In 2022, artificial intelligence (or A.I., for short) is more than a buzzword. Consumer-friendly A.I. tools have created innovative ways to generate content and analyze data — along with reasonable concerns about how the tools are designed and programmed.
But for those in the digital accessibility space, the potential benefits of A.I. greatly outweigh the risks. At the Bureau of Internet Accessibility, we’ve used A.I. as part of our automated audits for years.
We’ve also seen A.I. and machine learning simplify the hard work of digital accessibility, and over the next decade, we expect that A.I. will play an important role in improving online spaces for people with disabilities.
Of course, manual testing and remediation will always be essential for creating inclusive content. But by performing certain tasks quickly, inexpensively, and with human-like accuracy, A.I. tools have already made an impact.
Below, we’ll explore several ways that artificial intelligence can help organizations provide a better experience for users with disabilities. We’ll also discuss some of the limitations of A.I. — and provide a roadmap for creating better content for all users.
1. Automatic captions aren’t perfect, but they make captioning easier
Captions are essential for accessibility. Without accurate captions, your media isn’t useful for many people with hearing disabilities — or people who simply prefer to browse the web without sound.
The Web Content Accessibility Guidelines (WCAG), widely considered the international standard for digital accessibility, requires captions for all pre-recorded multimedia. Implementing captions isn’t especially difficult, particularly if you draft captions while writing your video scripts.
For longer videos and live media, this might not be a realistic option. Automatic captioning tools convert speech to text, and while they’re not perfect, they’re getting better: Machine learning models can limit errors by cross-referencing word pronunciations in their training data.
For WCAG conformance, captions must be 100% accurate — automatic captions simply aren’t there yet. YouTube’s automatic captions are roughly 60-70% accurate, and creators need to carefully review the output.
But eventually, A.I. models will be able to recognize intent, which will address some of the current issues with speech-to-text generation (such as recognizing different homonyms or non-speech sounds).
2. With human oversight, A.I. tools can improve code and markup
Many people use assistive technology (AT) such as screen readers to access the internet. These tools rely on accurate markup. Each web element needs to be properly identified with semantic HTML or WAI-ARIA (Web Accessibility Initiative - Accessible Rich Internet Applications).
With proper markup, AT can present the content to users in a way that makes sense. For example, if a user encounters a web form, a screen reader can tell the user that they’re viewing a form and identify the info needed for each form field.
A.I. tools like the OpenAI Codex and ChatGPT are designed to convert natural language to code, simplifying the work. This could be especially powerful when designing complex web applications — but developers still need a solid understanding of the basics in order to review the output.
A.I. models are prone to hallucination, and in our experiments with OpenAI, we found a number of inaccuracies, particularly when asking the model to generate WAI-ARIA.
Even so, A.I. models can limit work for talented developers, which removes a major roadblock for web accessibility: Many organizations believe that they don’t have the resources to create content in an accessible way.
When code and markup can be generated instantly, then edited to follow the best practices of accessible design, there’s no excuse for ignoring the 1.3 billion people worldwide who live with disabilities.
3. Image recognition tools may improve experiences for users with visual impairments
Missing image alternative text (also called alt text or alt tags) is a common accessibility issue. WCAG requires alternative text for all non-text content, as alternative text enables users to understand the purpose of visual content without viewing that content.
Writing alt text isn’t especially difficult. However, if you have a large website with thousands of images, creating individual descriptions for every single image may be extremely time consuming (another reason to prioritize accessibility from the first stages of product design).
At this time, computers aren’t great at accurately identifying images. In fact, many tests intended to distinguish human users from computers use image recognition — think of all of those “select all images containing a traffic light" prompts you encounter when browsing the web.
That’s already changing. Products like Microsoft’s Computer Vision use A.I. models to identify image content, extract text, and moderate user-submitted images.
As with captions and markup, A.I. image descriptions need to be 100% accurate in order to be useful. But as A.I. models gain access to a wide range of training data, accuracy will improve.
Eventually, image descriptions may be applied automatically, greatly improving experiences for people with vision disabilities.
Understanding the Limits of A.I. for Web Accessibility
A.I. tools may make web accessibility easier, but there are a few things that they can’t do: They can’t make judgment calls, they can’t provide 100% perfect output, and they can’t truly understand the experiences of users with disabilities.
As machine learning models improve, they may become indispensable resources. They might greatly reduce the work of accessibility, but they’ll always require human oversight. To create products for real people, you need to think about your users at every stage of the process.
Software tools can help — but with the number of accessibility lawsuits rising each year, you shouldn’t wait to adopt an accessibility-first approach. The benefits of accessible design greatly outweigh the costs, and by prioritizing users with disabilities, your organization will be in a great position to take full advantage of the A.I. revolution.