Karya: Manu Chopra on How to Center Dignity

Ambika Samarthya-Howard: Tell me who you are and a little bit about your organization. 

Manu Chopra: I’m Manu Chopra with Karya. We use AI to give earning and learning opportunities in low-income communities across the global South. AI is creating a whole new wave of jobs. AI is going to create trillions of dollars of wealth for rich people. It’s important for us to find ways to bring those opportunities to our community. Otherwise, you have a repeat of what happened to the internet, where you have a very unequal way of how that technology is affecting all of us. 

We believe that AI is going to create a lot of these new jobs. Our goal is to bring those new AI jobs to people in our communities. In doing that, we do three things. First, we enable them to earn more supplementary income. Second, we’re able to teach them skills that can lead to more jobs in the future to be better prepared for an AI-enabled future. Third, and most importantly, in the process of employing our community to build AI models and train the AI of the future, we make sure that the AI of the future and present works well for them, so that they’re not just forgotten. Today, we have brought digital work opportunities to over 100,000 people.

Our goal is to bring those new AI jobs to people in our communities.

– Manu Chopra

Ambika Samarthya-Howard: How are you employing them? Can you take us through a specific project that explains what those 100,000 people are doing?

Manu Chopra: We bring work opportunities to our communities that come out of a funded project from Google or Microsoft or the government of India. Today, over 80% of rural Indian households have access to a smartphone, although women’s access to the device remains more curtailed. Google, Microsoft, and other companies have our workers do a series of jobs that are required to build these AI models. 

These jobs can be as simple as building an AI model in a local language, since no one speaks their own language incorrectly. Of course, to build an AI model, we need lots of training data, not personal data. Those training datasets need to be built in Indic languages, which have historically been marginalized in global work in the natural language processing space. Our communities sit in their homes on their phones, simply reading out stories in their mother tongue. That’s the easiest task we offer. 

For the simple task of reading out sentences in your mother tongue, we are able to pay them nearly 20 times the Indian minimum wage, which is a little over ₹450 Indian rupees an hour, and that’s where the work starts. From there, the work keeps getting more complex, and then we can pay our workers up to 40 times the local minimum wage. The work is everything from building AI models in local languages, to fine-tuning AI models, to reinforcement learning, to human in the loop, to evaluations, to benchmarking. 

At the core of Karya’s work is to identify communities that happen to be low income which can start doing this work. Second, we then upskill them to do more complex work so we can take them up the food chain. We always want to work where it’s extremely critical to our clients, and thus they’re willing to pay high wages for it, so that we’re able to create local impact. Specifically we look at work where we believe the technologies will be used by members of a community within the next few months, so it’s critical that these technologies work well for our communities. 

At the core of Karya is the fundamental belief that low-income communities are excellent beneficiaries of AI models. All the good AI work you see is around building a healthcare chatbot, an agriculture chatbot, an education chatbot, or this chatbot is really needed, and it can make a dent, but our communities are also excellent builders of these AI models, not just excellent beneficiaries. When we employ our communities to build these AI models, we enable economic and learning opportunities. The resulting technologies also get better and are more inclusive for these people. It’s not rocket science. If you want to build inclusive AI, you have to employ the communities you want to include. It’s literally as simple as that.

Our communities are also excellent builders of these AI models, not just excellent beneficiaries.

– Manu Chopra

We have a project with the Gates Foundation where we are engaging with 30,000 low-income women who are building the largest gender intentional AI corpora in Indic language history. The result of that corpora is going to be, hopefully, that AI models are less misogynistic, less sexist. That can only be possible because we engage with women’s voices in the process of building these AI models. The question is, can we create this win-win situation where our communities get earning and learning wages, and the tech gets more inclusive for all of us?

Ambika Samarthya-Howard: So you can afford to pay people to build these models when a philanthropy asks for a specific AI language model on gender, or maybe an AI language model in rural Bangalore, and then you build it out? Is that where the funding comes from?

Manu Chopra: Yes. We bring wages to workers mostly through clients. The Gates Foundation funding is the only case where a philanthropic organization is paying us to create economic opportunities for our communities. All but one [of our AI projects] is a client. A company like Google or Microsoft will pay us to build datasets. If they pay us $100, $75 on average goes to our communities, and the other $25 covers our costs, so we break even on every project. Then we use philanthropy either to improve our capacity, do more work, or to do things that the market won’t fund. 

