The application of machine learning to accessibility is not new, but as AI explodes in popularity, it’s certainly getting more attention.
Auto-craptions are an application of machine learning, and the choice of term I made suggests something about what accessibility practioners generally think of AI in accessibility.
But that’s not the whole picture. The problems I see with AI and accessibility are only when it’s used directly on the end product. This is the same as the problems with AI in any field. If you only look at examples where a lawyer uses AI to generate their court brief or an overlay that uses AI (or at least markets their product as using AI), then you’re finding some of the most questionable situations.
Alternative Text and AI
I’ve recently seen a lot of new products applying AI to image alternative text, which I think is an area that requires a great deal of caution. MasterWP just released their product, EveryAlt, which analyzes images in your media library for alt text. It seems like a reasonable methodology: from reading the info, it only directly adds information in the media library, and it doesn’t override existing alt text unless you tell it to. 10up’s ClassifAI does a similar job.
I’ve written on this before, and I haven’t changed my opinions much since then, so you should possibly just divert to my previous automated alt text article for a moment. I’ll still be here when you get back.
That article is all about the kinds of things that AI misses when it creates alternative text. But that isn’t to say that there isn’t any benefit to automatically sourcing image descriptions. I’m using a different word here on purpose, however: automation can generate a description of an image. That’s not an alt attribute. An alt attribute is in reference to the specific usage and context of an image, and about the image itself only as it serves a purpose.
Let’s go back to talk about auto-craptions again. They are widely considered awful. But this isn’t because we shouldn’t use automation to generate captions; it’s because that generation is only the first step in the process. Automatic captions need to be edited.
What AI does for accessibility
Which brings me back around to what power AI does have for accessibility. It can be used to perform analysis or generate content, but it should not be the end product.
There’s an open ticket on WordPress core for revamping how the editor and media library handle alt attributes. It’s stalled, because it’s a pretty complicated problem to assess and resolve. One of the arguments is that WordPress should not retain a fixed alt attribute, because attributes should be set to purpose at the time of use. And I just made that argument above, so I have to give that some credit.
But sometimes you do want the same information. And sometimes the person who added the image knew things that the person using the image doesn’t know – like who is in the image, when the event occurred, or why this image is important.
And sometimes the person using the image can’t see it.
So the media library needs to store an image description; but this shouldn’t automatically be considered the alt attribute. AI could absolutely be used to generate an image description, which can be available for reference when writing alt descriptions.
Other uses of AI for Accessibility?
What are other hypothetical uses of AI for accessibility? It’s an interesting thought experiment to imagine. I think it might be best used in assistive technology, where it can learn with a user. Learning what the user wants and expects, optimizing for their preferred tools, and learn their needs.
It can also become part of a development toolkit. If you could ask a bot to build you a form and it defaulted to accessible HTML (HyperText Markup Language) and error handling, that could save developer time and prevent errors. The number of accessibility errors caused by developers presuming a demo was complete is hard to measure…
But all of this depends on expert training. It depends on tools being consistent – and not having your prompts slide into entropy because of new bad data.