It is possible to use AI to write alt text and scale accessibility. But there are a lot of open questions.
What is alt text?
Alt text is a written description of a digital image. The text is in the markup for the image, and doesn’t appear on the web page. Alt text is used to make visual works available to people who use a screen reader to access webpages.
Alt text is required for accessible websites and ebooks. The Web Content Accessibility Guidelines (WCAG) specifies equivalent text for images. It is more important than ever to meet WCAG standards. The European Union has set accessibility requirements for media based on WCAG. These standards go into effect in June of 2025.
Long-running blogs can need a lot of alt text
Blogs that have been publishing for a long time may have a backlog of thousands of images that need alt text. Dries Buytaert’s website has 10,000 images. Of those, 9,000 do not have alt text. Dries ran tests to see how well AIs could write alt text. He selected three images that presented challenges to the AI, and graded the results.
Ebooks will need alt text too
There are millions of ebooks in publisher’s backlists. And sometimes hundreds of images in a single ebook.
But what about hallucinations?
Nielsen Norman Group argues that hallucinations are common. In fact, they say that hallucinations are part of how LLMs work. They suggest two things to build trust while using AI tools. First, let people know that the text is AI generated, and then give them tools to verify the results.
Would this work for alt text? You could include a disclaimer somewhere that the alt text is derived from AI. I’m not sure how verification would work in this case. How would a person with low vision go about verifying the description of an image they can’t see?
Will using AI lead to more piracy?
Are LLM models training from the images they are processing? Would running the AI locally prevent theft? What is involved in setting up a local deployment?
Let’s keep looking for best practices for LLM generated Alt text
The ideal workflow balances AI with human oversight. And balances speed and accuracy case by case. How would you decide which images should have AI generated alt text vs human writing or review?
Maybe you could triage a blog or backlist for the most popular posts or titles. And then make those a priority. They would get the most human intervention. Images on less frequently-viewed posts would be AI generated for speed.
Another approach is to consider the potential harm of inaccurate alt text. For instance, alt text for images in a safety manual should be completely trustworthy.
While technology can scale accessibility, there are open questions about using AI to generate alt text for images. These questions include concerns about accuracy, potential for hallucinations, and copyright infringement.
Contact me if you are interested in efficiently adding metadata to your blog or ebook backlist.