Kate Crawford: The Woman Who Mapped the AI Atlas

AI is neither artificial nor intelligent. It is made from natural resources, fuel, human labor, infrastructures, logistics, histories, and classifications.

Nevada, 2015. Kate Crawford is standing at the edge of an open-pit lithium mine. The mine is enormous — a deep gouge in the earth several kilometres across, carved by industrial machinery to extract the lithium that goes into the batteries that power the laptops and servers and smartphones that run the AI systems she studies.

She is not here as a tourist. She is here as a researcher — as someone who has decided that to understand what AI really is, you need to understand what it is made of. Not just the code. Not just the algorithms. The physical stuff: the minerals in the ground, the factories where the hardware is made, the data centres where the computations run, the labour that extracts and assembles and maintains all of it.

What she is seeing is not what the AI industry’s promotional materials show. The promotional materials show elegant interfaces, helpful assistants, intelligent machines. What she is seeing is a hole in the ground. A wound in the planet. The material reality of the technology that the industry presents as weightless, frictionless, and clean.

The gap between the presentation and the reality is what Kate Crawford has spent her career documenting. The Atlas of AI — the book she published in 2021, one of the most important books ever written about artificial intelligence — is the product of this documentation.

---

Australia to New York: The Formation of a Critical Mind

Kate Crawford was born and grew up in Australia — in the specific Australian intellectual culture that combined British academic traditions with a more egalitarian, less deferential approach to authority and received wisdom. She studied literature and communications, developing an early interest in how media and technology shaped social life, before pursuing an academic career that would bring her to the United States and to the centre of the global AI conversation.

She completed her doctorate at the University of Sydney, working on the social dimensions of new communication technologies. The doctoral research established the intellectual orientation that would define her career: not primarily the technical dimensions of technology, but the social, political, and material dimensions — who builds it, who benefits, who is harmed, and what assumptions are embedded in the design.

She moved to New York in the early 2000s, joining the Microsoft Research lab there as a researcher. The Microsoft connection was significant: it gave Crawford direct access to the AI research community while maintaining enough critical distance to observe the field with the eyes of someone who was not primarily invested in its technological progress. She was inside the tent but not of it — a social scientist studying the institution from within.

At Microsoft Research, she became a senior principal researcher and eventually a cofounder of the AI Now Institute at New York University — the first academic research institute dedicated to studying the social implications of artificial intelligence. The AI Now Institute, which Crawford co-founded with Meredith Whittaker in 2017, became one of the most influential voices in the AI ethics and AI governance conversations, producing annual reports on the social impact of AI that were widely cited in policy discussions.

The AI Now Institute was a deliberate institutional intervention. Crawford and Whittaker recognised that the AI ethics conversation was being dominated by the AI industry and by academic AI researchers who were primarily focused on technical questions. An independent academic institute with genuine research capability and genuine critical distance could contribute perspectives that the industry-dominated conversation was missing.

---

The Atlas of AI: Mapping What the Field Prefers to Ignore

“Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence” was published in 2021 and immediately recognised as one of the most important books about AI to have appeared in years. It is not a technical book — Crawford does not describe neural network architectures or training algorithms. It is a political economy of AI: an account of the material, labour, and political dimensions of AI development that the technical narrative almost entirely omits.

The book’s structure mirrors the title. An atlas maps territory. Crawford is mapping the territory of AI — not the technical territory of algorithms and benchmarks, but the physical, social, and political territory in which AI is embedded.

The earth. The first chapter traces the physical materials that AI systems require — the minerals extracted from open-pit mines, the water consumed by data centre cooling systems, the carbon emitted by the energy consumed in training large models. Crawford spent time in lithium mines in Nevada and cobalt mines in the Democratic Republic of Congo, examining the physical reality of the supply chains that make AI possible.

The environmental critique is not just about carbon emissions — though the energy consumption of large AI models is substantial and growing. It is about the pattern of extraction: the concentration of environmental costs in specific communities and regions, often in the global south, while the benefits of AI accrue primarily to consumers and companies in the global north. The distribution of costs and benefits is not accidental; it reflects the global patterns of extractive capitalism in which AI development is embedded.

