Visual summary of operating lessons from Kate Crawford.

Kate Crawford studies the social and political implications of artificial intelligence. In Atlas of AI, she examines the resources, labor and data behind these systems and asks how their construction and use concentrate power. — Sydney Ideas: Atlas of AI.

Part 1: The Materiality of Artificial Intelligence

  1. On the nature of AI: Crawford treats AI as a material industry, built from minerals, energy, water, labor and data—not as disembodied intelligence in a cloud. — Renewable Matter: Calculating and Powers.
  2. On the illusion of the cloud: The cloud metaphor can hide the places, institutions and choices that make computation possible. Crawford argues that this abstraction obscures political responsibility. — Big Tech: Kate Crawford Interview.
  3. On redefining intelligence: She favors describing these systems as complex information processing. That framing avoids treating statistical computation as equivalent to embodied human intelligence. — Renewable Matter: Calculating and Powers.
  4. On probabilistic statistics: Her critique of enchanted determinism targets a contradiction: systems portrayed as mysterious or magical are also trusted to produce reliably certain answers. — Renewable Matter: Calculating and Powers.
  5. On magical thinking: Presenting AI as otherworldly can divert attention from its labor and planetary costs. She treats that framing as a political choice. — Big Tech: Kate Crawford Interview.
  6. On avoiding scrutiny: Crawford argues that abstract imagery distances AI from the energy, labor and capital behind it. Investigating those connections makes scrutiny more concrete. — Ding: Excavating the Politics of AI.
  7. On historical scale: She uses the pyramids as an analogy for the scale of AI infrastructure investment, urging reflection on what society is building and why. — Ground Truths: Kate Crawford.

Part 2: The Planetary and Environmental Costs

  1. On the business of mining: Crawford and Joler trace how mining and logistics shift environmental and human costs onto communities outside the places where technological value accumulates. — Anatomy of an AI System.
  2. On deep time: She contrasts geological timescales with short-lived devices and split-second convenience. That mismatch makes the material costs of computation easier to overlook. — Big Tech: Kate Crawford Interview.
  3. On the cost of convenience: A brief request to a voice assistant invokes mining, manufacturing, shipping, data processing and labor. The convenience is local; its supporting system is planetary. — Anatomy of an AI System.
  4. On metabolic media: Her metabolic-media metaphor connects the consumption of cultural material, energy and water with the production of synthetic media and environmental waste. — Eating the Future.
  5. On future energy consumption: She cites nation-scale electricity projections to illustrate potential AI demand. These are forecasts, not measurements of an already realized 2030 outcome. — Eating the Future.
  6. On obsolescence cycles: Crawford follows devices beyond use into disposal. Short replacement cycles make e-waste part of the AI lifecycle, not a separate downstream issue. — Ground Truths: Kate Crawford.
  7. On AI as an extractive industry: She begins the material story in mines and mineral supply chains. An account of AI limited to software leaves that dependency out. — Sydney Ideas: Atlas of AI.
  8. On unpriced externalities: She asks whose costs disappear from the price of a technological service, including pollution, hazardous work and the consequences of disposal. — Anatomy of an AI System.

Part 3: The Hidden Labor of Automation

  1. On the myth of automation: What looks like machine intelligence may depend on hidden human work. Crawford asks readers to examine the labor concealed behind the interface. — Reimagining the Internet: Kate Crawford.
  2. On crowd-sourced labor: The labor supply chain includes people labeling datasets and supplying human feedback. Outsourcing those tasks does not make them machine work. — Renewable Matter: Calculating and Powers.
  3. On algorithmic management: Her warehouse observations link automated picking targets and surveillance to workers’ physical strain, anxiety and diminished control over their working day. — Big Tech: Kate Crawford Interview.
  4. On proving our humanity: Crawford and Joler’s essay uses image-based reCAPTCHA as an example of unpaid work: proving oneself human could also help train image recognition. — Anatomy of an AI System.
  5. On unequal value: She contrasts the rewards captured by technology firms with the low pay and difficult conditions of contractors supplying essential human feedback. — Ground Truths: Kate Crawford.
  6. On continuous human input: Generative systems still rely on people evaluating responses and handling difficult content. Apparent autonomy can conceal that corrective human input. — Ground Truths: Kate Crawford.

Part 4: Data Extraction and Epistemological Violence

  1. On the bloodless word: Treating data as a free natural resource hides its connection to people’s lives and the risks attached to reusing it. — Ground Truths: Kate Crawford.
  2. On forcing categories: Putting people into machine-readable categories is a political act. Crawford asks who chose those boundaries and whose understanding they privilege. — Reimagining the Internet: Kate Crawford.
  3. On context and place: An image’s meaning depends on culture, context and interpretation. A dataset label is one reading of it, not its self-evident essence. — Excavating AI.
  4. On rapacious harvesting: Crawford and Paglen’s examination of facial datasets challenges the assumption that publicly available photographs are automatically available for unrestricted machine-learning use. — Excavating AI.
  5. On systematizing the universe: She questions what is lost when a complex world is organized into fixed classes. Making something computable is not a neutral translation. — Reimagining the Internet: Kate Crawford.
  6. On raw data as myth: Crawford and Paglen describe datasets as constructed interventions. Selection, labeling and taxonomy shape what a system can learn before an algorithm processes it. — Excavating AI.

