Digital Accessibility Research with Artificial Intelligence

I am experimenting a lot with ChatGPT and other Gn AI's. You can use it for useful things, for example for research into digital accessibility. ChatGPT stands for generative technologies here, you can also use Gemini (in my experience, not very reliable so far) or Perplexity.ai or another generative technology of your trust.

In general, two scenarios can be distinguished:

Finding the right criteria and requirements for specific problems
Having complicated requirements explained to you

Finding suitable criteria and requirements

For simple elements such as select boxes, it is quite easy to find the relevant information. I mostly use the Mozilla Developer Network, but there are numerous other reliable sources such as Stack Overflow or the W3 School. Things get difficult for complex elements like multiselect dropdowns or similar composite elements for which there is no HTML equivalent or simple ARIA. A simple Google search sometimes doesn't help because it often shows few to no or outdated articles.

So you ask ChatGPT or your trusted GenAI: "What do I have to pay attention to if I want to make multiselect dropdowns accessible?". Please note that at least the free version of ChatGPT is currently frozen at the beginning of 2022 and therefore cannot, for example, take into account any requirements from the final WCAG 2.2 from November 2023.

I'll spare you the answer from ChatGPT, you can try it out for yourself. The practical thing about this is that an example code is also shown. Please do not simply copy the code, as it is generic and contains a lot of DIVs.

In the second step, you can ask another GenAI, for example Gemini from Google or perplexity.ai. You should compare the answers. If the answers are the same, we are on the right track. Some GenAIs also provide sources for the results that you can study. This is what for eample the Bing copilot does.

In the third step, I would research specialist articles on the topic and compare the answers from the GenAIs with them. Make sure that the articles are up-to-date, come from reputable sources and address the WCAG criteria used.

Have we really saved time by taking all these steps, stupid? Not necessarily, but the chance of getting correct answers and implementing the element correctly has increased. The alternative would have been to take the first article from Google and use it. As a novice, however, I cannot know to what extent this article is correct or up-to-date.

GenAIs can of course also be used to have articles summarized. The shorter the article, the better the summary. In my experience, from around 6000 characters, the summary becomes very rudimentary. Here, too, you can of course cheat by copying a certain subset of the text into the chat window; in my experience, 3000 characters should work well.

Another option that I like to use for preliminary research is to ask ChatGPT about the WCAG criteria that apply to a certain element. In my experience, this is also quite reliable, although, as I said, you should always check the answers. Unfortunately, it does not currently work for EN 301549; here, too, only the WCAG criteria are mentioned, but not the requirements from other chapters.

As always with GenAIs, specific questions are better answered than generic ones. For example, do not ask: "What should I consider in barrier-free design?" Even a human could hardly meaningfully answer this question. Instead, ask: What should I consider when using colors in accessible design?" The conversation can be further deepened with the answers from ChatGPT if you need more information on individual aspects. Or you can have suitable articles recommended to you straight away. Perplexity.ai works quite well for this purpose.

The GenAIs' talkativeness is relatively annoying, which usually comes to the fore in the first answer. If you ask about colors in accessibility, you probably don't need to know why accessible web design is important.

Getting rules explained

ChatGPT can also help to get complicated rules explained. There are rules that have rarely been used, such as 2.5.4 Motion Actuation. ChatGPT can help here, for example by asking for an explanation of this rule and perhaps asking for examples.

As mentioned above, ChatGPT may not know the current rules from EN 301549. ChatGPT can help with this and other regulations or laws by simply copying in the text of the relevant paragraph and asking for a clear explanation. However, you have to rely on the explanation sounding plausible, as there is hardly any in-depth material on the European Accessibility Act or the EN so far, at least I have not come across anything that is not very superficial. In my view, a ChatGPT for accessibility would be very desirable.

Internal and limited resources

The commercial version of GenAIs often allows two things that can be interesting for internal use. First, there is the possibility of accessing internal resources. Let's say you have an internal wiki and a library, then the GenAI could be set to give specific answers for the respective organization. This is important for large organizations with complex designs or multiple libraries. For the A11Y professionals, this may not be relevant because they know these documents or may even have created them themselves. But we also want to make the work easier for other people or perhaps even external service providers.

But it may also be worth considering limiting the external resources used for retrieval. For example, if you say you only want to use reputable or freely accessible resources for the GenAI, this is also possible. The index can be expanded or reduced as needed.

Generate user stories

Another way to use ChatGPT is to generate user stories. User stories are a method in agile project management to describe a feature in more detail. For example, you can generate a story about what could happen if a user encounters a certain problem: for example, a missing form label.

Notebook LM as your own mini-GPT

Notebook LM is an application from Google that you can use to search through information you have compiled yourself. You first add up to 50 links, documents or notes and can then query them in a structured manner similar to ChatGPT. The risk of the GPT fantasizing is lower here, as it mainly draws on the resources provided. For example, I added the BITV, the EN 301549, the German BFSG, the German BFSG regulation, the German Disability Equality Act and the WCAG and thus have a pretty good overview of all the regulations and standards in Germany.

Getting error reports explained

Professionals try to write their error reports in such a way that they can be understand. This does not always work. If you encounter problems in an error report, you can simply copy the text into ChatGPT and ask for an explanation and possible solution.

This should also work with automatic checking tools such as the PDF Accessibility Checker, WAVE and similar tools. Unfortunately, there are relatively few resources other than for the web from which the GenAIs could be trained. Therefore, the answers to apps and PDFs may not be as good as for questions about the web.