On AI And ‘the people it often misses’
To mark Global Accessibility Awareness Day this week, Microsoft published a blog post wherein the company addresses the dire need for artificial intelligence models to properly account for people with disabilities. The post was written by Susanna Ray.
“Ask an AI tool for a picture of ‘someone at work,’ and you’ll often get a person at a desk, in front of a computer, maybe holding a coffee cup. When people with disabilities appear at all, the images sometimes have been jarringly wrong: amputees with extra limbs, blindness portrayed with blindfolds, those with dwarfism sporting huge, pointy ears,” she wrote on Thursday. “The problem isn’t intent; it’s absence. AI systems can only learn from the data they’re trained on and the criteria used to judge their outputs. For many groups—including the estimated 1.3 billion people living with disabilities—there hasn’t been enough focus on representation online to shape either.”
Ray goes on to mention Microsoft is working alongside numerous organizations, including Kilimanjaro Blind Trust Africa, based in Nairobi, to ensure “AI-generated imagery reflects how people with vision disabilities actually live and work.” In addition, the company also is closely working with other communities in an effort to “build libraries of images that depict them in a more accurate and representative way,” adding “those groups can offer the datasets to tech companies and others for training and testing AI systems, a process that typically requires thousands of examples.”
Most poignantly, and the crux, Ray writes “research in media and social science has long shown that how groups of people are portrayed shapes how society thinks about them, with direct influence on access to things like education and employment.”
Microsoft’s post is well worth a read in its entirety. As a person with disabilities, I’m mostly bullish on AI’s potential as an assistive technology from a pragmatic perspective. But it is equally important that AI agents be trained with the most comprehensive dataset possible—which entails accurately and authentically portraying the disability community. These systems aren’t sentient beings; they must be fed information, so their “aptitude” is only as good as the information they’re given. Likewise, while something like ChatGPT is effective in helping, say, someone code or write an email, its appeal—and its utility—is somewhat stunted if, for instance, it never shows a wheelchair user in a workplace environment doing their job like anyone else.
Reading Ray’s post had me drawing parallels to disability coverage in mainstream media. For the breathlessness with which tech news outlets have reporters cover AI, nary a word is said for the technology’s impact on the disability community. Likewise, nary a word is said for disability-in-tech coverage in general—stuff like AI is seen as more important, more traffic-worthy, with accessibility news only mattering during token events like Global Accessibility Awareness Day or themed “accessibility weeks.” While these special times on editorial calendars are a good idea in a vacuum, it’s the journalistic equivalent of shuttering the holiday lights in storage the first 11 months of the year until after Thanksgiving when it’s Christmastime—then they will go up.
Newsrooms, and AI models, must be better at dignifying people like me.
