I was watching an AI read my own book back to me.
Gemini, specifically. Three chapters in. Citing me to my face. Blowing smoke about how important this work is, what a vital contribution to accessibility the author makes. (The author was me, lol. Gemini did not know this yet.)
So I did what anyone in my position would do. Screw it, I told it. I’m Jessica Chambers.
Among the flattery that followed was a casual mention of Taffy – our Siamese cat. The OG. The one who appears in basically every illustration across the book and the WCAG guides and a fair chunk of the Silktide educational content. I mentioned it later, in a different chat with a different AI. That’s weird, right? The other AI was confident. Gemini’s hallucinating. There’s no Taffy. I almost let it go.
But human memory is fallible and twelve cats is twelve cats and I’m a verifier by nature, so I went back to the page.
The chapter Gemini had been reading was the one about accessibility overlays. The one about snake oil. The one explicitly about people gaming the appearance of accessibility for marketing purposes.
The body text of that chapter does not say “Taffy” anywhere. Not in the prose, not in the headings, not in the snake oil rundown. Taffy is nowhere in the visible content of the chapter.
But the alt text on the first illustration reads, in part: “Taffy, a Siamese cat, presents the viewer a bottle labelled Taffy’s amazing accessibility elixir… Taffy’s wearing a bowler hat and has a pointy moustache. He smiles with a wink.”
Three mentions of Taffy. All in the alt attribute. None on the page.
Gemini was not hallucinating. Gemini was reading the accessibility tree.
The proof was in the part of the page that exists for accessibility.
I am not here to give you a hot take on what AI is going to do to accessibility.
I am here to tell you we have been here before, and here is what we know about how it ends.
The loop nobody is watching
Here is the mechanism, stated as plainly as I can:
AI is now writing the accessibility tree, at scale, badly. And AI is now reading the accessibility tree, at scale, preferentially.
The same layer is both machine-generated and machine-consumed, and almost nobody in the accessibility community has clocked that this is the shape of the thing now.
When an AI agent reads a webpage, the cheapest, cleanest, most structured signal available to it is the accessibility tree. Not the DOM. Not the visual rendering. The accessibility tree. The thing we built for assistive technology (screen readers, voice navigation). The thing alt text and ARIA labels and semantic headings have always been feeding. Of course agents are using it.
And on the other side: when LLMs build sites – and they are building a lot of sites now – they reach for ARIA almost reflexively, often inappropriately, in ways that the maxim every accessibility practitioner knows would diagnose as doing it wrong. There is already a meaningful proportion of new web content where the accessibility tree was written by a machine that did not understand what it was doing.
Machines writing the metadata. Machines reading the metadata. One of the most consequential signals in the next decade of commerce is now flowing through a layer that almost no one is watching.
That’s the loop. The rest of this essay is about why it’s about to get bad.
I have seen this movie three times
I got my Microsoft Certified Trainer in 1997. I am old enough to remember Mavis Beacon and to have read Don’t Make Me Think when it actually came out. I was there for mystery meat navigation. I have watched, in real time, the following cycles play out:
Alt text was introduced in HTML 2.0 in 1995 as fallback for broken images and text-mode browsers. Accessibility didn’t get formally attached to it until WCAG 1.0 in 1999 – four years later. Once it became an accessibility tool, search engines noticed, and within a few years every product image had alt text like “buy cheap Nike shoes free shipping shoes nike sneakers.” The defenses caught up eventually. The cost in the meantime was paid by screen reader users who got keyword soup read aloud.
Meta keywords were never visible to users. They existed to inform search engines about the content of a page. Within a few years, every page on the internet had meta keywords stuffed with every imaginable term. Search engines stopped trusting them entirely.
Link farms. Cloaking. Schema.org abuse. Every single time someone introduced a structured, machine-readable signal to the web, the same thing happened on roughly the same timeline:
Useful → exploited → defended → either patched or deprecated.
Every. Single. Time.
The web has immune responses to metadata exploitation. Each one takes years to develop. The patient survives. The cost during the meantime is paid by whoever was relying on the layer when it got compromised.
ARIA is next. ARIA is already next. The only question is what gets paid in the meantime.
What’s different this time
Two things make this round structurally worse than the previous ones.
The first is the loop I already named. The generation and the consumption are happening in the same ecosystem. The pollution is being produced by the same kind of system that’s reading from it. The defenses for either side don’t exist yet. It’s not “AI breaks accessibility eventually.” It’s “AI builds the broken accessibility AND drinks from it.” The spam loop is self-feeding before anyone’s even built the spam detection.