Labour. The second major strand of the Atlas is the account of the labour that AI requires — labour that is systematically invisible in the AI industry’s self-presentation.

AI systems are not created by algorithms alone. They are created by the human labour of data collection, data labelling, and content moderation. The workers who label images so that computer vision systems can learn to classify them, the workers who annotate text so that language models can learn to parse it, the workers who review content to train models to distinguish acceptable from unacceptable — this labour is real, substantial, and often poorly paid.

Crawford draws particular attention to the “ghost workers” documented by Mary Gray and Siddharth Suri: the hundreds of thousands of workers, often in developing countries, who perform the micro-tasks that make AI systems work — labelling data, verifying AI outputs, moderating content — through platforms like Amazon Mechanical Turk and through content moderation firms contracted by the major technology companies. The labour is performed at scale, at low wages, often without benefits or employment protections, by workers who are invisible in the AI industry’s account of how its systems are built.

The labour critique extends to the higher-skilled workers at AI companies. Crawford documents the specific working conditions at Amazon’s warehouses — where AI surveillance systems monitor worker productivity to a degree that workers and labour advocates have described as dehumanising — and the specific labour relations at technology companies more broadly, including the advocacy by lower-paid contract workers for wages and benefits comparable to those of full-time employees.

Data. The third major strand of the Atlas is the account of how data — the training material for AI systems — is acquired, and what the acquisition says about the power relations embedded in AI development.

Much of the data that trains AI systems was not created for that purpose and was not provided with the consent of the people it concerns. Facial recognition systems are trained on photographs scraped from the internet — photographs taken of people in public spaces, posted for social purposes, not intended as training data for biometric systems. Language models are trained on text written by millions of people who had no knowledge of or consent for this use. The extraction of data for AI training, like the extraction of minerals for AI hardware, tends to flow from those with less power to those with more.

Crawford is particularly sharp on the history of training datasets. Datasets like ImageNet were assembled by researchers who made specific choices about what categories to include, what images to label as representative, and what labelling criteria to use. These choices were not neutral — they reflected specific assumptions about how the world was organised, what categories were natural, whose perspective was the default. The datasets encode these assumptions, and the AI systems trained on them encode them further.

Classification. One of the Atlas’s most powerful chapters examines the AI industry’s compulsion to classify — to reduce complex, contextual human realities to discrete, computable categories. Emotion recognition systems that assign faces to emotional categories based on facial expressions; gender recognition systems that assign faces to male or female categories based on visual features; affect analysis systems that assign job interview videos to “confident” or “nervous” categories — all of these systems attempt to extract computable categories from human realities that resist categorisation.

Crawford’s critique is not simply that these systems are inaccurate, though the evidence suggests they often are. It is that the entire project of classifying human emotional states or gender identities from facial features or body language is misguided — that the categories being imposed do not correspond to stable, measurable realities in the world, and that the act of imposing them causes specific harms, particularly to people whose lives do not fit neatly into the categories being used.

The state. The final major strand of the Atlas is the account of AI’s relationship to government power — the use of AI systems in surveillance, in law enforcement, in military applications, in the administration of social benefits. Crawford examines the specific use of predictive policing systems, of facial recognition by law enforcement, of algorithmic benefits administration, and of autonomous weapons systems.

The state power critique connects to the broader political economy argument: AI is not a neutral technology that governments and corporations can use for good or ill with equal ease. The specific capabilities that AI systems have — the ability to process large quantities of data about individuals, to identify patterns and make predictions, to operate at scale without human oversight for individual decisions — are capabilities that expand the reach of existing power and concentrate it in ways that are not easily reversed.

---

The Academic Context: AI Ethics vs. Critical AI Studies

Crawford occupies a specific position in the landscape of AI research and AI ethics — a position that is worth locating precisely, because it differs in important ways from the positions most commonly associated with AI ethics in the industry.

The mainstream AI ethics conversation — the conversation that happens in AI company ethics teams, in academic machine learning conferences, in industry-sponsored research centres — tends to focus on specific, technical problems: bias in training data, fairness metrics for algorithmic decisions, privacy protections for training data, transparency in automated decision-making. These are real problems that deserve attention, and significant research is being done on each of them.

