Karya Pays Rural Indians up to 20 Times Minimum Wage to Build AI in Their Own Languages

Most of the data that trains artificial intelligence is in English, collected by low paid gig workers who rarely see the value their work creates. An Indian nonprofit called Karya flips that model. It pays rural workers, most of them women, about $5 per hour to record speech and label data in languages like Kannada, Odiya and Marathi, roughly 20 times India's legal minimum wage. The work matters because AI largely fails in the thousands of languages spoken across India, and it pays enough to change lives in villages where a few thousand rupees can be several months of income.

Karya began in 2017 as a research project inside Microsoft Research in Bengaluru. In 2021, Manu Chopra, who grew up in Delhi's Shakur Basti neighbourhood before studying at Stanford, teamed up with Microsoft researcher Vivek Seshadri and civil society leader Safiya Husain to spin it off as an independent nonprofit. Their mission was straightforward: channel the value of AI data work directly into low income communities instead of extracting it from them.

Two Problems, One App

India is home to hundreds of millions of people who speak languages that AI models barely understand. Speech recognition, translation and voice assistants trained mostly on English simply do not work for them. At the same time, rural poverty remains deep and persistent. Karya attacks both problems with a single smartphone app designed for people with low literacy and unreliable connectivity, much of it usable offline.

The tasks are simple by design. A worker opens the app, records herself reading sentences in her mother tongue, transcribes short clips, or labels images. Karya packages this labeled data and sells it to major technology companies and researchers who need it to build Indian language AI. What makes Karya different is what happens to the money. Workers are paid far above typical data labeling rates, and when a dataset is resold, the workers who created it receive a share of the proceeds. The people whose voices train the models keep owning a piece of their contribution.

What the Money Means in a Village

The scale is still modest by tech industry standards, but the impact per person is dramatic. According to TIME, Karya has paid out about 65 million rupees, close to $800,000, to some 30,000 rural workers. More than 32,000 people across over 25 Indian states have used the platform, roughly 60 percent of them women, completing more than 35 million AI training tasks.

Individual stories make the numbers concrete. Preethi, a 22-year-old tailor in Karnataka, earned 4,500 rupees in three days reading Kannada sentences aloud, about four times her typical monthly income. In Raghurajpur, a craft village in Odisha, artisan Swarnalata Nayak earned about 4,000 rupees in a week simply by speaking Odiya into her phone, money she put straight into her children's school fees. No new hardware, no training centre, no migration to a city. Just a smartphone and a mother tongue.

Backing From Big Tech and Philanthropy

Karya's model has attracted serious partners. Microsoft uses Karya to source Indian language speech data. Google is working with the organization to build AI models for dozens of Indian dialects, and Google.org awarded it a $1 million grant to improve local AI models and develop a multilingual chatbot for rural users. The Gates Foundation funds projects that use Karya data to detect and reduce bias in AI systems. Karya has delivered over 10,000 hours of curated speech and text data to clients, and its stated goal is to reach 100 million people by 2030.

The organization runs on Microsoft Azure and uses Azure AI services to validate data quality, checking recordings and labels before they reach clients. That quality focus is part of the business case: data collected by native speakers who are paid well and trained briefly turns out to be better than data scraped or sourced at the cheapest possible rate. Fairness, in this model, is not a cost. It is the product.

A Different Answer to Who AI Pays

Economists expect AI to generate enormous wealth over the coming decades, and the standard critique is that the gains will flow to capital and to a small group of highly skilled workers. Karya offers a small but working counterexample. Chopra estimates that roughly $1,500 in savings can lift an Indian worker into the middle class, a threshold that might otherwise take generations to cross. Wages from the app have gone into school fees, home repairs, seeds for the next planting season, and savings accounts.

There are honest limits. Data work is task based and fluctuates with client demand, so it is income rather than a salaried career. Reaching 100 million people will require scaling far beyond anything Karya has done so far. But the core idea, that the people whose languages and voices make AI useful should be paid as its builders rather than mined as its raw material, is proving itself one village at a time. A woman in Odisha speaks Odiya into a phone she already owns, a voice assistant somewhere learns to understand her, and her daughter's school fees get paid. That is what redistributing the gains of AI looks like in practice.