Opportunities for AI in Accessibility

I was completely moved by Joe Dolson’s subsequent article on the crossing of AI and availability because I found it to be both skeptical about how widespread use of AI is. In fact, I’m very skeptical of AI myself, despite my role at Microsoft as an accessibility technology strategist who helps manage the AI for Accessibility award program. AI can be used in quite productive, inclusive, and accessible ways, as well as in harmful, exclusive, and harmful ways, like with any tool. And there are a lot of uses for the poor center as well.

I’d like you to consider this a “yes … and” piece to complement Joe’s post. Instead of refuting everything he’s saying, I’m pointing out some areas where AI may make real, positive impacts on people with disabilities. To be clear, I’m not saying that there aren’t real challenges or pressing problems with AI that need to be addressed; there are, and we’ve needed to address them, like, yesterday; instead, I want to take a moment to talk about what’s possible so that we can get it one day.

Other words

Joe’s article spends a lot of time examining how computer vision versions can create other words. He raises a number of true points about the state of affairs right now. And while computer-vision concepts continue to improve in the quality and complexity of information in their information, their benefits aren’t wonderful. He argues to be accurate that the state of image research is currently very poor, especially for some graphic types, in large part due to the lack of context-based analysis that exists in the AI systems ( which is a result of having separate “foundation” models for text analysis and image analysis ). Today’s models aren’t trained to distinguish between images that are contextually relevant ( which should probably have descriptions ) and those that are purely decorative ( which might not even need a description ) either. However, I still think there’s possible in this area.

As Joe points out, alt text editing via human-in-the-loop should be a given. And if AI can intervene and provide a starting point for alt text, even if the quick reads,” What is this BS?” That’s certainly correct at all … Let me try to offer a starting point— I think that’s a win.

If we can specifically station a design to examine image usage in context, this may help us more quickly determine which images are likely to be elegant and which ones are likely to be descriptive. That will help clarify which situations require image descriptions, and it will increase authors ‘ effectiveness in making their sites more visible.

While complex images—like graphs and charts—are challenging to describe in any sort of succinct way ( even for humans ), the image example shared in the GPT4 announcement points to an interesting opportunity as well. Let’s say you came across a map that was simply the name of the table and the type of visualization it was: Pie table comparing smartphone use to have phone use among US households making under$ 30, 000 annually. ( That would be a pretty bad alt text for a chart because it would frequently leave many unanswered questions about the data, but let’s just assume that that was the description in place. ) If your browser knew that that image was a pie chart ( because an onboard model concluded this ), imagine a world where users could ask questions like these about the graphic:

Are there more smartphone users than feature phones?
How many more?
Is there a group of people that don’t fall into either of these buckets?
What number is that?

For a moment, the chance to learn more about images and data in this way could be revolutionary for people with low vision and blindness as well as for those with various forms of color blindness, cognitive disabilities, and other issues. It could also be useful in educational contexts to help people who can see these charts, as is, to understand the data in the charts.

What if you could ask your browser to make a complicated chart simpler? What if you asked it to separate a single line from a line graph? What if you could ask your browser to transpose the colors of the different lines to work better for form of color blindness you have? What if you asked it to switch colors in favor of patterns? That seems like a possibility given the chat-based interfaces and our current ability to manipulate images in today’s AI tools.

Now imagine a purpose-built model that could extract the information from that chart and convert it to another format. Perhaps it could convert that pie chart (or, better yet, a series of pie charts ) into more usable ( and useful ) formats, like spreadsheets, for instance. That would be incredible!

Matching algorithms

When Safiya Umoja Noble chose to call her book Algorithms of Oppression, she hit the nail on the head. Although her book focused on how search engines can foster racism, I believe it’s equally true that all computer models have the potential to foster conflict, prejudice, and intolerance. Whether it’s Twitter always showing you the latest tweet from a bored billionaire, YouTube sending us into a Q-hole, or Instagram warping our ideas of what natural bodies look like, we know that poorly authored and maintained algorithms are incredibly harmful. A large portion of this is a result of a lack of diversity in the people who design and construct them. However, when these platforms are built with inclusive features in mind, there is real potential for algorithm development to help people with disabilities.

Take Mentra, for example. They serve as a network of employment for people who are neurodivers. They match job seekers with potential employers using an algorithm based on more than 75 data points. On the job-seeker side of things, it considers each candidate’s strengths, their necessary and preferred workplace accommodations, environmental sensitivities, and so on. On the employer side, it takes into account each work environment, communication issues relating to each job, and other factors. Mentra made the decision to change the script when it came to the typical employment websites because it was run by neurodivergent people. They use their algorithm to propose available candidates to companies, who can then connect with job seekers that they are interested in, reducing the emotional and physical labor on the job-seeker side of things.

More people with disabilities can be used to create algorithms, which can lessen the likelihood that they will harm their communities. Diverse teams are crucial because of this.

Imagine that a social media company’s recommendation engine was tuned to analyze who you’re following and if it was tuned to prioritize follow recommendations for people who talked about similar things but who were different in some key ways from your existing sphere of influence. For instance, if you followed a group of nondisabled white male academics who spoke about AI, it might be advisable to follow those who are disabled, aren’t white, or aren’t men who also speak about AI. If you followed its advice, you might be able to understand what is happening in the AI field more fully and nuancedly. These same systems should also use their understanding of biases about particular communities—including, for instance, the disability community—to make sure that they aren’t recommending any of their users follow accounts that perpetuate biases against (or, worse, spewing hate toward ) those groups.

Other ways that AI can assist people with disabilities

I’m sure I could go on and on about using AI to assist people with disabilities, but I’m going to make this last section into a bit of a lightning round. In no particular order:

voice recognition is. Researchers like those in the Speech Accessibility Project are paying people with disabilities for their help in collecting recordings of people with atypical speech. As I type, they are actively seeking out people who have Parkinson’s and related conditions, and they intend to expand this list as the project develops. More people with disabilities will be able to use voice assistants, dictation software, and voice-response services as a result of this research, which will lead to more inclusive data sets that enable them to use their computers and other devices more effectively and with just their voices.
Text transformation. The most recent generation of LLMs is capable of altering already-existing text without giving off hallucinations. This is incredibly empowering for those who have cognitive disabilities and who may benefit from text summaries or simplified versions, or even text that has been prepared for bionic reading.

The importance of diverse teams and data

We must acknowledge that our differences matter. The intersections of the identities we exist in have an impact on our lived experiences. These lived experiences—with all their complexities ( and joys and pain ) —are valuable inputs to the software, services, and societies that we shape. Our differences must be reflected in the data we use to develop new models, and those who provide it need to be compensated for doing so. Inclusive data sets produce stronger models that promote more justifiable outcomes.

Want a model that doesn’t demean or patronize or objectify people with disabilities? Make sure that you include information about disabilities that has been written by people with a variety of disabilities in the training data.

Want a non-binary language model? You may be able to use existing data sets to build a filter that can intercept and remediate ableist language before it reaches readers. Despite this, AI models won’t be replacing human copy editors anytime soon when it comes to sensitivity reading.

Want a coding copilot who can provide you with useful recommendations after the jump? Train it on code that you know to be accessible.

I have no doubts about how dangerous AI can and will be for people today, tomorrow, and for the rest of the world. However, I also think that we can acknowledge this and make thoughtful, thoughtful, and intentional changes in our approaches to AI that will reduce harm over time as well. Today, tomorrow, and well into the future.

Many thanks to Kartik Sawhney for supporting the development of this article, Ashley Bischoff for providing me with invaluable editorial support, and of course, Joe Dolson for the prompt.