Crawford’s approach is different in emphasis and in political implication. Where mainstream AI ethics tends to treat the problems it identifies as engineering problems that can be addressed through better techniques, Crawford’s approach treats them as political problems that require political solutions. Where mainstream AI ethics tends to work within the existing institutional structure of the AI industry — proposing reforms, developing guidelines, providing advice to companies and regulators — Crawford’s approach is more fundamentally critical of that institutional structure.

This difference is not merely academic. It has practical implications for what kinds of solutions are pursued. A research programme that frames algorithmic bias as an engineering problem tends to produce engineering solutions — better fairness metrics, better debiasing techniques, better evaluation frameworks. A research programme that frames algorithmic bias as a political problem tends to produce political solutions — regulatory requirements, collective bargaining rights for data subjects, redistribution of the benefits of AI.

Crawford’s influence has been primarily in the political direction — in the governance and regulation conversations rather than in the technical AI research conversations. The AI Now Institute’s annual reports have been regularly cited in Congressional testimony, in regulatory proceedings, and in legislative discussions. Crawford’s testimony and public advocacy have been influential in shaping how regulators and policymakers understand AI.

---

The Industry Relationship: Inside and Outside

Crawford’s position as a researcher at Microsoft and the AI Now Institute — simultaneously inside the AI industry and critical of it — has been both productive and complicated.

The inside position gave her access. She could observe the culture of AI research up close, could talk to the people building AI systems, could understand the institutional pressures and the assumptions that shaped the work. The access produced the specific, detailed, empirically grounded critique that makes the Atlas of AI more compelling than a critique written from purely outside the industry could have been.

The critical position made her sometimes unwelcome. The AI industry, like most industries, is more comfortable with criticism that proposes incremental improvements than with criticism that challenges the fundamental premises of what the industry is doing. Crawford’s critique was not always welcome at AI conferences, in AI company research labs, or in the broader AI research community.

She has navigated this tension with what appears to be deliberate strategy: maintaining enough institutional connection to have access and credibility, while maintaining enough critical independence to make the fundamental arguments that institutional insiders cannot easily make. The combination has made her one of the most influential voices in the AI governance conversation — more influential, in some respects, than purely technical AI researchers, because she addresses the political and social dimensions that technical researchers are less equipped to analyse.

---

The Materiality Argument: Why It Matters That AI Has Stuff

One of the most distinctive and most important contributions of Crawford’s work is the insistence on AI’s materiality — on the fact that AI systems are not weightless digital constructs but physical objects with physical impacts on the physical world.

The “AI as cloud” framing — the idea that AI is a technology that operates in the digital realm, distinct from the physical world — is pervasive in the AI industry’s self-presentation. The names of cloud computing services reinforce it: “the cloud,” “serverless computing,” “virtual machines.” The language suggests that AI operates in a digital ether that is separate from physical reality.

Crawford’s Atlas is a systematic refutation of this framing. Every aspect of AI development has a physical dimension. The training of a large language model consumes megawatt-hours of electricity, which is generated by burning fuel or capturing renewables, which requires physical infrastructure. The data centres that run AI systems consume enormous quantities of water for cooling. The hardware that performs the computations requires rare earth minerals extracted from specific places in the world. The labour that creates and maintains the systems is performed by human beings in specific physical conditions.

The materiality argument has several specific implications.

Environmental. AI development has environmental costs that are not included in its cost accounting. The carbon emissions from training a large model, the water consumption of data centres, the environmental damage from mineral extraction — these costs are real and are borne by the environments and communities near the physical infrastructure of AI, not by the companies and consumers who benefit from AI capabilities.

Labour. AI development depends on labour that is often poorly compensated, often invisible in the AI industry’s account of itself, and often located in developing countries that lack the regulatory protections available in the jurisdictions where AI companies are headquartered. Making this labour visible is a prerequisite for improving its conditions.

Political. The physical location of AI infrastructure is a political fact with political consequences. The data centres that run AI systems are in specific jurisdictions with specific laws about data privacy, labour rights, and environmental regulation. The supply chains for AI hardware involve specific relationships between corporations, governments, and communities. Understanding these physical and political facts is essential for governing AI responsibly.