The second is the part that should keep you up at night.
Google told us, years ago, exactly why they wouldn’t make accessibility a direct ranking factor. Per John Mueller in 2022: they couldn’t quantify it. They couldn’t define accessibility across sites in a way they could measure at scale, so they didn’t try.
AI agents quietly fixed their reason.
They don’t need a metric like Core Web Vitals. They don’t need to define accessibility in the abstract. The tree IS the metric. If the agent can complete a task using a site’s accessibility infrastructure, the site passes. If it can’t, it fails. The quantification problem that kept accessibility out of Google’s rankings for years has been solved – silently, in the routing logic of agents that aren’t in any way trying to do accessibility a favor.
Which means agent-success-rate – can this AI actually complete a transaction on this site – is about to become a ranking signal. Not on a Google SERP. Not on a star rating. Inside the routing decisions of the agent ecosystem itself. Invisible. Unpublished. Unappealable.
The exact mechanism might look different from what I’m sketching. But the incentive structure is already visible.
If your site’s agent-success-rate drops, you don’t get a notification. You don’t get a dashboard. You don’t get a metric you can A/B test against. You just get less traffic, slowly, with no signal as to why.
And recovery from this is structurally harder than recovery from any SEO penalty in history. The things that have always protected human-facing businesses – brand loyalty, sunk cost, the second-chance instinct, the willingness to forgive a glitch – don’t apply to agents. An agent that failed once and found a path that works does not come back to retest. There is no “we trust them again now.” There is no reconsideration request. The deranking is silent and the floor is hard.
Which means the incentive to keyword-stuff ARIA is about to become enormous. And the defenses against keyword-stuffed ARIA do not exist. Google evolved Penguin and Panda to catch SEO spam over years. There is no equivalent for ARIA spam – no algorithm watching for fake aria-labels, no team auditing the accessibility tree for trustworthiness. Until about ten minutes ago, the only consumer of the accessibility tree was a screen reader and the only adversarial actor was a lawyer.
That changes now.
The maxim
Every accessibility practitioner knows a three-line folk maxim about ARIA. Google can’t find the original source. It’s community property, fully unattributed, and honestly, that’s how I prefer it.
Never use ARIA unless you must.
Always use ARIA when you have to.
You’re doing it wrong.
The third line is what makes it land. Every “you’re doing it wrong” in the history of ARIA has been accidental. Developers misusing labels because they read one Stack Overflow answer and stopped reading. Well-meaning teams over-applying roles inside a JavaScript framework that has its own opinions. The university homepage with 5,552 ARIA elements that helped no one and broke quite a lot.
The ARIA Apocalypse is what happens when “you’re doing it wrong” becomes intentional. When the wrong is the point. The maxim doesn’t break. It just gets a darker reading.
Who pays
The cost lands on disabled users.
Every previous wave of metadata exploitation has had the same downstream effect – the people who needed the metadata to be accurate got keyword soup instead. The structured data they were relying on to navigate their daily lives got polluted in service of someone else’s marketing.
AI agents can recalibrate. Agents can be retrained. Spam detection can be added later. The screen reader user has no equivalent move. They were already using the accessibility tree. When it becomes spam, they live in it. The pollution lands on them first and hardest and they do not get the patch.
Margaret Atwood wrote: “Better never means better for everyone. It always means worse, for some.”
The “better” here is better agent-success-rate for sites willing to game ARIA. The “worse, for some” is the population the accessibility tree was built for in the first place.
That is who pays.
Move fast and dodge things
Silktide used to have a poster up that said Move fast and dodge things. I have thought about that line a lot since.
Because we know what’s coming. We have the playbook. The web has metabolized this kind of exploitation before – alt text, meta keywords, link farms, schema. Each cycle took years to defend against. Each one extracted a cost from whoever was using the layer when it got compromised. We have watched this movie three times.
I don’t have a clean prescription. I am not going to pretend I do. There is no magic policy that fixes this, no detection algorithm that’s been built yet, no compliance framework that catches an agent-routed site silently losing traffic to a competitor with better ARIA spam. Most of the work of the next five years is going to be done by people who haven’t realized the work needs doing yet.
But the cycle is the cycle. Move fast, break things. Move fast, break the accessibility layer. Move fast, break the people the accessibility layer was built for.
I am old enough to have seen this. Now you have, too.
Move fast and dodge things.