When Prime Minister Narendra Modi hosted world leaders and tech chiefs earlier this year in New Delhi, he declared that AI must be “democratized” and be a “medium of inclusion and empowerment, especially across the Global South.”
It is a convenient vision for Silicon Valley, which is in the midst of an ongoing landgrab for the lucrative market. Young, tech-savvy and mobile first, India has become one of the most important growth regions for AI, ranking behind only the US in usage for both OpenAI’s ChatGPT and Anthropic’s Claude.
But the key to both ‘inclusion’ and the business dreams of tech diffusion is overcoming barriers to speech. India has nearly two dozen official languages and more than a hundred dialects. If AI cannot close this gap, it will just become another technology that divides the English-speaking elite and everyone else.
True localization will depend on whether models can comprehend Bengali voice notes, Gujarati payment queries and code-switched Hindi-English business calls—all the messy and real-world spoken words that drive daily commerce and public life.
Over a billion people speak Indic languages. Yet one study found that GPT5 only achieved about 45% accuracy on a human-curated benchmark covering 11 of them. First generation AI tools were trained on internet text, the majority of which is in English.
Technical improvements and better data-sets have helped recent models improve in so-called ‘low-resource languages,’ those with less data to train on. But the language gap persists, especially for speech, forecast to become the next mass way of interacting with models.
“Voice is the most intuitive interface for humans, especially in more developing regions,” Sandeep Chinchali, co-founder of Poseidon, an Andreessen Horowitz-backed data infrastructure startup, told me. And South Asia, he added “uses voice for everything.” AI systems that cannot comprehend these interactions will be useless in automating this work, not to mention potentially dangerous in public services.
One problem is a lack of proper benchmarks for non-English models. Leading ones, for example, cannot even agree on what proper Bengali is.
Spoken Indic languages add another layer of difficulty: regional variants, background noise and frequent code-switching in technical and financial conversations.
Speech data for AI training requires accurate transcription, longer clips, varied acoustic environments, demographic and regional variants, as well as careful human review before they can truly improve AI models.
Systems trained on narrower data-sets often fail in the real world, where conversations mix local slang and borrowed English words in varied settings.
India Inc knows the stakes. Cracking the language challenge has become central to the country’s broader sovereign AI push. Building AI that works “at India’s scale” presents a massive opportunity, Pratyush Kumar, co-founder of domestic AI hopeful Sarvam AI, said in a statement this month.
In April, the startup so many are pinning their hopes on for catching up in a US-China race launched a new evaluation for Indic speech recognition, arguing that the standard metrics were not built for these languages and can distort how such systems are judged.
US tech giants are paying attention too. OpenAI unveiled a framework to evaluate AI systems on Indian culture and language. And the Indian government launched a translation platform that collects spoken data to improve multilingual models.
But crowdsourcing is no cure-all. High standards and human curation remain vital. A Stanford team warned in a paper last year that quality has become a key challenge when trying to scale such endeavours. It also raised ethical questions in a sector with a long history of poor pay and exploitation.
The government is already forcing high school students to learn three languages, including two indigenous ones. If the government is serious about making India an AI superpower, it should demand something similar from model builders and craft policy that ensures systems that can truly comprehend its linguistic diversity.
There is also a safety issue. As AI moves into schools, hospitals, courts and public service, language failures have consequences. Safety alignment tends to deteriorate when people engage with AI in low-resource languages. So the people most likely to be left behind by the revolution may also be the least protected from its risks.
Bridging the language divide is key to US tech cracking this market. And Modi’s promise that it will deliver “happiness for all, welfare for all” may ring hollow if the technology can’t understand the people it claims to empower.©Bloomberg
The author is a Bloomberg Opinion columnist covering Asia tech.