um but with this comes also extreme risk we are creating an infrastructure of disability discrimination because these tools are also being used to find match sort label measure optimize calculate analyze people at scale and with that comes deeper levels of mechanized unfairness and harm because the machine decision systems bring not just accessibility barriers as we've seen in digital systems and digital exclusion they come with a lack of representation unfair and presumptive labels and proxies and they're designed to filter you out eliminate your data and flag you as a threat if you're an anomaly so ai as i said earlier automates amplifies and accelerates existing patterns and i'm very very quickly given that we're already late uh tell you about one very quick um area that we're working on at the moment in our project called odd the optimizing diversity with disability if you take two applicants to a highly competitive job one person that is very very average and has a very average life and very average experience and another person that doesn't have an average exp set of experiences or data associated with them who has a disability but is otherwise much more qualified then what will happen is the the most unbiased the best system for choosing someone for the job that has been um certified and um audited to ensure that all bias is eliminated is nonetheless going to not choose the person that is more qualified simply because they are they don't follow the pattern of the individuals in the data set that have had success in the past job and so what happens is that we are following a pattern where within our infrastructure we are embedding uh this bias against anyone that doesn't fit the data set and i'll at the end i'll show a set of links but i won't get into this one um in as great a detail as i had wanted and the the other more alarming thing is what happens to the exceptions when systems actually get smarter one of the defenses of our ai decision systems regarding people is that we just need to give them more data we need to train them more but one of the things that i experienced in testing out systems that supposedly were getting smarter was that smarter isn't always better and this particular alarming experience happened when i was testing machine learning engines and when they were determining what cars should do at intersections when they encountered somebody that pushed their wheelchairs backward which was an unexpected behavior they chose to proceed through the intersection because they because of the assumptions they made about the trajectory of the individual when they were provided with a huge set of data regarding wheelchairs people in wheelchairs traversing through intersections then they decided to run my friend over with greater confidence so the smarter decisions were actually causing greater harm and the one of the reasons for this is that disability is the current ai strategies achilles heel um disability is the universal outlier it's uh it's at the edge of all other justice seeking groups and disability presents an entangled bundle of everything that can go wrong with current ai design um it's useful to know that um one thing about disability is that the only common data characteristic is sufficient difference from the average that things are not designed for you and so from a data perspective that means that you are not going to be recognized it presents in disability you have the culmination of diversity variability the unexpected complexity and entanglement and the exception to every rule or determination most companies and organizations assess risk using numeric metrics people with disabilities that feel the harm the most tend to be seen as a rounding error in the assessment and another way of looking at this is over the past 30 years i've been collecting data on diverse human needs the only way that i can plot this is in a multivariate 3d scatter plot i'm showing here a 2d version of it the needs of a population when plotted in this multivariate scatter plot looks like a starburst i've dubbed this the human starburst like a normal distribution eighty percent is clustered in the middle twenty percent of the space and twenty percent is distributed to the periphery in the remaining eighty percent of the space and it's out at that outer edge in that 80 percent of the space that you find the needs related to disability and what you'll also note there is that the data points in the middle are very close together meaning they're very similar the data points out at the edge are further and further apart meaning that they get more different and more different meaning that people with disabilities are more different from each other than anybody else is from each other in the middle and any statistical the determined prediction is highly accurate in the middle inaccurate as you move from the middle and wrong as you get to the edge and of course predictive analytics and decision systems are using that statistical determination and um what will therefore happen is that everything that is the ai is used to determine will be more difficult or inaccurate or wrong out at that outer edge even those wonderful um assistive technologies that we find if you're in an environment that is unusual and different it won't work those systems won't work well for you if you have a voice that is or speech that's different than average the systems won't work well for you and this particular pattern ripples through everything and of course we're applying ai to everything in our designs and our product designs in our knowledge truth and evidence in our education systems in our work these ai decision tools that are sorting and making decisions about people are permeate almost every decision that is made that is critical to our existence at the moment but the problem predates our artificial intelligence as far as the early 1800s and the invention of the average man ai mimics probabilistic reasoning statistical analysis and linear logic models which reduce diversity complexity and variability ai optimization automates amplifies and accelerates existing patterns of discrimination and so what we have is a self-perpetuating vicious cycle of discrimination reduction in what is seen as evidence and thereby what is proof data and even truth habituation of automated screening evaluation and decision making mechanization of decisions and dismissal of realities counter to assumptions because of course our ai tools are accurate consistent and efficient and one of the things that i'm seeing as well is that it's becoming a self-fulfilling prophecy machine self-fulfilling prophecy means a prediction is made about you and everything conspires to ensure that that prediction will be true we equate evidence with majority repeatability and statistical probability if you are not like the average probability is wrong we equate impact with a single measure for a large homogeneous number what about heterogeneous groups that need different measures the power tools of ai are thereby unable to handle diversity and complexity unprepared for the unexpected folk focus on matching human intelligence um take the turing test which tests whether you will mistake an a.i engine for a person and replicates thereby our own inadequacy and what about ai ethics measures isn't there a huge movement and all sorts of progress in terms of ai ethics but the issue is that those ai ethics measures don't address the particular bias against people with disabilities because it's more than data gaps and algorithmic bias it's how we treat our outliers and small minorities in predictive