`Don't use Silicon Valley as a template, India should rethink what AI should be'

Karen Hao, author of the New York Times bestseller, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI, has spent years looking at the human and environmental costs behind Silicon Valley’s rush to build AI infrastructure and models. She was in India recently for Synapse, a society and technology conference organised by curator Shoma Chaudhury, where she spoke to Sujit John on the "imperial" nature of big tech, and why India must forge its own AI path, focusing on small models and affordable infrastructure.

India is beginning to see this big build-out of AI data centre infrastructure. How do you see it?

What I have seen with a lot of countries that have this catch up mentality is, they're looking at the Silicon Valley model and saying we need to catch up to that. And that in and of itself is what I critique as imperial. You're using the idea that's coming from the empire and simply taking that as the template for you. Instead, you could just fundamentally rethink what AI should be for this country, and I went to a lot of sessions during the AI Summit, they were about open source models, about smaller models and there were people who brought up these really good points that for AI to work at scale in India, it actually cannot be large language models (LLMs) because it needs to run very cheaply, it needs to be able to run without internet on someone's mobile device so that a farmer can detect diseases for their crops, or doctors in rural areas can actually use this type of technology. If India thought from the ground up from that perspective, to catch up in the AI era could look completely different, where India already has all the ingredients that they need, a fantastic base of talent, the infrastructure capabilities for the smaller models, the high quality data for training these application-specific AI technologies. And then you wouldn't need these huge partnerships (with American companies) anymore and you wouldn't be actually giving away your sovereignty.

Are there major countries that are taking this small-model approach? What is China doing?

China's actually doing a lot of smaller models, because they're very compute constrained. And also China has a very different approach to AI development than Silicon Valley that has this mentality of just pushing technical advancement for technical advancement's sake. And now Silicon Valley’s run into this issue where they don't have product-market fit. So they're trying to find ways to convince people to buy their products. During the AI Summit, Brad Smith (Microsoft president) gave a keynote and he used this very specific turn of phrase: Govts need to help us generate demand for our technologies, he actually said out loud what is usually unspoken. Chinese companies have a different mentality. They can't afford to build technologies that no one's going to use because the venture capital model and the investment model are very different. VCs are not willing to wait years and years for their money to be returned. So they're much more thinking, what are the applications that meet our users where they are? And oftentimes those applications do not require the scale that is being built by the American companies.

But they are also building large language models…

They're also building it, but actually with significantly less computational cost. That's what we saw with DeepSeek, where it has the same capabilities but significantly cheaper to build and to run. And that's also why there are a lot of companies in the US now – not model developers but consumers of AI technologies – that are using Chinese models instead of the models from Silicon Valley because they're the same quality or sometimes better and they're just cheaper. I’m not saying everyone should be using Chinese models, but this is an example of how a country and the companies actually rethought from the bottom up what are the models that are going to work for us. And I think India has the potential to do the same thing.

In India, data annotation work that’s often done by very poor people, mostly women, tends to be presented as work that enables them to improve their conditions of life. But you have criticised it, characterised it as a form of "modern-day imperialism".

What I show in my book is time and time again, as the tech industry's appetite for resources accelerates exponentially, where do they go looking for those resources? Whether it's human resources, like the labour that they need to train their models, or physical resources, like the minerals that they extract, or energy resources, they always go to the poorest communities. India is a huge base for the labour that is supplied to these companies. And the way that they treat that labour is horrible. There was an investigation that just came out in The Guardian about women in India, from the poorest neighbourhoods that are being roped into doing content moderation for pornography, for child sexual abuse material. Because we are now talking about video generation models that can generate these kinds of things, companies that want to prevent their models from generating these kinds of things, are building content moderation filters that are being trained by humans in these extremely poor contexts. And when the women protested and said, this is affecting my well-being, my family's well-being, and breaking down my communities, because I was never told that this was part of the job, they're being told, you were told your job is data annotation, and this is data annotation.

What’s the answer to such content moderation issues?

The reason why there is all this harmful content that needs to be filtered out is because these companies are training their models on the entire internet. That's why there's harmful content. If you were talking about very application specific technologies like helping a farmer detect disease in their crops, why are you then training on pornography? That's not going to help detect the crops. You would just take curated photos of different types of diseases on different types of crops and train a model so you don't ever have to deal with content moderation. But because the companies take the opposite approach, they just hoover up everything, and then they expect people in the poorest neighbourhoods to be human shields for all of their users.