For example, our work on gender, while extremely critical to the future of technology and the future of people using our technology, [does not attract] enough commercial interest, so we have to rely on philanthropic funding to pursue those ventures. Same for our work in language poverty, same for our work in low resource intake languages. Not the Hindi, Tamil, or Telugus of the world, but languages spoken by tribal communities. We’ve been able to create a lot of impact by employing these communities and open sourcing these datasets to facilitate AI development, while still paying workers the wages they deserve, thus creating a doubly positive effect.

Ambika Samarthya-Howard: Sometimes you’re working with people who might be direct competitors, and also working with philanthropies to build capacity. If Google asks you for a specific dataset, and then Microsoft asks you for that same dataset, are you hiring two different sets of people to build out competing datasets? How do you navigate the broader trends in the ecosystem?

Manu Chopra: It’s not been a concern, in all honesty, and it’s not because of some amazing thing we have cracked. That’s the way the sector has always been. These companies compete, yes, but they also collaborate very often, even though it may not seem that way. We are the only nonprofit doing this work with a lot of for-profit companies that do dataset services. It’s very normal, if you go to any of the websites, to see how Google, Microsoft, Apple, Facebook, and others work with everyone. That was almost a practice in the sector before we came in. 

Now, one of the twists in our models is that we give, in principle, ownership of our datasets to our communities. We don’t actually profit from resales of data, but our communities do. Let’s say Google paid us $100, and $75 went to the workers, plus $25 to cover our costs. This dataset that is worth $100 in the market is owned in principle by our communities, and it’ll be listed on our website for Google. In most likelihood, Microsoft will reach out and say, “That $100 dataset you built for Google, I want it.” Which is a perfectly fair thing to ask. Google knows we can sell that dataset to Microsoft, which is why it’s priced so low at $100. It’s priced to be sold multiple times. When Microsoft comes in to say, “We’re happy to pay $100 for this,” all $100 will go back to our communities as royalties. We don’t take a cut simply because we didn’t do anything. We already broke even on the previous transaction. This is how we’re able to give royalties and keep paying our communities for the work they’ve done in the past.

Ambika Samarthya-Howard: Both as a business person and someone working in AI, how did you figure out this particular approach for this model?

Manu Chopra: Again, we got lucky in that exclusive datasets and non-exclusive datasets have existed in the industry before we came in. Exclusive datasets are where Google says, “Nobody else can touch this data. I’m happy to pay you 10X because I’m covering the cost of you being able to sell this to other organizations.” 

In the AI sector specifically, right now we’re in such early stages that nobody’s really building exclusive datasets, especially in Indic languages, because we’re doing the building blocks, and nobody wants to own the building blocks. When it starts getting to domain-specific stuff, which we will get to in a few years, you may start getting increased demand for exclusive datasets. Right now, the sector is mostly non-exclusive. For-profit data companies really love non-exclusive datasets, because it’s a one-time investment, and then you keep making money. Of course, in the case of a for-profit company, they keep all of it. 

For us to operate as a nonprofit with a singular goal of bringing as much money to our communities as possible, it makes a lot of sense for us to do royalties because our workers built that dataset. It’s their labor. They should have in principle ownership of it. Our clients also like this because it allows them to feel good about the impact they’re creating in the general process of doing their business, a win-win situation. You can go to our website and see all the datasets that our communities have already built. You can buy any of them and 100% of whatever you pay goes straight to the communities. That allows us to keep engaging with the communities. 

Ambika Samarthya-Howard: Did you come to this work from the development or tech world? How did you get comfortable with this economic model?

Manu Chopra: My cofounders are Safiya Husain, our chief impact officer with 10 years of experience leading M&E [monitoring and evaluation] at a major education nonprofit called STAR Education before joining Karya two and a half years ago, and Vivek Seshadri, our chief technical officer. Vivek and I co-founded Karya as an idea at Microsoft Research in 2017. I had just graduated from Stanford where I studied AI, and moved back to India. Vivek had just done his PhD from Carnegie Mellon, also studied AI, and moved to Microsoft Research, where he was my boss. In November 2022, we spun out Karya and became an independent organization.

Ambika Samarthya-Howard: How important was the capacity-building support and the Leaders Studio from Rippleworks? 

