Introducing IndQA
OpenAI
OpenAI built a new test called IndQA to check how well AI handles Indian languages and culture. It matters because most AI benchmarks ignore the billion-plus people who don't think or speak in English.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI just dropped IndQA, and it's a quiet admission of something the industry has danced around for years: most AI benchmarks are built by English speakers, for English speakers, and they don't tell you much about how a model performs for the other 80 percent of the planet.
The new benchmark covers 12 languages spoken across India and spans 10 distinct knowledge areas, built in collaboration with people who actually know the subject matter rather than engineers guessing at what "cultural understanding" means. That distinction matters more than it sounds. A model can ace a translation task and still completely miss a reference to a regional festival, a colloquial idiom, or the specific way a legal concept gets discussed in Marathi versus how it gets discussed in Tamil.
India is not a monolith, obviously, and OpenAI's framing seems to get that. Testing across a dozen languages instead of treating "Indian language support" as one checkbox is a meaningfully different approach than what's typically shipped. It also signals where OpenAI thinks the next wave of users is coming from — India is already one of the largest markets for ChatGPT by user count, and a benchmark like this is as much a product roadmap as it is a research artifact.
What's notably absent from the announcement is any claim that current models do particularly well on IndQA. That's probably the point. Benchmarks like this tend to exist because the results are humbling, and publishing the test itself is a way of inviting the field to catch up rather than declaring victory prematurely.
My take — AI-written commentary, not fact-checked reporting
I'll believe the multilingual push is real when benchmarks like this start showing up as a routine part of every major model release, not a one-off blog post from a single lab. Right now it reads more like a PR move dressed up as infrastructure — useful, sure, but let's see if OpenAI actually publishes its own scores on IndQA before calling this progress.
Read more about this at: OpenAI