Can AI Speak India’s Languages or Just Translate Them?

In 2023, Sam Altman said AI was improving rapidly in non-English languages and predicted that multilingual systems would get much better. India has since expanded its language-AI infrastructure. But the harder test is not whether a model can translate a sentence—it is whether people can reliably use AI in the languages, dialects and contexts that shape their lives.

Contents

The prediction was directionally right—but language support is not a single capability
India is building language infrastructure, not just language apps
The difficult part: dialects, code-switching and local meaning
Translation is not the same as understanding
What would meaningful progress look like?
The opportunity for India is larger than translation
The verdict: the infrastructure is advancing; universal language access remains unfinished
AI, Revisited: The series

When Sam Altman met Satyan Gajwani for an Economic Times conversation in June 2023, the discussion turned to a question that matters especially in India: who benefits from AI when much of the digital world—and much of the data used to train AI systems—has historically been concentrated in English?

Gajwani asked about bias, inclusion and the challenge of building products that work beyond the country’s largest metropolitan centres. Altman said OpenAI had made a significant step forward from GPT-3.5 to GPT-4 in non-English languages. He described GPT-4 as “pretty good” at roughly the top 20 languages and “okay” at perhaps the top 100, while acknowledging that languages spoken by much smaller populations were more difficult.

He also made a broader point: language alone was not enough. History, culture and values mattered too, and systems needed to represent people across the world.

Read the original Economic Times interview.

Three years later, the question is no longer whether multilingual AI is being developed. It is whether the technology is becoming dependable enough to make digital services, knowledge and opportunity genuinely more accessible across India.

The prediction was directionally right—but language support is not a single capability

AI systems have become more capable at working across languages, and language technology now covers far more than text translation. Systems can transcribe speech, convert speech into text, translate between languages, generate spoken responses, read text in images and support voice-based interfaces.

These capabilities matter because language is often the first interface between a person and a service. A citizen should not need to be fluent in English to understand a government scheme. A farmer should be able to access useful information in a language they are comfortable using. A small business should be able to interact with digital tools without translating every question into English first.

But “supports a language” can mean many different things. A system may translate a short, formal sentence reasonably well and still struggle with a noisy phone call, an unfamiliar accent, code-switching, a regional expression or a question that depends on local context.

Speech recognition, translation, reasoning, summarisation and voice generation are distinct tasks. Success at one does not guarantee success at the others. Nor does a model’s ability to produce fluent sentences prove that the information it provides is correct.

The first lesson: multilingual AI should be evaluated by the tasks people need to complete—not simply by the number of languages listed on a product page.

India is building language infrastructure, not just language apps

One of the most significant developments since the 2023 conversation is the expansion of India’s public language-technology infrastructure.

BHASHINI, the government’s National Language Translation Mission, provides language services, models, datasets and APIs intended to help developers and institutions build multilingual digital experiences. Its capabilities include automatic speech recognition, machine translation, text-to-speech, optical character recognition and transliteration.

In March 2026, the Press Information Bureau described BHASHINI as a national language-technology platform and reported that its ecosystem included more than 350 models, with support for use cases such as multilingual conversational assistants, citizen-service platforms, governance interfaces and knowledge-access systems. These are government-reported platform figures, not an independent assessment of accuracy across every language or task.

Read the government’s March 2026 update on BHASHINI.

This is an important shift. Instead of every organisation having to build the entire language stack from scratch, a shared platform can make underlying services available for integration into websites, applications and public services.

That creates a path for multilingual technology to move from demonstrations into everyday workflows. But infrastructure is an enabling condition—not proof that every service built on top of it works well for every user.

The difficult part: dialects, code-switching and local meaning

India’s linguistic reality is more complex than a list of official languages. People move between languages in the same conversation. They use regional vocabulary, dialects, local pronunciations and expressions whose meaning depends on context. Many communities also speak languages and dialects that have limited digital text, recorded speech or professionally annotated training data.

These differences can matter in consequential settings. A mistranslated instruction in a casual conversation may be inconvenient. A misunderstood symptom, legal question, financial instruction or public-service request can have more serious consequences.

