AI in Indian languages · How Founders Use AI
Where does AI still get Indian languages wrong, and how do I check its answer?
Dialects, names, numbers and culture are where it slips. In AI4Bharat and Josh Talks' Voice of India test, OpenAI's transcription models got over 55% of words wrong in Indian speech, and some models missed two words in three in Bhojpuri and Maithili. Translation can flip meaning: in 2025 Meta's auto-translate turned a Kannada condolence message into one that declared Karnataka's chief minister dead. To check, translate the answer back into English with a second tool and compare, look hard at names, dates and amounts in lakh and crore, and ask the assistant which words it was unsure of. For anything legal, medical or going out in public, have a fluent speaker read it first.
On the AI map, 29 October 2026 Open-weight modelsAvatars and voice
Go deeper8 sources
-
OpenAI, Meta speech models struggle with Indian languages
The Hindu BusinessLine · Article · Feb 2026
Reports AI4Bharat and Josh Talks' Voice of India test of speech recognition across 15 languages and many dialects, with error rates by company.
-
Global AI Speech Recognition Models Fail Indian Languages, Impacting Critical Services
OECD AI Incidents Monitor · Article · Feb 2026
The OECD's record of the Voice of India findings as a risk to welfare and health services.
-
Karnataka CM Siddaramaiah Calls Out Meta For 'Faulty' Kannada Translation
Inc42 · Article · Jul 2025
A real case of auto-translation reversing the meaning of a Kannada message, and the complaint that followed.
-
Introducing IndQA
OpenAI · Article · Nov 2025
OpenAI's test of 2,278 questions on Indian culture in 12 languages including Hinglish, written by 261 Indian experts, and why it is not a language leaderboard.
-
Script Gap: Evaluating LLM Triage on Indian Languages in Native vs Romanized Scripts in a Real World Setting
arXiv (research paper) · Article · Dec 2025
A study of real health messages in five Indian languages and Nepali, finding models handle romanized text up to 24 points worse than native script.
-
Indi-RomCoM: Code-Mixed Benchmark for Evaluating LLMs on Romanized Indic-English Instructions
arXiv (research paper) · Article · Jun 2026
A 2026 benchmark showing models follow instructions less well as messages mix more English into Indian languages written in English letters.
-
AI4Bharat at IIT Madras and Josh Talks AI Launches India's First Independent AI Evaluation Platform
AIM Network · Video · Aug 2026
Mitesh Khapra of AI4Bharat and the Josh Talks AI team on why AI has to be tested on real Indian accents, dialects and uses.
All 8 sources
-
Multilingual support
Anthropic · Guide · 2026
Anthropic's scores for Claude in Hindi, Bengali and other languages against English, with three tips: name the languages, use the native script, mind the culture.