---

The Emotion AI Critique: Why Classifying Feelings is Wrong

One of the specific applications of AI that Crawford has critiqued most forcefully is “emotion AI” — systems that attempt to recognise, classify, or predict human emotional states from facial expressions, voice, or body language.

The critique has two distinct strands.

The first is empirical: the science underlying emotion AI is weak. The dominant paradigm in emotion AI is based on Paul Ekman’s theory of basic emotions — the claim that there are six universal, biologically determined emotional states that are expressed in consistent facial expressions across cultures. Ekman’s theory, which was developed in the 1970s, has been extensively criticised by emotion researchers who find that emotional expressions are highly variable across individuals and cultures, that the same expression can mean different things in different contexts, and that the theory of universal basic emotions is not supported by the neuroscientific evidence.

Emotion AI systems trained on datasets of labeled facial expressions will learn to classify faces in ways that match their training data, but the training data encodes the specific cultural assumptions and methodological limitations of the research that produced it. The systems will not measure emotional states; they will classify facial expressions according to a theory of emotion that is scientifically contested.

The second strand of the critique is political: even if emotion AI systems worked as their proponents claim, the applications to which they are being put raise serious concerns about autonomy, privacy, and the use of surveillance to manage workers and job applicants.

Emotion AI is being deployed in hiring — to analyse video job interviews and classify applicants’ emotional states, confidence levels, and personality traits. It is being deployed in workplace monitoring — to track workers’ emotional states and engagement levels. It is being deployed in educational settings — to monitor student attention and emotional engagement with learning material.

All of these applications share a common structure: they subject people to surveillance during situations where they have limited choice about participation, and they use that surveillance to make or inform consequential decisions about the surveilled person, without that person’s meaningful consent and without reliable scientific validation of the measures being used.

Crawford’s critique connects the specific technical problem (the science is weak) with the specific political problem (the power asymmetry is significant) in a way that is more powerful than either critique alone. A perfectly accurate emotion recognition system that was deployed to surveil workers and job applicants without their consent would still be problematic. An inaccurate emotion recognition system deployed with consent in a context where the subject had genuine power to refuse would be less harmful. The actual applications combine both problems.

---

The Research Programme: AI Now and Its Reports

The AI Now Institute — which Crawford co-founded with Meredith Whittaker in 2017 and which she has continued to lead — has produced a series of annual reports on the state of AI and its social impacts that have become reference documents for policymakers, journalists, and advocates working on AI governance.

The AI Now Report 2018 was the first major report to systematically document the use of AI in government services — in child welfare, in criminal justice, in employment services — and to raise specific concerns about the accountability gaps, the lack of public oversight, and the specific harms that automated decision-making in these contexts could cause.

Subsequent reports documented the growth of the AI surveillance industry, the labour conditions of platform workers who supported AI development, the environmental costs of large-scale AI training, and the specific policy responses (and non-responses) to the social harms of AI.

The reports have been influential in specific policy conversations. The concerns documented in AI Now reports about facial recognition and law enforcement contributed to the moratoriums on government facial recognition use that were enacted in several American cities. The labour concerns documented in the reports contributed to advocacy for better protections for platform workers. The accountability concerns documented in the reports informed the development of the EU AI Act’s requirements for transparency and human oversight in high-risk AI applications.

The AI Now Institute also pioneered the practice of tracking and documenting specific incidents of AI harm — creating a repository of documented cases in which AI systems had caused measurable harm to specific people or communities. This practice — of insisting on the concrete and the specific rather than the theoretical and the general — has been influential in shifting the AI ethics conversation from abstract principles to specific harms.

---

The Political Economy of AI: Power and Its Distribution

The deepest argument in Crawford’s work is about power — about who has it, how AI development reflects existing power concentrations, and how it creates new ones.

The AI industry is dominated by a small number of very large companies that have access to the data, the compute, and the talent required to develop frontier AI systems. These companies are American (primarily) and Chinese, they are publicly traded or venture-backed, and their primary accountability is to their shareholders and investors rather than to the people whose data trains their systems or whose lives are affected by their systems’ decisions.

