Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence
Authors: Kate Crawford | Year: 2021
AI is not immaterial: it is a planetary-scale infrastructure of extraction — of minerals, energy, data, and human labor — and understanding it requires tracing these supply chains rather than treating it as software floating in the cloud.
Core Ideas
Crawford’s central argument is that the dominant framing of AI as “intelligence” is a mystification. It obscures the physical reality: AI systems are constituted by lithium mined in Nevada, cobalt extracted in the Congo, electricity generated by fossil fuel plants, servers cooled with billions of liters of water, and data labeled by workers earning below minimum wage. Making this material geography visible is the book’s primary contribution.
Each chapter traces one layer of AI’s physical infrastructure. The Earth chapter follows raw material extraction: the rare earth minerals, lithium, and cobalt that go into the batteries and chips that AI runs on, tracing the environmental and community damage at extraction sites. The Labor chapter examines the hidden human work: the mechanical turk data-labeling pipelines, the content moderators who view traumatic material, the delivery workers whose routes are optimized by algorithmic systems that surveil and discipline them.
The Data chapter examines the politics of data collection: whose images, words, and behaviors are collected, under what terms of consent, and for whose benefit. Crawford shows that many foundational training datasets were assembled without meaningful consent and encode the biases and power relations of the contexts in which data was gathered. The classification of people by AI systems — for hiring, lending, criminal justice — reifies and amplifies these embedded structures.
The State chapter traces the relationship between AI development and military/surveillance infrastructure. Crawford argues that AI is not a neutral technology that can be applied to either good or bad ends; its development is entangled with specific governmental and corporate power interests that shape what problems get solved and whose interests are protected.
Significance
Crawford’s book is the most systematic account of AI’s political economy available. It provides the vocabulary for a materialist critique of AI that goes beyond algorithmic fairness (are the outputs equitable?) to ask structural questions about who benefits from the system as a whole and who bears its costs.
For AI practitioners, the book is a reminder that choices about what to optimize, what data to use, and how to deploy systems have consequences that extend far beyond the performance metrics on a benchmark. The computational and environmental costs of large-scale training are now widely discussed (e.g., energy consumption of GPT-4 training), and Crawford’s framing of AI as material infrastructure is a key reference for these conversations.
Related Papers
- Turing 1950 — the original question “can machines think?” that Crawford’s book puts in political and material context
- Kaplan et al. 2020 — scaling laws that quantify the compute and data hunger Crawford critiques as environmentally and politically consequential
- Sutton 2019 — “scale is all you need” as the engineering philosophy Crawford’s political ecology problematizes