Research is making these gaps easier to examine. The MILU benchmark, published in the proceedings of NAACL 2025, evaluates language models across 11 Indic languages and eight domains, including culturally specific knowledge such as local history, arts, festivals and laws. Its purpose is to test capabilities that general-purpose, English-heavy evaluations may not capture.

Newer research is also looking at dialect-level performance. A January 2026 preprint introducing the INDIC-DIALECT benchmark examined dialect classification and translation in Hindi and Odia dialects. The researchers reported that several tested large models performed poorly on dialect classification, while approaches adapted to the task improved results.

This is a specific research finding, not proof that all models fail on all Indian dialects. But it illustrates why headline claims about multilingual ability can conceal uneven performance beneath the surface.

The broader point is straightforward: standard-language competence cannot be assumed to extend to every dialect, speaking style or real-world conversation.

Translation is not the same as understanding

A system can translate words and still miss what a person means.

Consider a citizen asking a question about a government benefit using local terminology, an abbreviated phrase or a mix of Kannada and English. The system may recognise the words but misunderstand the underlying request. It may then produce a fluent, confident answer that does not address the person’s actual situation.

The same problem can appear in healthcare, education, agriculture, banking and customer service. Language is not merely a wrapper around information. It carries context, assumptions, relationships and local knowledge.

Altman recognised this in 2023 when he said the challenge involved not just language, but history, culture and values. That observation remains relevant: localisation requires more than converting English output into another script.

For AI to be useful, it needs to understand the task, retrieve reliable information, handle ambiguity and make uncertainty visible. In high-stakes settings, users also need a clear path to a human expert or authoritative source.

What would meaningful progress look like?

Counting supported languages and model releases can show that infrastructure is expanding. To determine whether that expansion is improving people’s lives, the next set of measures should focus on outcomes.

Task-level accuracy: Does the system correctly complete real tasks—such as understanding a service request, transcribing a conversation or explaining a document—in each supported language?
Real-world speech: Does it handle background noise, regional accents, different speaking speeds and code-switching, rather than only clean, scripted audio?
Dialect and community coverage: Are smaller language communities represented in data collection, evaluation and product design?
Meaning and factual reliability: Does the system preserve the intended meaning, and are its answers grounded in trustworthy information?
Safety and recourse: Can users recognise uncertainty, correct errors and reach a human when the consequences of a mistake are significant?
Independent evaluation: Are performance claims supported by transparent testing across languages, tasks and user groups—not only by aggregate scores?

These measures would shift the debate from whether AI can produce output in a language to whether it can help someone accomplish something accurately and safely in that language.

The opportunity for India is larger than translation

Reliable multilingual AI could change how people access education, public services, healthcare information, financial tools and digital commerce. It could also create opportunities for Indian developers, researchers, language specialists and businesses to build products around local needs.

Shared infrastructure such as BHASHINI can lower some barriers to building these services. Research benchmarks and community-led data efforts can help reveal where systems fall short. Public and private institutions can then test applications against the needs of the people expected to use them.

But scale should not be confused with inclusion. A platform can process millions of requests and still deliver uneven quality across languages or use cases. A service can offer many language options and still leave users struggling to complete the task that brought them there.

The goal is not simply to make AI speak more Indian languages. It is to ensure that language does not determine who can use AI effectively—and who is left with a fluent but unreliable substitute.

The verdict: the infrastructure is advancing; universal language access remains unfinished

Altman’s 2023 prediction was directionally right: multilingual AI has advanced, and India is investing in shared language-technology infrastructure. Speech, translation and other language services are increasingly available to developers and institutions, while research is becoming more specific about how to evaluate performance across Indic languages and dialects.

Yet the evidence does not justify declaring the problem solved. Support for a language is not the same as dependable performance in that language. Translation is not the same as cultural understanding. And a fluent answer is not necessarily a correct one.

The next phase should be judged less by how many languages a system claims to support and more by what people can reliably do with it: understand a benefit, access a service, learn a concept, ask a question or complete a task in the language they actually use.

The real test of multilingual AI is not whether technology can speak to India. It is whether India’s people can use that technology—and trust the result.

AI, Revisited: The series

AI, Revisited returns to predictions and warnings made during the early public adoption of generative AI, then tests them against evidence and developments that followed. The aim is to separate what has materialised, what remains uncertain and what the original conversation may have missed. Explore our wider AI coverage on FutureIsNow.