Democratising AI needs a radically different approach

‘We cannot change as long as we keep viewing Big Data and deep learning as the holy grail’ 
| Photo Credit: Getty Images/iStockphoto

The growing power and influence of Big Tech companies is a concern for policymakers worldwide. To break Big Tech’s hold over the Artificial Intelligence (AI) ecosystem and democratise AI development, India, like many other countries, is investing in sovereign cloud infrastructure, creating open data platforms and supporting local start-ups. However, these efforts are unlikely to be enough and may even deepen Big Tech’s dominance.

Challenges of Big Tech dominance

The enormous computational costs of building deep learning models make it nearly impossible for smaller players to compete. Deep learning is now the popular form of AI because it has generalised capabilities. But this is precisely also what makes it computationally expensive. As of 2023, Gemini Ultra was the most costly model, costing about $200 million to train. To make it affordable, any new entrant would necessarily be beholden to Big Tech companies for compute credits. These costs also incentivise Big Tech companies to keep advocating for deep learning as the future and pushing out larger and larger models — it locks in their position as the dominant actors and provides the primary revenue stream through which they can recover their costs.

Some recent policy proposals suggest investing in public compute infrastructure or developing a federated model, taking a leaf out of India’s Digital Public Infrastructure model. However, it is not enough just to provide alternate infrastructure. This infrastructure also has to be competitive with the Big Tech offering. Big Tech companies offer a wide range of developer tools which make workflows easier and more efficient, and these tools are optimised for their cloud infrastructure. Along with access to cloud infrastructure, they give companies access to the latest algorithmic models, making tasks such as image or video analysis far easier, along with tools to simplify data preparation and labelling. Big Tech’s end-to-end service offering makes development cheaper and easier and drives up the costs of switching to other providers.

Big Tech’s data monopoly is even harder to contend with. These companies access a continuous data stream across various domains, social interactions and geographies. This “data intelligence” is likely to be more sophisticated than what other players can achieve, giving them a substantial competitive edge. Unsurprisingly, many smaller AI companies find their end game is to sell to Big Tech, further entrenching the cycle of dominance. While public data initiatives aim to democratise data access and create a more equitable playing field, they often fall short. Open data initiatives are prone to commercial capture, where the better-resourced actors — here, Big Tech with its advanced computational infrastructure and data intelligence — are positioned to best leverage these open data architectures.

The shift toward deep learning as the most popular form of AI has also meant that commercial firms, particularly Big Tech, now dominate AI, and academia has a diminishing role. Industry players now have more academic publications and citations and are shaping the direction of AI research.

Prioritising a theory of change

We need a radically different approach to AI development that does not aim to compete with or replicate the Big Tech model but changes the rules of the game altogether. As long as we are locked into a ‘big-data’ and ‘larger is better’ imagination of AI, we will only keep chipping away at an exploitative model of commercial surveillance and even a wasting of precious public resources.

A model of AI development whose starting point is a theory of change, i.e., understanding the causal mechanisms through which various factors link together and developing hypotheses about how potential interventions may contribute to change.

In this model, domain expertise and lived experience guide AI development rather than statistical patterns in Big Data alone. This knowledge and experience are harnessed to develop theories of change and build purpose-driven and smaller models that reflect frameworks for progressive change. Data collection is then targeted and curated to test and further refine the theory of change. By championing “small AI”, firmly anchored in a theory of change, we can carve out a space for AI development that is inherently more democratic and effective.

Historically, significant advancements in medicine, aviation, or weather forecasting typically relied on theory-driven models, where hypothesis testing and scientific rigour in fields such as biology, physics, and chemistry were prioritised over sheer volumes of data. In our obsession with ‘bigger is better’ we seem to have forgotten this entirely.

Another missed opportunity

We need to change course urgently, and we cannot do that as long as we keep viewing Big Data and deep learning as the holy grail. On this current path, we only increase our dependence on Big Tech. The recently signed Global Development Compact is a missed opportunity to re-think the current paradigm. While it makes all the right noises about democratising AI, it ultimately falls back into the same trap of assuming that if countries build large enough data sets and are given access to computational power, we will magically be able to achieve the Sustainable Development Goals and address Big Tech monopolies.

Urvashi Aneja is the Founder and Director of Digital Futures Lab