This concentration of power in a few private companies developing technology that affects everyone is, in Crawford’s analysis, a political problem as much as an ethical one. The problem is not just that the companies might make bad decisions — though they sometimes do — but that the decision-making about the development and deployment of AI is happening in a context of extreme power asymmetry, without the democratic accountability that decisions of this consequence should have.

The counterarguments are familiar and worth acknowledging. Market competition constrains the ability of any single company to behave badly. Regulatory oversight provides some accountability. The researchers and engineers at AI companies often have genuine ethical commitments that constrain company behaviour from the inside. The global marketplace gives consumers some power to choose AI products and services that are more consistent with their values.

Crawford’s response to these counterarguments is that they are insufficient to the scale of the power concentration and the consequences of AI development. Market competition does not provide adequate accountability when the companies are as large and as dominant as the major AI companies are. Regulatory oversight is limited by regulators’ technical capacity and by the lobbying influence of the regulated. Individual researcher ethics are constrained by the institutional pressures of commercial organisations. Consumer choice is limited by the oligopolistic structure of many AI markets.

The political economy argument points toward specific policy responses: antitrust enforcement to limit concentration, data governance frameworks that redistribute the value of data toward its generators, regulatory frameworks with enforcement teeth, international governance mechanisms that can manage the race dynamics. These are not primarily technical solutions but political ones.

The deepest argument in Crawford’s work is about power: about who has it, how AI development reflects existing power concentrations, and how it creates new ones. The political economy framing produces political solutions — antitrust, data governance, regulatory enforcement — that the engineering framing of the same problems cannot generate.

---

The Criticism of Crawford: What She Gets Wrong

Any honest account of Crawford’s work must acknowledge the specific criticisms it has attracted — criticisms that come both from within the AI industry and from researchers who take the AI ethics and governance questions seriously.

The technical critique. Some AI researchers have argued that Crawford’s analysis lacks technical depth — that her accounts of how AI systems work, and of what specific technical choices mean for social outcomes, are sometimes imprecise or misleading. The critique is that social scientists critiquing technology they do not fully understand can misidentify the technical sources of social problems, leading to policy responses that do not address the actual causes.

This critique has some merit. The Atlas of AI is not primarily a technical document, and there are passages where the technical claims are less precise than a computer scientist would prefer. But the core arguments — about materiality, labour, data extraction, and power — do not depend on technical precision about AI architectures, and the imprecision in specific technical passages does not undermine the political economy analysis.

The negativity critique. Some AI researchers have argued that Crawford’s work is systematically negative about AI — that it presents AI’s harms without adequate acknowledgment of its benefits, and that this asymmetry distorts the overall assessment.

This critique reflects a genuine tension. Crawford’s work is critical; it is designed to surface dimensions of AI development that the industry narrative obscures. A work designed to surface overlooked negatives will necessarily emphasise those negatives, and a reader who only reads Crawford’s work will not have a complete picture of AI’s impact.

But the critique misses the point of critical research. The AI industry and mainstream AI research produce copious accounts of AI’s benefits. The specific contribution of critical AI studies is to produce accounts of the costs and the politics that those mainstream accounts omit. The appropriate response to the negativity critique is not that Crawford should be more balanced, but that the full picture requires reading both the industry’s account and the critical account.

The agency critique. Some critics have argued that Crawford’s analysis underestimates the agency of people within the AI industry — the researchers and engineers and executives who have genuine ethical commitments and who are working to make AI development more responsible. On this view, the emphasis on structural forces and power relations obscures the specific choices that individuals make and the specific ways that individuals can push for better outcomes within institutions.

This critique also has some merit. The Atlas of AI is primarily a structural analysis, and structural analyses can underweight individual agency. But structural analysis and individual agency are not mutually exclusive — understanding the structural context within which individual choices are made is essential for evaluating how much agency individuals actually have and what kinds of choices are available to them.

---

The Legacy: What Crawford Has Built

Kate Crawford’s legacy in the AI field is primarily one of intellectual infrastructure — the concepts, the research programme, and the institutional capacity that have made critical AI studies a serious field with real influence on AI governance.

The conceptualization of AI as political. Crawford’s work, more than any other single contribution, has established the framing that AI is a political technology — that it reflects and shapes power relations, that its development is embedded in specific political economies, that governance requires political as well as technical analysis. This framing is now common in policy discussions; it was not the default framing ten years ago.

