denkwerk TechTalk
A11y-AI: AI for empathic assessment of accessibility
Artificial intelligence, especially the GPT-4 language model, is on everyone's lips. At denkwerk, we used generative AI to test the accessibility of websites - and to generate an empathetic result tailored to the restriction.
When it comes to modern web design, accessibility is an extremely important point in the specifications. After all, 1.3 billion people – around one sixth of the world's population – struggle with some form of severe disability. In Germany, around 10 percent of the population is affected by disabilities. Visual impairment is a type of disability that is extremely relevant. For people with no or severely impaired vision, using websites can quickly prove to be a hurdle.
Barrier-free tests with A11y-AI
At the denkwerk TechTalk, our Senior Software Developer Miguel Franken presented a new method for testing the accessibility of websites: our newly developed A11y-AI service enables testing in a completely new, empathetic way. Many websites are difficult to use for people with disabilities, and the rate for web stores is as high as around 75 percent. According to Miguel Franken, the problem is usually of a holistic nature: stores and websites are often not universally developed because there is a lack of awareness that people with disabilities must also be able to use the site - and that the restriction affects a not insignificant proportion of the population. In the interests of equal participation, the Accessibility Improvement Act (Barrierefreiheitsstärkungsgesetz, short: BFSG) will therefore also apply in the private sector from June 2025. This means that non-accessible digital services can be penalized. However, this should not be the reason for addressing the issue. At denkwerk, we therefore try to awaken empathy for our customers' target groups among decision-makers.
Accessibility testing is very technical
Existing methods for testing are often very technical and not very attractive, neither for developers nor for users. That's why we at denkwerk thought about how we could make testing more empathetic. The solution is the A11y AI: it uses GPT-4 and the Axe Core to create reports in "normal" language from the perspective of a person who is actually impaired. This can help providers of digital services to make their websites more accessible: The idea is that they can use A11y to better understand why optimization is necessary.
"Accessibility is still not a matter of course in the realization of software projects.
A11y-AI uses the possibilities of Large Language Models to present the effects of barriers in the web on people with disabilities in a sensitive and personal way.
In this way, we want to create empathy for the realization of accessible web projects."
But how does it work? The system consists of three pillars: a persona, the accessibility check and the context, i.e. the content of the website. All of this is brought together using GPT to provide an analysis in natural language, which the persona appears to have written itself.
The solution: an artificial persona
In a first step, we have therefore developed an artificial person ("persona") that has a background story and a certain type of restriction. Claudia was developed as the first persona – others will follow: Claudia is 63, an architect, and suffers from macular degeneration, a disease that severely impairs central visual acuity: People with this disease can only see sharply at the edge of their field of vision, and the disease leads to blindness in the medium term. The persona is therefore severely visually impaired.
In a second step, the system uses Axe-Core tool to analyze the website entered and outputs a report. Axe-Core currently analyzes 103 accessibility rules, but the generated A11y report is very technical. Therefore, our development team has designed a preprocessing that translates the technical output into non-technical language.
Together, A11y-AI creates an accessibility report in natural language, which appears to have been created by "Claudia": she points out problems and explains why they are relevant for her and possibly for other people with visual impairments. The system also recognizes rule violations that are so serious that assistance systems cannot be used, for example. The text is divided into three sections: Introduction, the actual report – with details of the barriers found – and the conclusion.
And so, after entering a website URL, "Claudia" determines the difficulties that can arise for a person with a severe visual impairment. Through the persona "Claudia", the system is empathetic and shows that the web must be made accessible. This approach helps decision-makers better understand why the optimization makes sense – and thus people with disabilities have a much better online experience.
