A look at Indian nonprofit Karya, which sells AI training data and redirects all the profit to its workers, who retain ownership of the data they create
I hope Karya's team keep the balance as they grow in the business world! — https://time.com/... #aiethics #aifairness — [images] Bluesky: Margaret Mitchell / @mmitchell.bsky.social : Really beautiful piece from @billyperrigo.bsky.social concerning some of *the most important* variables in AI right now: Data labeling and compensation. https://time.com/... Twitter: @milagrosmiceli : Karya tries to debunk this myth by paying $5/hour (20-30x local minimum wage). This is great in comparison to other BPOs but incredibly cheap for clients who concentrate wealth and power in the AI industry. So the question is: if not Karya, who profits from these workers' labor? Arvind Narayanan / @random_walker : @Sahasrangsu_G There is a detailed explanation in the article. Indian law prevents them from being a nonprofit and doing what they do; so they've set up two companies, one for profit and one nonprofit, which (AFAICT) together function essentially as a nonprofit. @asyoulikeit78 : Sceptical bc of my years of exp with tech innovators, big and small, using rural India as experiment fuel. But hopeful seeing a startup founder in India acknowledge caste / limitations of tech-solutionism for a change! Curious how the data ownership contracts work though. 🧐 Peter Lee / @peteratmsr : How to make AI data work more fair to people and a more effective path out of poverty? The nonprofit, Karya (a spinout from @MSFTResearch) in India, is gaining early but real traction with its creative, socially-aware, approach. Great coverage by @TIME https://time.com/... @milagrosmiceli : Finally, I want to highlight that AI is always “BY the people.” But before claiming that it is also “FOR the people,” we need to ask who owns the technologies that are fuelled by data workers. The answer probably doesn't include “the people.” Billy Perrigo / @billyperrigo : The workers behind AI rarely reap its rewards. One nonprofit in India is trying to fix that. After years of reporting on the darker side of the AI industry, I traveled to India with one question. Could its alternative model really work? https://time.com/... [image] Dr Meena Kandasamy / @meenakandasamy : This is tech journalism that covers AI from the ground up, when we are so used to thinking about it as policy trickle, or a top-down approach. Great read @milagrosmiceli : This isn't philanthropy. The AI industry needs this labor as much as impoverished communities need jobs. So the question remains: who produces the models that require these workers' expertise and who profits from the “low-skill” narrative? @milagrosmiceli : While companies predate impoverished communities, the fact that they do so to get specific expertise (e.g. language-related) at low prices remains undiscussed. Data workers' unique expertise is vital for AI. Framing their work as low skill serves to justify miserable wages. Nikhil Pahwa / @nixxin : People have forgotten this: in 2008, @GoogleIndia launched a toll free number, which people could call and get responses to search queries, provided by a human on the other end of the line. Google used it to train voice to text for Indian accents. Shut down in 2010. F Arvind Narayanan / @random_walker : In India, a shockingly different approach to data work: -nonprofit -pays 20-30x minimum wage -workers retain ownership of data they create (!) -helps build AI for their mother tongue, benefiting locals The challenge, of course, is signing up AI companies. https://time.com/... Arvind Narayanan / @random_walker : The big question is how to put pressure on AI companies to source their data work ethically. What levers do we have? LinkedIn: Deepak Seth : AI's Transformational Impact at the “Bottom of the Pyramid” — A few weeks ago in a webcast talking about Future of AI (https://www.linkedin.com …
Context & Ripple Effects
Data labeling has run on a cost-minimization model for a decade: Samasource's Nairobi operation was doing this work for Google and Microsoft back in 2018, and the pattern since has been scale plus opacity — Scale AI's Remotasks subsidiary outsources roughly 240K annotators while the workers themselves stay invisible.
The low-water mark came with TIME's investigation into OpenAI paying Kenya-based Sama $2/hour to label toxic content for ChatGPT. Karya inverts that structure: it sells training data commercially but routes all profit back to workers, who keep ownership of what they produce, at $5/hour — 20-30x local minimum wage yet still cheap for clients. The catch, per its own team, is signing up AI companies willing to pay for ethical sourcing.
First-order effects
- Karya's workers gain something no major labeling contractor offers — data ownership plus profit redirection — while Karya itself must convert ethics-conscious buyers into actual contracts against rivals competing on price.
Second-order effects
- Buyers like OpenAI and Scale AI face a pricing-and-reputation benchmark: Karya's $5/hour rate makes existing arrangements harder to defend publicly, forcing either wage convergence or quiet avoidance of scrutiny.
Third-order effects
- If worker-owned data cooperatives prove commercially viable, the annotation layer of the AI supply chain could split into an ethically sourced premium tier and a commodity tier — with regulation or procurement standards deciding which one wins volume.
The trend: AI's data supply chain is being forced to choose between cost-minimized outsourcing and worker-compensated sourcing as labor conditions move from hidden cost line to reputational variable.