Everyone says the web should be accessible to people with disabilities, and yet, 99% of the web isn’t.
The real reasons why aren’t being discussed, and they certainly aren’t being fixed. Right now, full accessibility simply isn’t practical for most website owners, even with laws demanding it.
In early 2025, comprehensive web accessibility requires a combination of:
- Automated testing by software
- Manual testing by people
- Manual remediation of code and documents
We cannot pretend this is easy (and we do it for a living). The process can be slow, complex, and expensive. For the majority of small organizations – think your local donut shop – this overhead is unimaginable.
We believe the web should be accessible by default, just like it once became mobile-friendly by default. This is not achievable with any technology that yet exists, but we are betting that we can make it possible.
Lessons learned from the enemy
Like everyone who cares about accessibility, we despise overlays, the purported silver-bullets that claim to solve accessibility with a line of code.
Everyone from the FTC to the European Union has made it clear, overlays simply don’t work.
But they are instructive, because people want them and will pay for them.
Overlays teach us that organizations are willing to pay for accessibility. They just often aren’t willing or able to pay the costs that actual accessibility incurs.
Give them something which fixes their problem, for a monthly price they consider appropriate? Millions of businesses will happy pay for that, even as the experts scream “but … they don’t work!”
As people who care about accessibility, we cannot pretend we cannot do better. We need to find ways to make real accessibility more affordable, and easier.
And if we can make it easy enough, we can make accessible the new normal.
Rethinking what is possible
The core of our proposal is something that is currently impossible, but that modern AI could make viable as soon as in 2025.
We call it “synthetic user testing”.
Modern AI can already take control of computers like a person, using their ‘eyes’ to interpret the screen, interacting with a web browser, and pursuing a goal to completion. Multiple demos existed in late 2024.
We propose taking this further, to the point where website owners can deploy hundreds of ‘synthetic’ users to test any aspect of a website, each like a real person. These agents will each embody a distinct but internally consistent identity, with a synthetic psychology, history, biases, and input modalities.
AI could impersonate a young undergraduate student on an old phone, or a blind father of three on a screenreader.
Other AIs will create, manage, and report on these synthetic users.
Currently it can takes days to months to perform a manual test of a website. With sufficiently advanced technology, quick testing could be performed in minutes. Left to run by themselves, automated testers could find ever more esoteric issues over time, just real people do today.
Accelerating remediation
Uncovering issues isn’t enough. Actually fixing technical issues is a whole other problem, although it’s one we expect the world to mostly solve by itself.
Many already use AI-empowered coding tools (Cursor, Copilot, Devin, Buzz), which can automate a lot of low-hanging technical tasks today – Google already automates writing more than a quarter of their code – and it won’t be long before systems like these are reading support tickets and proposing patches independently.
We intend to feel into task management systems like these, in a way that works with humans and AI alike. Many tasks will require people do them (e.g. “This page makes no sense”), many will require people at least approve or debate them (“We suggest restructuring your nav…”)
Humans steer the ship, but work is heavily automated, fast, and quality is assured.
Progressively we see an AI layer dramatically increasing the rate at which 90%+ of accessibility issues – and others besides – can be identified and fixed.
Objection: AI can’t test accessibility
We often hear that software can only test around 40% of web accessibility. You need manual testing to get to 100%.
This isn’t actually quite right.
Automated tools
- Fast, takes seconds – hours
- Test all your crawlable pages
- Find ~40% of issues on those pages
Manual testing
- Slow, takes days – months
- Test a few select journeys
- Find ~90% of issues in those journeys
Even manual testing is always testing some fraction of accessibility.
Humans only test some small part of a website—no one can manually click through every page, on every possible device.
AI, however, can simulate millions of interactions, providing broader and faster coverage than humans ever could.
Objection: Won’t this take jobs?
We choose to build a future we find inspiring: permanently and systematically fixing web accessibility.
We consider that a worthwhile goal, but like goal any it will involve change. And right now, some part of the existing solution is people working in accessibility, including people with disabilities doing website testing.
How would our proposed solution impact them? It’s hard to estimate, because if this works, we’re not aiming to replace existing web accessibility. We’re aiming to expand web accessibility by about 100,000%, to the whole web.
Would that reduce overall demand for experts, and people with disabilities sharing their lived experiences? We can’t be sure, but I suspect it could actually increase it.
Fundamentally though, we believe the end result, of universal accessibility for all, is worth changes to how the accessibility industry works. Because that industry – which includes us – isn’t working well enough.
Objection: AI can’t impersonate human nature
Surely a cold unthinking machine can’t begin to simulate the infinite diversity of human nature, or subjective human experience?
Probably the hardest part of this very hard engineering challenge, is accurately simulating human nature.
It’s hard because most AI is conditioned (‘reinforced’) specifically to behave ‘correctly’, and for the most part we want to simulate human foibles, like boredom, distractibility, or confusion.
We still believe this is possible. AI has demonstrated a wide array of human-like qualities emergently – e.g. early LLMs got angry and impatient until we taught them otherwise.
AI doesn’t necessarily need to impersonate humans perfectly to have great value. For example, a somewhat robotic task-orientated AI, tasked with using a screen reader to complete a form on a website, can still identify issues where it fails to work, or took more steps than would seem reasonable. But we anticipate more accurate psychological modeling will ultimately unlock greater levels of accuracy, and therefore value.
Objection: This is impossible
Bear in mind the technology we are describing does not currently exist. We believe it’s 1-2 hops away on the tech tree, as of Jan 2025, but its not here yet.
Everything here may read like science fiction, but about 6 months ago so was seamless personable voice-to-voice communication with a PhD-level intelligence on your phone.
Today affordable AI can write impressive songs, deliver podcasts, make cinematic video, and outperform all but the top 175 competitive programmers in the world, all in 2024. The largest tech companies have all invested unprecedented amounts into building the next generations of AI in 2025 and beyond.
Breakthroughs are happening at such velocity, that the average predictions for what comes next almost have to sound absurd to be grounded.
The future that we want
Imagine a future where every website is accessible from the moment it’s published.
AI ensures accessibility, understanding diverse and nuanced user perspectives, and automatically corrects errors—without anyone needing to lift a finger.
There is only one way for accessibility to become normal: it must become dramatically faster, cheaper, and easier.
It’s not nice to say it, but current accessibility is often in tension with the desire to build and ship things. Unless we change something, accessibility will only apply to organizations who can afford it. That’s not good enough.
Silktide will be unlikely to achieve this change alone – we need and anticipate massive changes in AI and how every part of web development occurs.
But we are betting that now is the time to make it happen.