Manu Chopra: It’s been amazing. Rippleworks has been one of the most dignified experiences we’ve had because they care deeply about being of service to the organizations they support, especially for an organization like ours, which is growing so fast. We were 25 people last year, we’ll be 120 people in two months, and 660 in a year. To run this organization requires traveling for fundraising or clients, because we are fundamentally a two-sided marketplace. We have to identify people in the communities, mobilize them, and build a proper operations and sales-facing team. 

Our clients, Google and Microsoft care deeply that we operate as a nonprofit and care about the impact. They also demand, and rightfully so, excellent datasets of the highest quality possible. There’s no quality discount you get by being a nonprofit. In fact, you have to work hard to prove that we can create impact and build high-quality datasets at the same time. We need world-class mentorship. 

What Rippleworks did was to put us in touch with someone who had run multiple large operations systems and built them from the ground up for Uber in India, and Snapdeal in India. We were able to work with him over a period of four or five months, and learn from his experience on building an operations team to identify and upskill a million low-income communities remotely at scale in India over the next three years. 

Despite our backgrounds, none of us have done that. There’s no training I’ve ever received to build the systems and processes in place. [It was helpful] to have someone hold us accountable and give us ideas. Because of the Rippleworks support, we launched Operations 2.0 at Karya, which allows us to go from 100,000 [people working for us in communities] to 1 million over the next few years. It’s been a tremendous success already, and will only become more successful in the future, and that happened because of Rippleworks. 

A lot of organizations give us financial help, which was, of course, so appreciated, especially as a young organization. Rippleworks provided us with the best mentorship we’ve ever received, because it connected us to someone who knew the Indian context and had run a large operations team in India. They came up with a system that was just the right amount of stress. It was just the right amount of work, and it allowed us to get the things done. It was a ‘10 out of 10’ experience, in all honesty.

Ambika Samarthya-Howard: Why do you call that relationship ‘dignified’? Were they good listeners, or is there something else about it that could be replicated?

Manu Chopra: Every interaction we had with everyone at Rippleworks has been one of complete respect and dignity. That stands out in an ecosystem that often does not have that, in all honesty. [I don’t want to] contrast this as better or worse to other funders, because we’ve never met a donor we have not enjoyed working with. What RippleWorks has succeeded at doing is building systems of dignity, through things like mentorship, by recognizing where we are coming from and yet pushing us on where we need to go to get to our goals. 

Ambika Samarthya-Howard: How do you know the new Ops system they helped you create has been working?

Manu Chopra: We track metrics like, how quickly is work happening? Is our unit cost improving? Is team satisfaction improving? We do eNPS [Employee Net Promoter Score] anonymously to get a sense of how comfortable people are feeling about the work they do and if they feel a sense of clarity or confidence in where the team is going. 

Ambika Samarthya-Howard: How do you view impact? How do you know when something is working? 

Manu Chopra: I would defer this question to Safiya [Husain], our chief impact officer. In general, we basically look at something called PERMA+4 [a framework for work-related wellbeing and sustainable work performance] to understand M&E at Karya. We’ve done two randomized controlled trials [RCTs] with J-PAL [a global research center working to reduce poverty by ensuring that policy is informed by scientific evidence], as well as multiple impact reports and research. Workers answer questions about their emotional well-being, and we track their satisfaction or dissatisfaction with their wages and the work they’re doing, as well as what they want to see in the app. We collaborate with research organizations that research labor rights to make sure our workers are having a dignified experience on the platform.

Ambika Samarthya-Howard: Are there any requirements or touch points you have with funders that are particularly helpful in promoting trust? When is it good to have some requirements?

Manu Chopra: I haven’t thought much about that, to be very honest. All the funders we’ve had have been very kind and understanding of just how busy we are in a year one or two organization, in cases where we are delayed in sending a quarterly impact report or something like that, which I really appreciate. We’re just now building a fundraising team and have hired a second fundraiser at Karya outside of me, as well as a philanthropic partnerships team. Karya is an organization where every week there’s 10 new things happening, and funders always ask how they can help us expand. I’ve been blown away by just how incredible the [funding] ecosystem is, from Gates, to Google, to Microsoft, to Google.org, to McGovern Foundation, and Rippleworks, these are ‘10 out of 10’ organizations with ‘10 out of 10’ people.

