Available Isn’t Accessible
| AUTHOR | Gopal Tomar |
| DATE | July 20, 2026 |
| CATEGORIES | Artificial Intelligence AI Governance Platform Governance Principles |
Artificial intelligence has never been more available. Today, anyone with a smartphone or an internet connection can ask a chatbot a question, generate an image, translate a document, or summarise a report in seconds. It is easy to mistake this widespread availability for accessibility.
But availability and accessibility are not the same thing.
When we hear the word accessibility, we often think of wheelchair ramps, screen readers, Braille displays, or people with visible disabilities. These are important, but they represent only one part of a much larger idea. The World Health Organization (WHO) estimates that around 1.3 billion people live with significant disabilities, yet accessibility has always been about more than disability alone. At its core, it asks a simple question:
Can the people a system is designed for actually use it in a meaningful way?
That question becomes far more difficult when applied to AI.
The W3C Web Content Accessibility Guidelines (WCAG 2.2) describe accessibility through four principles: systems should be perceivable, operable, understandable, and robust. In practice, AI quickly exposes how difficult these principles are to uphold.
A chatbot may be operable because anyone can type a question into it. But is it understandable if it provides an answer without showing where that answer came from? A map may look polished, but is it reliable if the dataset, projection, scale, or uncertainty remain hidden? A legal or healthcare assistant may sound convincing, but how accessible is its advice if users cannot verify the source, check whether it is current, or understand its limitations?
This is where many AI systems fall short.
The issue is rarely the interface; it is the lack of verification.
That failure appears in three ways.
First, AI can produce incorrect answers with remarkable confidence.
Second, users often have no audit trail—they cannot see which sources were used or where the model inferred information.
Third, performance is inconsistent across languages, cultures, and contexts. A model that works well in standard English may struggle with regional languages, mixed-language conversations, or disability-specific terminology.
Seen this way, accessibility is not just about receiving an answer. It is about having the ability to question it.
This distinction also separates assistive technology from universal design. Assistive technologies—such as screen readers, captions, speech-to-text software, or Braille displays—address specific needs. Universal design aims to build systems that work for as many people as possible from the beginning.
Think of a wheelchair ramp. It certainly benefits wheelchair users, but it also helps older adults, parents with strollers, delivery workers, and travellers carrying luggage. The same principle applies to AI. Features such as citations, plain language, multilingual support, correction mechanisms, and transparent source trails do not benefit just one group—they improve accessibility for everyone.
Accessibility, however, is not only a design challenge.
It is also a question of scale.
Today’s largest AI models promise to serve millions of users across countless tasks. Their capabilities are impressive, but so are their costs. These systems are expensive to build, highly centralised, difficult to scrutinise, and often opaque. Behind every model lies infrastructure that consumes enormous amounts of electricity, water, land, specialised hardware, and capital. Recent reporting from India has also linked rising electricity demand to the rapid growth of data centres, AI infrastructure, and electric vehicles.
This raises a broader question: who benefits from AI, and who bears the cost of making it available at scale?
The answer is not to assume that smaller models are automatically better. They can also be biased or poorly evaluated. But they encourage a different way of thinking.
Participatory AI starts with the people who will actually use the system. Instead of assuming that one model can solve every problem, it involves communities, domain experts, and intended users in defining the problem, shaping the data, evaluating the outputs, and improving the system over time. The OECD AI Principles reflect the same idea by emphasising transparency and explainability, ensuring that people understand when they are interacting with AI and can challenge its outcomes.
Some Indian examples show why this matters.
Myna Mahila Foundation’s Myna Bolo is designed around women’s sexual and reproductive health questions, local languages, literacy levels, and the need for confidential access to information. According to the organisation, the system has been trained on 170,734 real questions asked by women from the community. It is not built around an imagined “average user”; it is built around the people it intends to serve.
Adalat AI demonstrates the same principle in a different domain. Rather than becoming a general legal adviser, it focuses on specific tasks such as court transcription, workflow support, digitising records, and real-time case updates. Because its purpose is clearly defined, its outputs can be reviewed and verified within an existing judicial process.
Both examples point to the same lesson.
Accessible AI is not simply about reaching more people. It is about being understandable, verifiable, and accountable to the people who depend on it.
The conversation around AI is often driven by scale—how many users it can reach, how many languages it supports, or how quickly it generates answers. Those are important achievements, but they are not the same as accessibility.
For me, accessibility begins with trust.
A system is not accessible simply because it is online or widely available. It becomes accessible when people understand how it works, know where its answers come from, recognise its limitations, and have the confidence to question or correct it.
The future of AI should not be judged only by how intelligent our models become, but by how accountable they remain to the people who rely on them. If a system can provide answers but cannot be inspected, challenged, or corrected, then it has not truly expanded access—it has merely expanded availability.
Because in the end, accessibility is not measured by how many people can open an AI system.
It is measured by how many people can understand it, trust it, and use it with confidence.
Until AI can do that, we should be careful not to confuse availability with accessibility.