Latest WebAIM Million project shows that AI is making web less accessible. We can blame the bias from training data – all the inaccessible code out there that was used to train AI. But we can also note that people focused on accessibility can get good results out of the same AI.
This is not new to many using and testing AI possibilities – AI is not just a tool, it can also be, and often is, an amplifier. The problem is that, by default, it amplifies inaccessibility. As it is biased towards it. But when a human operating AI understands accessibility it can lead AI to better results. Same AI, different human (prompt). I wrote about shifting accessibility left in AI as well. But we still need the human operator or AI agent manager to understand accessibility.
Skills, instructions, harnesses, feedback loops – they all help, but are not enough – and as you probably noticed they are not helping consistently. We can not delegate the accessibility knowledge to AI and leave it at that. As with other fields – human needs to understand, lead, verify the output of AI.
Testing and verifying code is valid, automatically, is not difficult and does not even need AI – often it is even better to just use deterministic static testing without AI to get consistent results. But understanding what the code really does, especially when we also add human perception, operability, understanding and possibilities of assistive technologies is very difficult.
It has the potential, but only if we care about accessibility for people
Well – again nothing new – we have accessibility scores to learn from. If we let AI to “fix” the score – it will do it, eventually. We will get the 100/100. And we can be sure that page is still not accessible, and far from usable.
Chasing the score for the score itself is wrong. And we do not have automatic testing that reveals if we chased the score or actually made things accessible. Chasing scores is chasing static test coverage.
Adding “230VAC” as alt text for image showing “110VAC” in home appliance user manual passes the score of automatic tool and potentially destroys the appliance in the real world.
Passing a score versus real world example
So – if you are using AI (or not), and you want to pass the agentic browsing score, and then you dive in the requirements and see that they require “well-formed accessibility tree that helps AI agents to navigate and interact with the page”, please make sure what this well-formed accessibility tree actually means.
As Lighthouse is open source we can check the agent-accessibility-tree.js audit source (opens in new window) and we can see the target rules:
- button-name
- input-button-name
- input-image-alt
- label
- link-name
- select-name
- document-title
- aria-allowed-attr
- aria-allowed-role
- aria-command-name
- aria-conditional-attr
- aria-dialog-name
- aria-hidden-body
- aria-hidden-focus
- aria-input-field-name
- aria-prohibited-attr
- aria-required-attr
- aria-required-children
- aria-required-parent
- aria-roles
- aria-text
- aria-toggle-field-name
- aria-tooltip-name
- aria-treeitem-name
- aria-valid-attr
- aria-valid-attr-value
- duplicate-id-aria
- definition-list
- table-duplicate-name
- tabindex
- autocomplete-valid
- presentation-role-conflict
- svg-img-alt
All these can be found in the accessibility audit (that has some more) and you can find them in axe-core as well, just for clarity.
As we can quickly see – it is up to developers to implement them – and if the developer focuses only on fixing the issues with AI – we can almost guarantee barriers for people. Some parts are also not only up to developers and require planning in advance (content and UI UX designers) and some are only made possible by developers but used by content providers – showing that accessibility (tree) is a group responsibility!
Changes are AI will just burn more tokens and figure it out.
Chasing scores may look well on some dashboard, discriminating people, getting complains or even legal action because of people chasing scores instead of making things accessible is a real threat.
So, basically, we are in the same situation as before – chasing scores for the scores will mean inaccessibility. Fixing accessibility for people will make things better for AI.
And we can use this to promote some WCAG on level AAA. Context-bearing links and button names should become a default now.