Ambika Samarthya-Howard: What hasn’t worked? What are some lessons you’ve learned both for yourself and the organization, and also in terms of funders?

Manu Chopra: There are lots of things that haven’t worked on the project side because lessons [happen] there on a daily basis. I certainly entered the space as a ‘techno-utopian,’ i.e. the sense that tech can solve many of our problems. I now call myself, as Kentaro Toyama puts it, a recovering technoholic. In Kentaro’s book “Geek Heresy,” he says that technology is an excellent amplifier of human intent and capacity, [but] not a substitute. That’s a lesson we had to learn. 

In the early days, especially in 2017 or 2018, we had a pilot and thought that since we have an Android app, people could just download it. We thought we didn’t need to tell everyone to come together and do the hard work of selecting who gets the work. [We thought we could just] work with the nonprofit on the ground, come up with some heuristic process to decide which person should work for Karya over someone else. 

Every village has a WhatsApp group these days, so we just put out a message in the middle of the day to say that the first 100 people who came in could do this work. [But] every single person who signed up was a man, and most of them had the same last name, because the people with access to phones in the middle of the day are privileged, and knowledge spreads through channels of power. What happened was one person saw the message and called his family members, all of whom were in upper-class communities. They were low-income, yes, and we had no problem with employing them, but obviously, we would have wanted those 100 people to be as diverse as possible. The resulting AI model works better when the datasets are diverse. 

That was a big mistake and we learned that’s not how to do mobilization. We have to mobilize through nonprofits, which doesn’t necessarily scale as well. We have to reach out to communities that need this work, not just those who want it. Then we had to come up with an access code system to link certain access codes to the gender requirements we had, to make sure we were working with at least 60% women. That was a very early lesson, and there are 1,000 lessons like this on the project side when you’re trying to make something work in our communities. How do you do it in a manner that meets the communities where they are, and gives them these opportunities in a way that improves their agency and not have an undignified experience? There are lots of lessons there. On the funder side, in all honesty, maybe we’re just too young as an organization to have lessons on what doesn’t work well with funders.

Ambika Samarthya-Howard: When was your first support from Rippleworks?

Manu Chopra: I think it was six months ago.

Ambika Samarthya-Howard: Do you have other funding besides Rippleworks and the Gates Foundation?

Manu Chopra: Gates was our first investor, 14 days after Karya started, with a $2.1 million grant. Google.org is the second biggest donor. They recently gave us a $1 million grant. We also have funding from the Patrick McGovern Foundation, Microsoft Philanthropies, LinkedIn, SVCF, Fast Forward, 100X, DRK, and Mulago. I’m sure I’m missing some, but those are the big ones.

Ambika Samarthya-Howard: How did you raise the gender balance and get more women to join?

Manu Chopra: India has so many incredible nonprofits. Indians are so aspirational, always willing to learn new things, and word of mouth helped. We just opened referrals on the app, and that’s been going well because every person refers 50 people.

Ambika Samarthya-Howard: You built a for-profit revenue model, but that hasn’t stopped you from getting grants from philanthropies. How do you sort that out? It’s rare for a nonprofit to also be for-profit, and give wages back. How do you think about that internally, and how do you communicate that externally?

Manu Chopra: Karya operates as a nonprofit. If we get $100 from the client and give workers $75 while we keep $25, donors love the fact that their grants are not going to workers as wages because that’s not the best use of their money. They love that their grants are used to improve our capacity to enable us to bring more wages to the market in our communities. They recognize how big the [AI] sector is. 

By some estimates, north of $17 billion a year is already being spent on AI datasets, [but] nothing of it comes to low-income communities from these very big organizations. [Donors] are very committed to help Karya get to a stage where we can bring a significant chunk of that [revenue] for wages to workers, just from the market. All the donors we work with care deeply about market or government as a pathway to scale. 

At Karya, we do both. Government is both a payer at scale for us, and also a doer at scale for us. The same is true with the market being a huge payer at scale, and [donors] really like this. They understand we need philanthropy, simply because there is so much to do. The fact we have a market model has been the biggest thing that’s helped us raise philanthropic capital. Philanthropists care deeply about a pathway to scale. We are doing this exercise with Bridgespan right now to work on a strategic roadmap. Every philanthropist I meet asks me, “For every dollar I give you, how many dollars go to the workers?” For them to invest in a model like Karya, you have to do better than that. They care deeply about the impact we can create with the wages we give to our communities.

