|AI is making a huge difference in the world, but developers must be diligent and mindful of how they build AI and models, making adjustments if things do not go well / Photo by: Daniil Peshkov via 123RF|
Even if we celebrate the International Day of Persons With Disabilities each year, we must remember that AI is not a panacea, explained Shari Trewin of Venture Beat, a website on transformative technology. AI is making a huge difference in the world, but developers must be diligent and mindful of how they build AI and models, making adjustments if things do not go well.
Sure, there are cases of unwanted bias in AI systems, but sadly, many developers are unaware of the need to treat marginalized groups fairly, especially with regard to race and gender. More unfortunately, disability is one aspect of diversity that has been ignored. Whether our disability is temporary or permanent, technology should accommodate this part of the human experience.
Statistics on People With Disabilities
Having a disability places an individual in the world’s largest minority group, stated Disabled World, an information and statistics website on world disabilities, disability research studies, and more. Currently, around 10% (650 million people) of the world’s population live with a disability. Around 80% of persons with disabilities live in developing countries, as revealed by the UN Development Program (UNDP), the UN’s global development network.
Furthermore, the World Bank, a unique global partnership focused on combating poverty worldwide, disclosed that about 20% of the world’s poorest people have some kind of disability, as mentioned by Disabled World. They tend to be perceived in their own communities as the most disadvantaged. People spend on average about eight years or 11.5% of their life span living with disabilities in countries with life expectancies over 70 years.
Alarmingly, disability rates are higher among groups with lower educational levels in OECD (Organization for Economic Cooperation and Development) countries. An average of 19% of less-educated individuals has disabilities compared to 11% among individuals with higher educational attainment.
It is estimated that 13.3% of 18- to 34-year-old Americans have a disability of some kind and are already suffering from worse employment outcomes, standing at 6.4%, according to a 2017 report by the Census Bureau.
How AI Improves Accessibility for People With Disabilities
Smart Home Speakers and Voice Assistants
Smart home speakers like Google Home, Amazon’s Echo series, and Apple’s HomePod is not just for people who want to queue up their favorite podcasts without moving away from their couch, noted Kyle Wiggers of Venture Bear. For people with disabilities, smart home speakers are a godsend. At TechShare Pro, a U.K.-based conference focused on AI, disability, and inclusive design held in November 2017, Ellie Southwood, chair at The Royal National Institute of the Blind, commented that the Echo Dot made her feel included as she spent less time searching for things online. She could also multitask while online.
One hobbyist incorporated a Raspberry Pi development board and the Alexa Voice Service, Amazon’s third-party speech recognition platform, to place voice controls to a motorized wheelchair. These are two notable examples of how such technologies can change the lives of people with disabilities. Smart home speakers become more powerful when paired with smart home appliances. This means people with sight loss and physical ailments can switch on lights and adjust the temperature using their voice.
Speech-to-Text and Text-to-Speech
Voiceitt, an app for people with speech impediments, learns speakers’ pronunciations over time. The app then normalizes abnormalities in exportable audio and text. On the other hand, Google’s DeepMind division leverages AI to generate closed captions for deaf users.
DeepMind’s AI algorithm watched more than 5,000 hours of television and analyzed 17,500 unique words, per a 2016 joint study with researchers at the University of Oxford. The model outperformed a professional lip-reader as it translated 46.8% of words without error in 200 randomly selected clips. The professional only managed to translate 12.4%.
“Voice recognition was developed for disabled people, but it’s the hot item at the moment and is useful for everyone,” Kiran Kaja, technical program manager for search accessibility at Google, said at the 2017 TechShare Pro.
Automatic Image Recognition
Screen-reading programs aid blind and vision-impaired people in navigating websites. Most websites contain images and not every photo has an appropriate alt text or title. But AI is capable of classifying images automatically.
To illustrate, Facebook developed captioning tools that describe images to visually impaired users. Google’s Cloud Vision API can understand the photos’ context, so it might label a picture of a jack-o’-lantern with “pumpkin,” “Halloween,” and the like.
|Smart home speakers like Google Home, Amazon’s Echo series, and Apple’s HomePod is not just for people who want to queue up their favorite podcasts without moving away from their couch / Photo by: goodluz via 123RF|
Is There Any Way to Tackle AI Bias?
Alex Engler of non-profit public policy organization Brookings Institution argued that AI training data should include many people with diverse disabilities. It’s possible for an AI software to include a diverse range of individuals, but it will take effort. After all, disability is not just a simple concept “with a small number of possible values.” It has many dimensions, and disability also varies in intensity and impact, which usually changes over time.
Disability also depends on the context. For instance, the issues faced by a visually impaired person navigating around a city contrasts that of someone in a wheelchair. This means the data used to describe an individual with a disability may look unique.
People with disabilities know that revealing a disability can be risky. Engler listened to a group of students discussing the advantages and disadvantages of disclosing their disabilities. One chose not to disclose their disability because they thought it would reduce their chances of landing an internship. Another revealed his disability so that his needs could be accommodated for the application process.
Disability information is sensitive. The data used to train AI models does not always contain information “but may reflect the presence of a disability.” Some people with disabilities want to contribute their own data set to help test or train AI systems, but it comes with personal risk. This risk makes them “re-identifiable” due to the “unusual nature” of an individual’s situation even if the data is anonymized.
Include People With Disabilities in the Development Process
They should be part of the development process to help developers identify who the outliers might be as well as the people who might be affected by the solutions they are formulating. For instance, a voice-controlled service might not work well with people with speech impairment or deaf speakers since their voices are not well understood.
Additionally, an online assessment test based on test times might be unfair to individuals who use assistive technologies to access it. Therefore, it is best for developers to seek out and work with the affected stakeholders on creating a fair and equitable system.
Since people with disabilities represent the world’s largest minority group, we have to be more inclusive when training AI models and developing new products. We should collaborate with people with disabilities to ensure that their needs are also being catered to by new technologies.
|Since people with disabilities represent the world’s largest minority group, we have to be more inclusive when training AI models and developing new products. We should collaborate with people with disabilities to ensure that their needs are also being catered to by new technologies / Photo by: Fabio Formaggio via 123RF|