AI and Accessibility

by

In the past few years, and more so, the last few months, AI has matured into a vital tool and critical role in our lives, affecting the way we learn, work, interact with each other, and even relax. It is powerful and yet still in its infancy. The possibilities of how AI can assist and improve our lives are boundless. One way AI is improving life is through accessibility. Although fantastic strides in digital accessibility were seen during and directly after the Covid-19 pandemic, technology is still far from inclusive. Many websites, platforms, and software still require the use of a mouse for navigation, incorrect information is often communicated to screen readers, especially when math and science are involved, and although alternative description for images is now the norm, it is often incorrect or insufficient. AI is working to address and remove these barriers and many more, leveling the playing field and creating equality for people of all different abilities. However, we must remember AI’s use in accessibility is still immature, and although it has an unlimited list of accessibility issues it can address in the future, it currently creates some of its own accessibility issues such as AI systems not being accessible themselves, incorrect information and alternative descriptions, too much textual information, illogical organization of content, writing inaccessible code, and more. Due to these accessibility issues with AI, we must not forget that AI is just a tool. It can help us greatly with accessibility, but we must still be vigilant at ensuring what we present is fully accessible ourselves. Below you will find more information on the pros and cons of AI and accessibility in addition to some tips to help ensure your AI content is accessible.

Accessibility and AI Pros

Accessibility Testing

  • AI can facilitate and automate the accessible design process, in addition to ensuring updates and changes are accessible.
  • AI can automate and accelerate the accessibility testing process.

Captions and Audio Description

  • With AI speech recognition and natural language processing, video and audio captions can be completed quicker and with greater accuracy than ever before.

Facial Recognition

  • Passwords, CAPTCHA, and other authenticating methods have presented barriers in accessibility; however, AI’s real-time ability to recognize identity, emotions, and expressions has improved both accessibility and the authentication process for most users.

Image Recognition

  • AI can recognize objects, scenes, photographs, and even text within images, greatly improving user experience for those with visual impairments.

Lip-reading Recognition

  • AI’s lip-reading abilities are improving the accuracy of live captions greatly. In fact, AI’s lip-reading abilities are over seven times greater than that of humans. In 2016, Oxford University put human lip-readers up against AI. Humans had an average error rate of 47.7%, while AI’s error rate was 6.6%!

Navigation

  • AI is helping with navigation both in the built world and in the digital world. Technology can assist a digital user on which sites are accessible, while AI navigation systems can provide users real-time guidance, suggest accessible routes, and offer accessibility information on facilities in the physical world.

Personalization

  • AI powered websites can adapt their design and functionality to the user’s specific set of needs. For instance, it could increase font size and contrast for a user with low-vision or automatically turn on captions or open transcripts for a user with a hearing impairment.

Real-time, Automated Translations

  • AI has improved the accuracy of real-time translations greatly. These automated translations can be as simple as reading text on a page to translating between two languages, and as complicated as converting written text to sign language (Hand Talk Plug In).

Text Summarization

  • Summarizing large chunks of information into abstract-sized bites can be very beneficial for those with cognitive impairments such as dyslexia and ADHD, or those with memory issues, low-literacy abilities, non-native speakers or even those with limited time, emphasizing the importance and benefit of accessibility to everyone.
  • Previously, AI’s summarization abilities were extractive and only based on the text available, but now, Salesforce has designed AI that is abstractive, using text not in the original content, creating more concise, meaningful abstracts.

Voice Recognition

  • AI’s voice recognition technology allows those with physical and mobility disabilities to interact and control digital technology and other tools with only their voice.

Accessibility and AI Cons

Accuracy and Alternative Text

  • Whether we’re dealing with accessibility or simply content creation, AI can and does make mistakes.
  • Although AI is getting better at generating alt text for images, it is not always accurate and often doesn’t provide descriptions in the proper context. For example, in 2023, the Bureau of Internet Accessibility conducted a test using ChatGPT: “We asked ChatGPT to write alternative text for an image of an apple. The text read: “A red apple with a stem and a leaf on top, against a white background.” That’s decent alternative text — but the model added details (such as the white background, stem, and leaf) that we didn’t provide.”

Data Bias

  • AI learns from the data it is trained on. Inferences from data often rely on averages which reflect the most common entity not the most representative nor inclusive.
  • AI algorithms are only as good as the data they are trained on. If an AI algorithm for captions is created in New York by a native New Yorker, only using data with local accents and colloquialisms, it won’t understand the accents and terms being used at an event in Alabama which is in need of live captions.
  • AI can misidentify labels if the data fails to provide all possibilities. For example, if AI is provided with images of postal carriers who happen to all be of one gender, AI may assume that all postal carriers are only of this gender.

Dependence and Trust

  • Some users may become overly-dependent on AI technology, putting them at risk if the AI technology goes down or has issues.
  • A 2017 study conducted at Stanford University showed those with visual impairment overtrust AI generated alternative text. In the study, a black and white photo of Hillary Clinton walking onto a stage with a crowd in the background had been captioned by AI as “I am not really confident, but I think it’s a man doing a trick on a skateboard at night.” The majority of those who were blind were fairly confident that the photograph was a man on a skateboard. This study amplifies why accurate alternative text, not simply alternative text is so very important.

Hyperlinks / Credibility

  • AI may link to sites that are not credible as references or present these links in their entirety, instead self-contained, live links.

Inaccessible Code and Content

  • The Bureau of Internet Accessibility in testing ChatGPT found accessibility issues in both the HTML code (subheading tags not nested properly) and the ARIA code (labeled something “true” to “false.”)
  • Just as AI can confidently present incorrect information as correct, it can present inaccessible content as accessible.

Lack of Awareness

  • Technology developers may be unaware of all their users’ needs, thus, excluding those with disabilities due to the lack of accessible features. Others simply do not have the funds to include accessible AI technology in their products.

Lengthy Sections of Text

  • Often AI can present the user with large blocks of text. Not only is this less than exciting for most users, but it can present issues for users with cognitive, attention, and visual impairments.

Organization of Content

  • In addition to often tagging headings out of order, AI can present information in an order that is not well suited to your audience. Reorganization in a logical, easy to consume layout is often needed.

Privacy

  • Those with disabilities, due to needing AI features such as facial and voice recognition, may be exposed to privacy violations at a greater percentage than the average user.

Tips for Making Your AI Content Accessible

  • Don’t assume your AI content is accessible because it tells you it is! Always check your content.
  • Read the generated alternative text. Is it in the correct context for how you are using the image? Is it extraneous? If so, shorten it.
  • Check your headings and subheadings. Are they labeled or tagged? Are they nested properly?
  • Are bulleted lists, tables, graphs, etc tagged as such?
  • Is your content broken up into easy to digest blocks?
  • Break content up further by adding images, videos, bulleted lists, etc, but make sure you are adding accessible content.
  • Check color contrast with an analyzer such as the Paciello Group’s Color Contast Analyzer.
  • Ensure videos and audio have captions and/or transcripts available.
  • If you used AI to write code, check the work with a screen reader, such as NVDA.