Yes, let's teach LLMs accessibility, but also provide the companies using them with better strategies

A prominent vendor of accessibility tooling made the case for teaching AI agents about accessibility, because developers use them and therefore this teaching is a way to shift left.

Shifting left, just so we're on the same page, is a strategy for increasing accessibility at scale. It moves efforts from being late in the process, like fixing an app long after it's gone live, towards being as early in the process as possible, like being present in developer tooling. Linters, design systems and assistance in authoring tools are all examples of this. And I see how in-editor LLMs with contextual hints about your codebase are too.

I agree that harm caused by developers who use AI (whether forced or not) needs mitigation. I also welcome initiatives that increase the amount of factually correct data into the training data. And I am sure companies that specialise in accessibility can be helpful here, sharing their knowledge via MCP servers. This makes it easier for “agents” to “understand” common concepts and increases the likelihood of the right information making its way to people building websites. And that's a good thing.

(As an aside: a lot of the “understanding”, LLM vendors also gain from scraping people's content, including this blog and that tool vendor's knowledge base, without permission, violating their copyright, ignoring their robots.txt and costing them high data bills— yes I am salty about that)

Bazookas

But, one thing doesn't sit well with me. In terms of increasing accessibility at scale, teaching LLMs to improve accessibility is akin to killing a fly with a bazooka. As in… an enormous amount of parameters is required to have a large language model that is trained and fine-tuned sufficiently to output coherent answers, and then there's a huge amount of compute needed to output quickly. While the difference between accessible and inaccessible code is knowing to use the right HTML, attributes and keyboard behaviours, often a couple of lines of code.

Don't get me wrong, I realise part of the industry only want to work with bazookas, and I get that we must therefore optimise for bazookas. I don't mean that sarcastically in any way, I think it's great companies are doing this.

But my point is, as accessibility practitioners, we have a much more direct way to improve accessibility through “shifting left”.

As soon as we're done teaching LLMs, we should probably spend most effort on teaching the organisations we work with.

Teaching organisations

We can teach them to spend those resources on learning and keeping up to date their knowledge of HTML, ARIA and how people use the web. To channel that energy into copy-pasteable or installable components. Into easy to read guidance. Into authoring tool functionality. They would not be alone, lots of organisations are taking such opportunities, and have been for many years.

We can teach them to not force developers to use AI to write code, and use processes that are more efficient, more reliable and less of a monumental strain on the environment.

We can teach them to value the professionals that understand accessibility deeply, and warn them about the business, legal and moral risks if they were to replace those professionals with bots.

Conclusion

Of course, we should welcome efforts that teach LLMs about accessibility, but my hope for our industry is that we don't forget the effectiveness associated with hiring and training specialists who understand deeply how people with disabilities use the web and act on that.