Philanthropists care deeply about a pathway to scale.

– Manu Chopra

What second-order impact can we have by making these technologies more inclusive? Say, ChatGPT understands Telugu, understands Tamil, and understands the context of people to work. How does that help our communities? Has access to information changed things in meaningful ways? How does that happen? There is a lot of interest from our philanthropic donors in making sure we are able to use the market wherever it is willing to pay for wages for our workers. Where it is not, philanthropy is more than willing, especially on things like gender and language. Philanthropy is very willing to bridge that gap. 

Ambika Samarthya-Howard: Any insights that you would share, any advice you’d give to somebody doing this work?

Manu Chopra: The time has come for innovative philanthropic models. I’m working with one of our prominent donors on a book about bold philanthropy. Nonprofits should think deeply about what their pathway to scale is, whether it is through government or the market. Every donor would agree that we are entering a very difficult fundraising environment. There’s always been more need than money available for the work we all want to do. Whenever possible, if you’re able to get market models without compromising the integrity of the impact, that’s great. Nonprofits are rightfully incentivized by the right donors to focus entirely on impact. 

There are several data companies in the world that do $1 billion in revenue. Karya isn’t trying to prove that the data market is valuable, that’s been proven multiple times over. Despite these data companies making over $1 billion in revenue every year, the workers who are employed by these data companies can make as low as $0.30 an hour, and this has been well reported. Here, I’m very cynical. I feel they just simply do not have the right incentives in place, and if they can get away with it, they will. 

The solution cannot be to just entirely rely on the market, because wages in the AI labor market, for example, are a market failure. We have to counteract that. In general, I’m a big fan of models that combine the speed and scale of a for-profit with the thoughtful impact of a nonprofit. I love what Rocket Learning is doing in education. It’s a straight government pathway collaboration. I love what The Noora Project is doing with healthcare and the government as a pathway. 

There are lots of organizations I look up to which are on very clear pathways that aren’t reliant on raising $50 million every year, which is just very hard to do as a nonprofit in India. Our largest nonprofits are run with such integrity and such passion and such commitment, and they have very large teams, so it takes a lot of effort for them to raise the money they [need]. It’s very hard for young organizations to get to that level. 

In all bluntness, it’s easier to be a $20 million organization in the Bay area than it is to be a $20 million organization in India. Even within India, it’s much easier for people like me, privileged people sitting in Bangalore, to do this for incredible communities, communities and organizations on the ground. What are other pathways to scale that rely on government or market, while still building the philanthropic ecosystem here in the country?

Ambika Samarthya-Howard: From your point of view, what are three things you would say you need to scale and to sustain your work?

Manu Chopra: First is building the team that our communities deserve. That is priority number one this year. We’ve just hired a chief revenue officer, a director of operations, a director of people’s success, and other key hires, but we need to hire more people. Second is sales and getting bigger contracts from Google and Microsoft to prove to them that our communities are low-income communities, [but] not low-talent communities. We think of an AI builder as a fancy person sitting in SF or Bangalore or some tech hub somewhere coding, but the world employs 5 million data workers. Why aren’t these people considered data AI builders? Without their work, nothing we do can happen. Third is raising the money we need to do our work. Building AI models, building an AI organization, and being the only nonprofit in our sector is expensive, and we’ve been very lucky so far, but we have to keep constantly raising more philanthropic capital to improve our capacity so that we can get to a point where, hopefully, we will not need philanthropic capital at all in a few years.

Ambika Samarthya-Howard: Thank you, Manu. 

Manu Chopra: Thank you so much.

Ambika Samarthya-Howard (she/her) is the Solution Journalism Network’s Chief Innovation Officer: She leads on innovation and technology, leverages communication platforms for the network strategy and creates cool content. She has an MFA from Columbia’s Film School and has been creating, teaching and writing at the intersection of storytelling and social good for two decades. She has produced content for Current TV, UNICEF, Havas, Praekelt.org, UNICEF, UNFPA, Save the Children, FCDO, Global Integrity and Prism.

* This interview has been edited and condensed.

View More Interviews