The materiality framework. The insistence on AI’s physical, material dimensions — the minerals, the data centres, the labour, the environmental costs — has become a standard element of serious AI analysis. The organisations that produce AI now face regular questions about their carbon footprint, their data centre water consumption, their supply chain labour practices. These questions became standard partly because Crawford and her colleagues made them impossible to ignore.

The AI Now Institute. The institutional contribution of the AI Now Institute — as a source of independent, rigorous, empirically grounded research on AI’s social impacts — has been significant. The annual reports, the specific advocacy work, the documentation of AI harms, and the policy engagement have produced measurable policy outcomes.

The training of researchers. The researchers who have worked at the AI Now Institute and in Crawford’s networks have carried her intellectual framework into academic positions, policy organisations, journalism, and advocacy. The field of critical AI studies that she helped establish is now a recognisable research discipline with its own journals, conferences, and research programmes.

---

The Ongoing Work: What Crawford Is Doing Now

Crawford’s research and advocacy continue, focused increasingly on the geopolitical dimensions of AI development — on the specific ways that AI development reflects and shapes the competition between the United States, China, and other major powers, and on the governance challenges that this geopolitical dimension creates.

The geopolitical analysis extends her earlier work on power and political economy to the international level. The AI race between the United States and China is not just a technological competition; it is a competition over who will control the infrastructure of the digital economy, who will define the standards for AI development and deployment, and whose values will be embedded in the AI systems that will increasingly shape global life.

Crawford’s research on this question connects to her earlier work on military AI — the use of AI in autonomous weapons, in intelligence analysis, in cyber operations — and to the broader question of how AI governance can function in a world where the major AI powers are in strategic competition. The governance question, in its international dimensions, is the hardest version of the problem that Crawford has been studying throughout her career. It requires the kind of international cooperation that is hardest to achieve when the parties are in strategic competition. It requires governance frameworks that can manage the specific risks of AI without being captured by the commercial interests that are most invested in avoiding governance. And it requires the kind of democratic legitimacy that international institutions have historically been poorly equipped to provide.

Crawford’s contribution to thinking about these questions — through her research, through the AI Now Institute, through her public advocacy — is ongoing and important. The field of AI governance is still being built, and the specific concepts, frameworks, and institutions that will allow AI development to be governed responsibly are still being developed. Crawford’s work is part of that building.

---

The Irreplaceable Perspective: What Only an Outsider-Insider Can See

What makes Crawford’s contribution to the AI field distinctive — what cannot be replicated by researchers who are purely inside the technical AI community or purely outside it — is the specific perspective that comes from having genuine access to the inside while maintaining genuine critical distance from it.

The purely technical AI researcher sees the technology clearly but may not see the politics — may take for granted the institutional structures and the political economies in which the technology is embedded, because those are the water in which the researcher swims. The purely external critic sees the politics clearly but may not see the technology — may attribute to the technology properties it does not have, or miss the specific technical features that drive specific political outcomes.

Crawford sees both. She has spent enough time inside the AI research world to understand the specific technical choices and their specific implications. She has maintained enough critical distance to see those choices as choices — as decisions that could have been made differently, that reflect specific interests and assumptions, that produce specific distributions of benefit and harm.

This combination is rare and valuable. It is what allows the Atlas of AI to be specific rather than generic, grounded rather than abstract, empirical rather than speculative. It is what makes Crawford’s critique hard to dismiss as uninformed, and hard to co-opt as an internal reform programme.

The AI field needs this kind of insider-outsider critique more than it needs more of what it already has — more technical papers, more benchmark results, more incremental improvements. The most important questions about AI are not technical questions; they are political, social, and moral questions. Answering them requires the kind of work that Crawford does: rigorous, empirical, politically serious, and genuinely independent of the industry it examines.

The irreplaceable perspective is the insider-outsider one. The purely technical researcher sees the technology clearly but not the politics. The purely external critic sees the politics clearly but not the technology. Crawford sees both — and the combination is what makes her critique specific rather than generic, grounded rather than abstract, empirical rather than speculative.

---

Further Reading

---

---