AI-coding and accessibility
How well is it working and what can HUMANs do
A few days ago, the PM of India, Hon. Narendra Modi, Addressed the world at the France AI summit. At the beginning of his speech, he shared an interesting fact,
If you upload your medical report to an AI app, it can explain in simple language, free of any jargon, what it means for your health. But, if you ask the same app to draw an image of someone writing with their Left hand, the app will most likely draw someone writing with their Right hand. Because that is what the training data is dominated by.
source: Transcript from ministry of external affairs.
I am sure most of us will try this small experiment… because I did it too…
First, I tried Google’s Gemini. It is easily available if you have a Gmail account and it is right ‘there’ on any Android phone.
I was both disappointed by the results and amazed by the ignorance. I tried Openai’s Dall-E, the biggest name in the AI space. They too DID disappoint!
After trying various tools for a while, and validating the hypothesis, I started digging deeper into the stats…
The Left handed population.
It is a well-established fact that the world is unfair to left-handed people. In earlier times, the use of the right hand was so deeply embedded in society, that parents would force their children to use their right hand to write and do stuff. Look at the word itself. it is called the RIGHT hand! I am one of those cases who was forced to write with his right hand after my grandparents discovered my inclination towards preferring my left hand to hold the pencil.
Bonus story, after I broke the bones in my right hand multiple times, I realized that I was ambidextrous!
And now, back to the main point.
The global statistics suggest that the number of left-handed people is currently 12%, with some regionality. This number rose from a mere 2% in the 1860s. This chart taken from Penn State University should highlight the global rate of change of left-handed people.
If we consider technical advancements in photography, people started people in action somewhere around 1960. This means that the training data fed to the AI models had on average 12% data on left-handed people. Despite that, the AI is unable to generate an image of a left-handed person, even when the instructions mention ‘a left-handed student'!
PwD population
In the early days like the 1950s, when technology was not a household item, people with disabilities were not able to function normally in society. That means the reporting of the disability was also less. In the last few years, we have seena significant increase in life expectancy. That also resulted in more people with problems associated with age, like loss of vision, reduced hearing, or motor impairments.
For instance, the total number of visually impaired people in 1960 was about 2% of the global population, whereas in 2025 that number is estimated to hit 14%. I'm going to go out on a limb here and assume that all the other disabilities would follow a similar pattern.
The quality of the source
Global statistics say that only 2-4% wof ebsites are accessible to users with disabilities. Although this seems like a stretch, a study conducted by webAIM with the top 1 million web pages concluded that 96.3% of them were inaccessible. Statistically, if we assume that all the code running on all the websites is available for the AI to train itself, 96% does not contain any accessibility considerations, or have wrongly implemented accessibility. As a result, when AI generates code for an app, the chances of it getting the accessibility right, is optimistically 4%!
However, in the early example we saw that even 12% of data doesn't allow AI to generate the image of a left-handed person. So with just 4% accurate data the chances of AI getting the accessibility of an app right, are practically zero.
So what can we do?
Relax, it is not all bad news.
AI is great for rapid prototyping
The AI is fast in creating concepts for rapid alpha testing. You can test the concepts of your business idea scarily fast. This gives you a simple application, which may or may not be accessible. But with this, you can go to the investors and get the right amount of funding, for proper development with accessibility best practices included.
Ai dependent development comes with a limitation. AI is great only for demo quality apps. They stop getting better after 70% quality mark. The article titled The 70% problem by Addy O covers all these limitations very well.
AI is learning the best practices
The AI tools are learning not just from the data sources, but also from best practices. Tools like lovable.dev already produce react codes that perform pretty well on the accessibility aspects. Cursor as well is getting there. PS. Claude needs to work on their publishable artefacts. They are barely accessible.
AI is not only learning to code, but also learning to understand the context and analyzing images. This will allow AI to maybe generate better alt texts for the images on the websites.
Big tech is investing a lot on accessibility
Many countries are opting to make the accessible products and services a legal requirement. USA and Canada already have it, while EU is getting ready, with EAA coming in force on June 28, 2025. With this, we're going to see a big shift in the way that apps and websites are designed and developed. It's like a big, accessibility-themed party, and everyone's invited. We're already seeing some pretty amazing developments in the world of accessibility tech, like voice assistants and augmented reality. For example;
Apple Introducing Eye tracking
Google introducing magnifier
Android 15 getting equipped with the focus hearing
Airpods that can be used as hearing aid.
And as these technologies continue to evolve, the possibilities for creating truly inclusive experiences are going to be endless.
Agentic AI
The AI is also getting into the agentic model. This means, the AI will be able to perform actions on your behalf. This is still a new tech and openAI just announced this one. I can’t confidently say that this will allow users to perform tasks just by asking AI to do it, but it might reach there. However, does the development have to follow the accessibility / element-labelling best practices to ensure that the AI can operate the apps? This we can’t find out unless we are willing to pay $200 per month to use the openAI pro version!
Wrapping up.
Making apps accessible has been a challenge and given the way the world is using AI, it will be some time till the AI developers figure out how to build more inclusive products. Till then, the accessibility champions worldwide have a responsibility of being the ‘HUMAN’ in the loop for any AI based development.