A friend of this author has been blind since birth. He uses Artificial Intelligence (AI) almost every day — to read a document; to make sense of a photograph; or to get through a government form his screen reader cannot open, all without relying on a sighted person for help. Watching what it gives him, independence in the small things most people never have to think twice about, the promise of this technology is impossible to dismiss. At the same time, it is impossible to assume that this promise will reach disabled Indians on its own. AI models are moving into government services, recruitment and healthcare. Behind them, data centres and power lines are rising as the Centre courts $200 billion in AI investment for the next two years.

Disabled people are being brought into the conversation late, if at all.

In the November 2024 Rajive Raturi judgment, the Supreme Court held that accessibility is a facet of the fundamental right to life and dignity, and that India’s accessibility rules were toothless; they were merely polite suggestions whereas the Rights of Persons with Disabilities Act, 2016 intended binding ones. The court gave the Centre three months to frame mandatory standards. Nearly two years later, the petitioners are back in court because too little has changed. Disabled people should not be made to adapt to systems built without them; AI is the newest such system.

The Chief Commissioner for Persons with Disabilities penalised 155 establishments, government ministries among them, for websites and apps a disabled citizen simply cannot use. While the 2016 Act has required accessible digital services since 2019, compliance remains the exception. This is the ground on which we are now deploying AI.

Inherent bias

AI is not neutral. When researchers ran 21 language models through AccessEval, a benchmark spanning nine kinds of disability, the models became more error-prone, more negative in tone, and more likely to stereotype the moment disability entered the question. A study of the widely used image model CLIP found it 15 percentage points less accurate on photographs taken by blind and low-vision users than on web images, and that objects central to disabled lives, such as a white cane or a Braille display, appeared up to 17 times less often in the datasets used to train such models. You can see this bias in everyday digital interactions as well. Tell a mainstream chatbot you are blind and want to become a software engineer, and it often opens with some version of “I’m sorry you’re blind.” The person asked a career question but the model answered as if they had reported a loss.

Further, a survey of 2,462 users of the NClude platform, which uses these same models to help disabled people complete job applications and government forms on inaccessible websites, revealed findings that were striking: 1,313 users said they had completed a task previously closed to them, but only 543 managed it through AI alone. The rest needed a staff member to step in. A system without that fallback leaves users to improvise a workaround.

The same neglect is being set in concrete. India’s data-centre capacity is projected to more than quadruple by 2030, to 6.5 gigawatts or more. States compete for these projects with power subsidies and duty waivers; Maharashtra this year relaxed its renewable-energy requirement for data centres from 100 to 51%. Rarely does anyone ask what the new load does to a grid already straining at summer peaks, or what the servers’ heat adds to cities that are dangerously hot to begin with.

For millions of disabled Indians, this is not a distant environmental debate. Some depend on powered wheelchairs, oxygen machines or other equipment that cannot be switched off when the grid strains. Many are the first to be stranded when a power cut or an inaccessible emergency alert throws daily life into disorder.

None of this is a case against AI, or the infrastructure a growing economy needs. Disabled people should not be asked to surrender tools they depend on in the name of sustainability. Any AI system the government deploys should be tested for disability bias. Training datasets need genuine disability representation, gathered with consent. And State incentives for data centres should carry conditions on renewable sourcing and on grid and heat impact.

Disabled Indians must have a voice in helping to decide what this technology becomes, from the data that trains the models to the grids that keep them running.

Navin Aade works on climate, disability inclusion and AI at NClude