Part 5: Bias and the Politics of Classification

  1. On the term bias: A bias-only framing can miss the deeper question: which worldviews does a system treat as normal even when it is working as designed? — Big Tech: Kate Crawford Interview.
  2. On neutrality: The presumption that automation is more objective can encourage misplaced trust. Crawford warns that human oversight alone does not automatically create accountability. — Ding: Excavating the Politics of AI.
  3. On constructing reality: Race and gender labels in training sets do not simply discover natural divisions. Their design can impose contested social categories on people. — Excavating AI.
  4. On the phrenological impulse: She calls the inference of inner character or state from outward appearance a phrenological impulse, and argues that such claims require scrutiny. — Nature: Time to Regulate Emotion AI.
  5. On affective computing: Crawford challenges claims that facial movements reliably disclose emotion regardless of context. She calls for independent evidence and safeguards before consequential use. — Nature: Time to Regulate Emotion AI.
  6. On historical continuity: Crawford and Paglen’s historical comparison links contemporary facial classification with earlier attempts to measure and rank people. A statistical method does not erase those political inheritances. — Excavating AI.
  7. On technical fixes: Changing demographic balance or deleting offensive labels does not resolve every problem. The purpose and underlying categories of classification also need examination. — Excavating AI.

Part 6: Power and AI as a Registry

  1. On power structures: For Crawford, understanding AI requires examining who controls its infrastructure and institutions, rather than treating the technology as an independent social actor. — Reimagining the Internet: Kate Crawford.
  2. On democratizing AI: Making tools easier to access does not by itself redistribute control over datasets or cloud infrastructure. She questions that limited meaning of democratization. — Anatomy of an AI System.
  3. On time and control: She links algorithmic control of work to control over time. Efficiency targets can constrain workers’ agency, not merely improve a technical process. — Big Tech: Kate Crawford Interview.
  4. On dismantling structures: Invoking Audre Lorde, Crawford questions whether building more of the same infrastructure can undo its power relations. Engineering is not the only political response. — Reimagining the Internet: Kate Crawford.
  5. On asking the right questions: Her guiding questions are who benefits, who is harmed and how power becomes concentrated. Those questions broaden governance beyond individual technical failures. — Big Tech: Kate Crawford Interview.
  6. On the limits of ethics: Ethical principles without enforceable consequences are insufficient. She argues for centering affected communities rather than letting corporate frameworks define the agenda. — Ding: Excavating the Politics of AI.
  7. On billionaires and power: She argues that concentrated AI infrastructure gives a small number of firms unusually broad influence. Its public consequences demand a public debate. — State Library Victoria: Mapping Planetary AI.
  8. On defining knowledge: Training data and classification shape the worldview a system can produce. Crawford warns that flattening complexity can make excluded perspectives harder to recognize. — Renewable Matter: Calculating and Powers.

Part 7: The Myths of AI and Ground Truth

  1. On ground truth: Training labels are human judgments rather than an untouched ground truth. Their assumptions need examination alongside the models trained on them. — Excavating AI.
  2. On scale as authority: A bigger dataset is not necessarily a more representative one. She asks where the material came from and which populations or viewpoints are overrepresented. — Ground Truths: Kate Crawford.
  3. On human labelers: ImageNet’s labels depended on human classification work. Crawford and Paglen examine how that process could turn uncertain judgments into authoritative-looking categories. — Excavating AI.
  4. On epistemic limits: She distinguishes statistical prediction from the fluid, relational nature of human intelligence. Simplifying experience into classes limits what the resulting representation captures. — Renewable Matter: Calculating and Powers.

Part 8: The Future of Model Autophagy

  1. On the slop tide: Crawford warns that synthetic media can undermine the audience and advertising relationships supporting human-created online content, while rewarding cheap attention-seeking production. — Eating the Future.
  2. On Model Autophagy Disease: She highlights the risk of models degrading when too much generated output feeds later training. That risk is not proof that every system must collapse. — Eating the Future.
  3. On forensic literacy: Her method follows AI beyond its interface into mines, warehouses, data work and disposal. Investigating those connections can reveal costs that software-only analysis misses. — Ground Truths: Kate Crawford.
  4. On systemic degeneration: Her essay connects demand for larger datasets with pressure on available human-created material. That pressure motivates new sourcing and synthetic-data strategies, with unresolved trade-offs. — Eating the Future.
  5. On structural regulation: She argues that governance must address labor, resource extraction and institutional power as well as data protection, rather than regulate an abstract technology in isolation. — Big Tech: Kate Crawford Interview.
  6. On Auditable Resource Reporting: Crawford calls for public reporting of AI’s energy and water use, backed by independent environmental audits. Her proposal makes accountability a matter of measurable operating practices, not voluntary sustainability promises alone. — Nature — Generative AI’s Environmental Costs
  7. On Acting Before a Future Breakthrough: In 2024, she argued against relying on hoped-for energy breakthroughs to address AI’s current footprint. More efficient systems and data-centre design were actions available to pursue while longer-term energy research continued. — Nature — Generative AI’s Environmental Costs
  8. On the path forward: Start with the problem a community needs solved, she advises, rather than assuming more AI is the answer. Some uses may warrant refusal altogether. — State Library Victoria: Mapping Planetary AI.