Introduction
A few days ago, Gemini’s automated scheduling system pushed me a daily accessibility-related article. The article described a service whose core technology allegedly uses AI to add web accessibility features during frontend rendering, while also offering personalized shortcut settings for users to adjust the website experience.
At first glance, this sounded exciting.
“AI that can add accessibility techniques during web rendering?”
I thought perhaps some new technology had finally been invented and applied in a practical way.
But after looking into it more carefully, my excitement quickly turned into disappointment — and eventually, anger.
From Excitement to Deep Disappointment
The article that Gemini collected turned out to be, in essence, a sponsored article.
The service mentioned in the article talked about accessibility widgets — a type of solution that has already been widely questioned and has not been proven to meaningfully solve accessibility problems on the web.
Then the company explained how its technology was supposedly better than traditional widget-based solutions.
But in reality, what they were doing was simply adding accessibility-related DOM enhancements during page rendering. As for the “personalized shortcut settings” they provided, that part was basically no different from other accessibility widgets. The tools may vary, but the nature of the solution is largely the same.
How do I know?
Because I actually went to the company’s official website and also tested some of the e-commerce websites they claimed to support.
I opened every automated testing tool I had, ran the tests, and then followed up with my own manual accessibility review.
Here are some of the problems I found:
The service claimed that its AI could generate image alt descriptions based on images. But in practice, the descriptions were very generic — something like “a woman wearing a dress.” That kind of description provides very limited value.
There is a well-known accessibility principle: No ARIA is better than bad ARIA. But this technology seemed to add ARIA attributes almost everywhere it could, without considering whether they were necessary. For example, even buttons that already had visible text descriptions were still given additional ARIA labels.
The heading structure was not used in a proper order.
Color contrast still failed to meet WCAG 2.2 AA standards, not to mention the more advanced color-related expectations that may come with WCAG 3.0 in the future.
As for the widget part, it followed the same fundamental problems as many other similar products, so I will not go too deeply into that here.
In other words, even the most common accessibility issues seen across the web were not properly fixed. The site could not even pass basic automated checks like Lighthouse and axe-core. At that point, there was no need to even move into the deeper manual testing stage.
The problems on the company’s own website were already concerning. But when I tested the real customer websites in the domain they claimed to have specifically trained their technology for, the results were even more frustrating.
In some cases, even basic keyboard focus indicators were missing.
At that moment, I truly felt deceived.
I was genuinely angry.
From Anger to Gratitude
But after the anger passed, I started to see the situation from another angle.
At least this company tried something new.
Throughout history, progress has often been built on failure. Without failed attempts, there is no foundation for future success.
With that in mind, I am still grateful that they took concrete action to implement their idea and launched a real product used by real customers.
Yes, the product has some effect.
But it also introduces more problems.
Still, they took a new step. Maybe this approach will improve over time.
After all, humans have always wanted to save effort while producing more. That desire has led to many great inventions.
Conclusion
AI is not omnipotent.
And AI that has not been properly trained is even more dangerous when applied carelessly.
Behind the sugar coating of AI, I still believe that products like this should go through complete testing before being released.
Or perhaps AI should be used more carefully in the development process itself — closer to the source of the website — rather than being applied as a patch after the fact.
Even better, accessibility should be planned from the design stage, where standards, structure, interaction, and content can be considered from the beginning.
I once experimented with adding a widget to a website. But instead of filling it with a bunch of setting buttons, I made it an AI chat interface. Through conversation, users could change different aspects of the website’s presentation.
The experiment worked.
But the cost was extremely high.
And another issue appeared: users started focusing on the chat window itself, which created new accessibility and usability problems.
So yes, AI may be useful.
But AI should not become a beautiful wrapper around an incomplete solution.
Accessibility is not something that can simply be sprinkled onto a website after it is built.
And it should never be treated as a marketing feature covered in AI sugar.