analytics and what happens in ai ethics auditing tools is that there's a comparison of a cluster of a protect bounded protected identity group or a small minority with the treatment of the average or the typical group and what happens to disability because people with disabilities cannot be clustered are not alike they um individuals with disabilities within these auditing tools tend to fall through the crops and they're stranded at the edges so the bias goes undetected so um you say that or i say that there is no common characteristic of disability a lot of people will say well why don't we ask for self-identification but this goes into my class of simple solutions to complex problems which tend to cause cobra effects the unintended consequences that happen when you think a solution is simple when we use linear thinking when we are stuck in the rut of monocausality not considering the complex adaptive system that we live in and the cobra effect of self-identification is that it excludes unrecognized disabilities humans choose less um qualified people in protected groups in that ai hiring system in fact because what they're trying to do is they're trying to choose um or actually that's probably too complex to explain fully but if you have a question about um how this works within hiring screening i can tell you more about it but the other thing is that it causes cultural and age bias because in countries where there is greater stigma associated with disability or if your background comes from stigmatizing uh disability then you're likely less or you're less likely to identify as having a disability one of the other risks that is associated with the use of ai and its use of big data is that people with disabilities also tend to be the most vulnerable to data abuse and misuse and the privacy protections that have been developed that we're assured will make the systems more just are actually don't work if you have a disability the privacy protections also tend to eliminate the helpful data that allows the ai to treat disability better and one thing to note about disability is most people with disabilities have to barter their privacy for essential services and have been doing so all of their life so um what's the opportunity here because there is actually a great opportunity that has not been explored innovation in the service of outliers and small minorities benefits everyone and that curb cut advantage that everyone here in the field already knows about and if you don't know what curb cut is just look it up applies to ai as well because we live in a complex adaptive system and accelerating flux it's entangled and complexly connected it is rife with feedback loops um change is viral it's not just complicated or amenable to engineering and we are stuck on a local optima from complexity theory what we're doing is we are hill climbing we're optimizing the patterns of the past and we're eroding the slope we're not going to get out of this crisis in order to get out of the crisis we have to find the global optima and to do this we need to include the people that are at the edge who have the better view of the whole terrain are more diverse and are not invested in failing strategies not the people in the complacent middle none of us are safe until all of us are safe so um what we can do if we actually create truly innovative ai and truly innovative process is we can recognize the existential importance of human difference and resolve the tension between diversity and inclusion that we're currently facing because if you look at the view from that edge that i was talking about that's where you find innovation that's where you detect weak signals the first signs of the unexpected crises to come and that's where the people have experience of complexity if we ignore that edge then we do great harm and this is what we try to do in inclusive design inclusive design is not about fixing people it's about ensuring that the systems that we all use are going to be inclusive of everyone so we have three dimensions first of all recognize that we're all different and build systems that integrate the and this um that difference and ensure that people are experts in their own difference secondly um create inclusive process constantly ask who is missing co-design with the individuals that hold the diverse perspectives and that have the greatest difficulty or can't use the current design and strive for benefit for all because right we need to recognize that we're intervening in a complex adaptive system and what we do rather than attempting to find a winning solution what we try to do is to expand continuously the systems that we have to include more and more of our differences and this is these are some of the strategies that we're looking at in our we count project uh where we are trying to shape data science address data gaps and biases co-design protections against data abuse and misuse and co-create more equitable decision supports we provocate with models like the lawnmower of justice where we take the top off the gaussian curve to remove the privilege of being the same as the majority we flip the perspectives we flip the prioritization with tools like our inverted wordle where it's not the popular majority word that grows and moves to the middle but the novel unique minority word or phrase um our odd project is about preventing the creation of monocultures and ai hiring tools by ensuring that we move from data exploitation to data exploration finding a diversity of individuals that can be hired and fill a job and we're also exploring bottom-up community-led data ecosystems where the individual who produces the data is the one that owns and governs how the data is going to be used in decisions because what we've realized is that intelligence that understands recognizes and serves diversity may lead to may be better able to respond to the unexpected detect risk adapt to change transferred a new context has greater longevity may reduce disparity and may help lift us out of our current and future crisis and i would love to continue the conversation thank you so so much judah um so i think at this point uh i we've certainly seen a comment in the chat about how people are seeing this could be something they can apply to uh describe the concepts that you're talking about to to their own groups and i think you know it's so incredibly relevant when you think in this time where the world has gone through so many different very challenging transitions and continues to that um to be able to understand that as we gravitate towards solution finding methods and mechanisms and opportunities as you say we be really mindful that we're not actually just going ahead and building an infrastructure for exclusion uh so i think i i think that you use the solution there um i think one of the things that we've realized is that we need to not think about solutions or fixes these problems are so complex and the minute we say we fix something or we solve something then uh it moves on and there's additional people that we've not thought of and it makes it harder for the individuals that are not addressed through the approaches that we've developed to actually uh obtain accessibility and justice absolutely absolutely so thank you actually for pointing that piece out too is that even it's about a paradigm shift for us entirely around how we even engage with pieces thank you so much for today yuda we really appreciate the time i wanted to let everybody know that we are indeed um recording this event so all attendees are going to be contact

2022-07-17 20:11

Other news