An Introduction to Accessibility Compliance and AI Assistance

Accessibility is a complex field that requires knowledge of all physical and psychological disabilities, as well as a working understanding of various national laws. As various compliance deadlines approach, such as the EAA and Title II of the ADA, companies need to adhere to accessibility standards. Many turn to an AI solution to get up to speed. While AI is powerful and can help with tedious and time-consuming tasks, it works best in tandem with human review. This blog will explore the types of accessibility measures a company can take, how AI can help or hinder their implementation, and the importance of human review for AI-generated material.

Alt Text: What It Is, and How Can AI Help or Hinder?

Alt text is a common component of accessibility. It is a short piece of writing, typically 150 characters or fewer, that is included with an image’s HTML code on a website. Most users without low vision would only notice the presence of alt text when an image fails to load on a webpage.

However, the alt text is always there. Screen readers, assistive devices that can read out a webpage for people with low vision or for those who are blind, will read out the alt text if it’s included in the image’s HTML. Alt text is not inherently present in an image; if it isn’t included, screen readers will read out the file name of the image. This can be very disorienting, and outright annoying, to users with low vision. Web developers may turn to AI to write the alt text for them, but they should be aware that it won’t do everything for them.

Many AIs are able to parse images and output descriptions, much faster than what a person could reasonably do. However, it can’t determine what is actually important in that image; it can only predict what’s important based on its training data. This prediction can be correct, but it isn’t guaranteed to be so.

Let’s use an example. Imagine you work at a publishing company. You are tasked with creating alt text for an online textbook about food science. Accessibility work is more time-consuming than you anticipated. To alleviate the workload, you feed all of the images into an AI, which outputs a description, and copy those descriptions into the alt text, without review.

Cons of AI-Generated Alt Texts

The problems with AI-generated alt text:

1. Not All Images Need Alt Text

Many images do, especially if they convey information not in the text of the page. However, images that are non-essential and decorative need no alt text. An example of a non-essential image could be a geometric pattern on a gradient background, where the image is purely decorative and not needed to understand the point of the passage. When working with decorative images, the alt text attribute needs to be defined as blank in the HTML code. This way, non-essential images can be skipped by the screen reader, rather than have it read out the image’s filename.

2. The AI Has No Context for How the Images Will Be Used

If one of the images is a woman in a kitchen using a mixing bowl, but the AI was trained on hair and makeup photography, it will struggle to describe the kitchenware at all. It may output something like, “A woman with a medium-bob haircut in a kitchen,” whereas, given the context, a better alt text might read “A woman using a mixing bowl to prepare food.”

An AI can output technically correct alt text very quickly, faster than what a person can write, but it will not understand why it’s doing it; a good alt text writer will. The alt text may be usable, but it won’t be guaranteed to be accessible. The AI you use must be trained on the type of work you will be doing; a committed edtech partner can help you choose the best AI for your needs.

Transcripts and Captions: What They Are, and How AI Can Help or Hinder

Another use of AI in accessibility is for captions and transcripts. Both are written versions of video content, including both spoken words and non-speech sounds, so they can be experienced by people with hearing disabilities.

Captions are incorporated into and synchronized with a video file itself, typically using an SRT file, while transcripts are text documents included alongside a video. Many people use subtitles on streaming services even when they aren’t hard of hearing. Or perhaps you need to watch a video, but your internet connection is too slow; a transcript included alongside the video would be very helpful. AI can assist in creating captions and transcripts by taking in an audio file and putting out text. Although this process helps optimize the workload of web developers, it won’t do everything.

Let’s go back to our hypothetical about the food science online textbook. You’ve moved on from writing alt text, after having to quickly re-do a good portion of work because you didn’t check the output of the AI you used.  Now you’re tasked with creating captions for some video modules. These, too, are time-consuming tasks. Because you had to redo the alt text, you’re behind schedule, so you feed the videos into a transcription AI and get out a lot of transcriptions very quickly. You copy and paste the generated text into a large number of SRT files without reviewing them. This approach, too, has a few issues.

First, AI transcription does not understand the finer points of grammar. Most AIs will handle periods, commas, and capitalization fairly well, but many AIs struggle with other types of punctuation. For example, quotation marks are both common and important for clarity. Quotation marks are also inaudible; you cannot hear when a quotation mark is used. So when an AI transcribes a video that happens to include a lot of discussions about short stories or people directly quoting others, all of these necessary quotation marks will be absent. As with alt-text, context is key, and a person is going to understand the grammatical context of speech more than an AI trained purely on transcription.

Second, much like people, AI can also “mishear.” When talking to one another, people can gloss over minor mispronunciations because we understand what the conversation is about. Even if we don’t understand what’s being said, we can always ask questions. An AI cannot do this; it will analyze the waveforms it is given and make a guess at what the word is based on its training data.

AI has an especially difficult time transcribing words that aren’t pronounced how it expects them to be pronounced. For example, if a speaker in a video has a cleft palate, which may affect certain sounds, the AI tool will incorrectly transcribe what they’re saying. Given that to train an AI tool, it must be graded against a rubric, its training data must be clear and precise, with no room for uncertainty.

Going back to our culinary learning module example, if a video is discussing how to use a colander, an AI transcription might incorrectly identify the phrase “sleek colander” as “Zeke calendar.”

If your attempt at accessibility isn’t correct, then it isn’t accessible.

Why Context Is Key in AI for Accessibility

Although AI tools can help alleviate much of the tedious work of accessibility compliance, they’re no replacement for human review. To a user with low vision, inaccurate alt text is no better than no alt text, and poorly written captions are sometimes worse than no captions at entirely. If your AI-generated accessibility measures contain errors, then it’s ultimately still not accessible.

Therefore, manual proofreading of AI output is essential. In the same way that a car’s cruise control doesn’t prevent you from hitting the gas or the brakes, it’s important to remember that AI isn’t a replacement for your responsibilities; it’s a supplement. You’re the one in the driver’s seat, so to speak, and the tools you have at your disposal, AI or otherwise, aren’t going to drive themselves.

Context is key to accessibility compliance. A person tasked with determining and figuring out the context will do a better job of it than an AI. After all, an AI is designed and programmed for a very specific task, and it has no knowledge of why it is doing something if it hasn’t been told so. People, on the other hand, are not so rigid in their knowledge; our “training data” is the sum total of every experience we’ve ever had, everything we’ve ever learned. People can’t help to put things into context; context is how we know what we know.

Whether you work in the accessibility field or you simply want to sell a product to a wide audience, working in tandem with AI, and not just using it as a one-size-fits-all solution, is the key to making truly accessible products.