Prime Minister Narendra Modi’s declaration that artificial intelligence must be “democratized” and turned into a force for inclusion speaks directly to India’s place in the next technological order. It is an ambitious statement, and rightly so. India cannot remain a passive consumer of artificial intelligence while the rules of the future are written elsewhere.
But there is a harder truth beneath the optimism. AI cannot democratize India if it understands only the English speaking elite. If AI speaks fluently to the elite but stumbles before the majority, it will not democratize India. It will only modernize exclusion.
This is the central challenge that India must confront before the rhetoric of inclusion becomes another familiar promise lost between policy speeches and market excitement. Artificial intelligence may become cheaper, faster and more widely available. It may enter classrooms, hospitals, courts, banks, agriculture, public service delivery and small businesses. But if it fails to understand how Indians actually speak, it will serve the already privileged more effectively than the people who most need technological access. That would not be inclusion. It would be a digital reproduction of social hierarchy.
India is already one of the world’s most important AI markets. Its young population, smartphone usage, digital payments infrastructure and growing comfort with online services make it attractive to global technology companies. For Silicon Valley, India is not a distant opportunity. It is central to the next phase of AI growth.
But India is not a market of one language, one accent or one cultural register. It is a republic of speech communities. Public life here is conducted through Hindi, Bengali, Gujarati, Tamil, Telugu, Marathi, Malayalam, Kannada, Assamese, Odia, Punjabi, Urdu, Meiteilon and many other languages. It is also conducted through dialects, accents, mixed grammar, borrowed English words, local idioms and everyday code switching. This is not linguistic disorder. This is India’s ordinary reality.
A trader may discuss inventory in Gujarati, payments in English terms and family matters in Hindi. A student may study science in English but understand concepts better in Bengali or Assamese. A farmer may describe a pest problem in a local dialect that does not appear in formal textbooks. A government beneficiary may not type a polished prompt but may speak into a phone using natural, broken, regionally inflected language. If AI cannot understand these forms of communication, it cannot claim to understand India.
The language problem has deep roots in how modern AI was built. The first generation of large models relied heavily on internet text, and the internet has long been dominated by English and other high resource languages. This gave English a structural advantage. Languages spoken by millions, even hundreds of millions, were treated as low resource languages because they were poorly represented in digital training data. This is one of the great contradictions of the AI age. A language can be socially vast but digitally weak.
The problem becomes more serious as AI moves from text to voice. In India, speech is not a secondary interface. It is often the most natural and inclusive interface. Many citizens who are uncomfortable typing can speak. Many small businesses operate through phone calls and voice messages. WhatsApp voice notes, spoken payment confirmations, customer service calls and oral instructions already form a major part of India’s digital economy. For technology companies, voice is the next market. For India, voice is a public interest question.
A model that misunderstands a shopping query may inconvenience a customer. A model that misunderstands a medical complaint, a legal statement, a banking instruction or a welfare request may create real harm. The difference matters. Once AI enters essential services, linguistic failure becomes more than a product weakness. It becomes an institutional risk.
Indian speech is difficult because Indian society is difficult to simplify. It carries background noise, regional pronunciation, incomplete sentences, local slang, caste and community references, religious terms, English technical vocabulary and emotional context. A person may begin a sentence in one language, shift to another, and finish with English administrative words. This is how India speaks. A model trained on clean, narrow or urban datasets will not be enough.
This is why benchmarks are vital. India cannot depend only on evaluation systems designed elsewhere. Standard speech recognition metrics may not capture the realities of Indic languages. A transcription may appear acceptable to a foreign evaluator while distorting meaning for native speakers. A model may perform well in formal Hindi but fail in ordinary mixed Hindi. It may translate words and still miss intention.
India needs its own language evaluation infrastructure. It needs benchmarks designed by Indian linguists, computer scientists, universities, public institutions, community experts and native speakers. These benchmarks must test real speech, not merely textbook sentences. They must include regional variations, noisy environments, code switching and social context. Without this, companies will continue to make broad claims about localization without facing serious scrutiny.
The issue is also central to India’s sovereign AI ambition. Sovereign AI is often discussed in terms of domestic models, national datasets, computing infrastructure and strategic autonomy. These are important. But sovereignty also includes linguistic control. A country that cannot process, preserve and build technology for its own languages has not achieved meaningful technological independence.
Language is not merely a medium of communication. It is a storehouse of memory, culture and authority. It carries humour, respect, grief, anger, negotiation and identity. When AI fails a language, it fails more than vocabulary. It fails a social world.
This matters especially for India’s smaller languages and dialects. If market incentives alone decide the future, companies will prioritise large and profitable language groups. Smaller speech communities will remain poorly served. Over time, this may create a new digital hierarchy in which some languages become technologically powerful while others are pushed further to the margins. A democratic country cannot allow that outcome to be treated as natural.
There is also a question of data ethics. Building language AI requires speech data, transcription, annotation and human review. This labour is often invisible. It is frequently performed by people who are paid little and given limited control over how their contributions are used. India must not build an inclusive AI future through extractive data practices.
Citizens who contribute speech data should know how their voices will be used. They should be fairly compensated. Their consent should be meaningful. Sensitive linguistic and community data should be handled with care. Data collection from poorer regions must not become a quiet transfer of cultural and linguistic value to distant corporations. If Indian voices train commercial AI systems, Indian citizens must not be treated merely as raw material.
The government has a major role to play. AI systems used in public services should be tested in Indian language conditions before deployment. Public procurement should require disclosure of performance across relevant languages and dialects. Safety testing should include low resource languages, not only English. Education, healthcare, agriculture, law and welfare systems must not use AI tools that are linguistically unreliable for the very citizens they are meant to serve. This is not an argument against innovation. It is an argument for responsible innovation.
Private companies will naturally pursue scale and profit. That is expected. But India must define the terms of access to its market. Technology firms that benefit from Indian users should invest seriously in Indian languages, Indian evaluation systems and Indian safety standards. Localization cannot remain a promotional slogan. It must become a measurable obligation.
There is a precedent in education. Indian students are expected to learn multiple languages because multilingualism is part of national life. AI builders who want to operate at Indian scale should be held to a similar standard. They must learn India’s languages, not simply translate English into them.
This is where India can also offer leadership to the Global South. Many societies across Asia, Africa and Latin America face similar challenges: rich oral traditions, multilingual populations, limited digital text, uneven English access and high dependence on speech. If India builds AI for its own complexity, it can provide a model for other multilingual democracies.
But that requires intellectual confidence. India should not merely import the assumptions of Silicon Valley and adjust them at the margins. It should build from its own realities. The starting point must be the citizen, not the product demonstration.
Can a farmer ask a question in his own dialect and receive reliable advice? Can a student learn in her own language without being pushed toward inferior AI support? Can a patient describe symptoms naturally and be understood safely? Can a pensioner access welfare information without English? Can a small shopkeeper use AI through ordinary speech rather than formal prompts? These are the real tests of AI inclusion.
The Prime Minister’s vision of AI as empowerment will be judged not by summit declarations but by such everyday encounters. India’s AI future will not be decided only in data centres, startup offices or global boardrooms. It will be decided in the voices of ordinary people who expect technology to meet them where they are.
If AI speaks fluently to the elite but stumbles before the majority, it will not democratize India. It will only modernize exclusion. India does not need artificial intelligence that merely performs well in English. It needs artificial intelligence that can listen to the country in its own languages.
Until then, the promise of AI for all will remain incomplete.
Naorem Mohen is the Editor of Signpost News. Explore his views and opinion on X: @laimacha.