The global footprint of generative AI

Recently, I came across the Cartography of Generative AI, a fabulous resource created by the Estampa group.

This resource shows how chatbots and other gen Ai interfaces are part of a dense, intricate and vast global system of physical resources, human labour, data flows, and economic structures. 

The right area of graph depicts the physical foundations of AI. Starting from the top, raw materials such as lithium, cobalt, copper, and rare earth elements are mined in countries such as the Democratic Republic of Congo or Chile. From there, the minerals travel across the world to be transformed into chips, graphics processing units and servers, in factories located in Taiwan, and represented in the bottom right hand corner of the graph.

The top left and middle area represent the data infrastructure. It depicts how text, images and other digital content are scraped from the internet and used in training data centres. It’s the back end of AI, and includes Big Data platforms, AI start-ups and other commercial entities that process and monetise the data. Much of this infrastructure is concentrated in the Global North, where access to capital and connectivity is greatest.

The middle left is the labour layer of generative AI. Developers and researchers in tech hubs like Silicon Valley, design, build and fine-tune AI systems. In turn, data labellers and annotators, usually based in low-income countries in the Global South, clean and classify data so that the AI models can learn to “see” and “understand”. Moreover, many others work as content moderators, reviewing and blocking violent content.

Finally, the bottom left and central area depict the users, that is the individuals and organisations who use generative AI for peripheral or core activities. The prompts, uploads, dialogues and reactions create new data that are fed back into the system, training the next generation of models. Thus, there is a self-reinforcing cycle whereby users are both consumers and producers in the generative AI system.

Estampa’s overview helps me see how generative AI products like ChatGPT, Midjourney and so on rely on a global web of data, materials and labour. It could inform interesting class discussions about how this interconnectedness shapes innovation, or how power dynamics might shape the future of these products. What do think about this resource?