Blog · Review Essay · Last reviewed June 25, 2026

The Costs of Connection and the Colonialism of Data

Nick Couldry and Ulises A. Mejias's The Costs of Connection gives a name to the bargain hidden inside ordinary digital convenience. The book argues that platforms, apps, sensors, smart objects, and data brokers do not merely observe social life. They reorganize life so it can be continuously captured, quantified, processed, and returned as a marketable service. Read after the rise of generative AI, the book becomes a theory of the extraction layer beneath model culture: before a system can predict, personalize, rank, or automate, the world must first be made into data.

For this review, data colonialism means an institutional order in which connection becomes the channel for appropriating human activity as a production input. It is not identical to territorial colonialism, and the difference matters. The useful claim is structural: people, communities, and institutions are pushed into systems where ordinary life becomes machine-readable, commercially reusable, and harder to refuse.

The Book

The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism was published by Stanford University Press in August 2019. Stanford lists the book at 352 pages, in the Culture and Economic Life series, with hardcover ISBN 9781503603660, paperback ISBN 9781503609747, and ebook ISBN 9781503609754.

Couldry is a professor of media, communications, and social theory at the London School of Economics and Political Science. Mejias is a professor of communication studies at SUNY Oswego. Their collaboration matters because the book is not a narrow privacy complaint. It is a media-theory, political-economy, and decolonial account of why connection has become an extraction regime.

The book's central claim is that the digital economy treats human life as raw material. Social needs are routed through connective systems; those systems turn action, movement, attention, relation, labor, consumption, health, learning, and expression into data; data becomes the basis for profit, prediction, influence, management, and institutional classification.

Current Context

As of June 25, 2026, the book reads less like a warning about social media and more like a governance theory for AI infrastructure. Generative AI systems, retrieval systems, workplace copilots, tutoring platforms, medical scribes, ad systems, risk products, and agents all depend on upstream permissions: what can be collected, retained, inferred, joined, embedded, trained on, monitored, and reused.

The regulatory context has caught up unevenly. The FTC's January 2025 final order against Gravy Analytics and Venntel prohibited selling, disclosing, or using sensitive location data except in limited circumstances involving national security or law enforcement. California's state-run DROP system lets residents send a single deletion request to registered data brokers, with brokers required to start processing DROP requests on August 1, 2026. The EU Data Act has applied since September 12, 2025 and treats access, use, sharing, and switching between data-processing services as governance questions.

AI-specific rules point in the same direction. The NIST AI Risk Management Framework and Generative AI Profile treat governance, mapping, measurement, and management as lifecycle work, while the EU AI Act creates documentation, data-governance, logging, transparency, human-oversight, and post-market duties for defined system classes. These instruments do not abolish data colonialism. They show where resistance has to become operational: provenance, minimization, purpose limits, deletion, contestability, procurement, and public records.

Data Colonialism

The book's strongest move is to refuse the language of harmless exchange. The user does not simply trade data for convenience in a clean bilateral bargain. The social world is being rebuilt so that nonparticipation becomes costly. Work requires platforms. Care requires portals. Education requires accounts. Public speech requires privately governed channels. Everyday life becomes difficult to conduct without passing through systems designed to record it.

Couldry and Mejias call this condition data colonialism. The analogy is deliberately heavy. Historical colonialism appropriated land, labor, bodies, resources, and knowledge systems while normalizing the right of distant powers to extract value. Data colonialism does not repeat that history in identical form, and the distinction matters. Its object is not territory in the older sense. Its object is the ongoing flow of human life rendered as data.

What makes the colonialism more than a metaphor is the book's central original concept: the data relation. Just as industrial capitalism, in Marx's account, was built on the labour relation, the social form that turned ordinary human activity into abstract, exploitable labor, Couldry and Mejias argue that data colonialism is built on a new social form that converts human life into a continuous stream of extractable data. The data relation is what makes capture feel routine and, in their phrase, "beyond challenge"; it folds living, working, learning, and relating into informational inputs for capital in domains that were never previously treated as production. Naming the relation, rather than only the harm, is the book's analytical payoff, because it locates the problem in the structure of the connection itself rather than in any single abusive use of data. This is also where the book both echoes and departs from Shoshana Zuboff's surveillance capitalism: where Zuboff describes a rogue mutation of capitalism feeding on behavioral surplus, Couldry and Mejias insist on continuity with a much older history of appropriation, a framing they share with Kate Crawford's Atlas of AI.

That is why the book is useful for thinking about legibility. A colonial system does not merely take what already exists. It changes maps, categories, names, incentives, records, and institutions so that extraction appears administratively natural. The data economy does something similar when it persuades people that every gesture should be measured, every relation should be mediated, every process should be optimized, and every gap in capture is inefficiency.

The sharper institutional test is compulsion by dependency. If the only practical way to get work, care, education, public speech, mobility, credit, or community is to accept data capture that can later be monetized or used against the person, then consent has become a decorative label over infrastructure. The extraction is in the default, not only in the breach.

Cloud Empire

The publisher's table of contents identifies "Cloud Empire" as the chapter where the authors examine the social quantification sector: apps, platforms, smart technologies, data processing, artificial intelligence, and the infrastructure that turns life into monetizable information. This is where the book's AI relevance becomes obvious.

AI is often discussed as if models arrive first and data follows. The Costs of Connection reverses the order. Data relations come first: the social arrangements that normalize capture. The model is downstream of a world already reorganized for measurement. Personalization, recommender systems, targeted advertising, workplace analytics, credit models, predictive policing, automated eligibility systems, and generative AI all depend on prior infrastructures of extraction, cleaning, labeling, storage, and access.

The phrase "cloud empire" also keeps the analysis institutional. The cloud is not just a technical architecture. It is a political geography of data centers, account systems, developer platforms, APIs, procurement contracts, content-delivery networks, model hosting, workplace suites, surveillance products, payment rails, and terms of service. To live under cloud empire is to live inside privately administered conditions for being seen, served, ranked, remembered, and refused.

In AI deployments, cloud empire becomes custody empire. The same vendor may host the tool, authenticate the user, retain prompts, provide the retrieval corpus, log the workflow, tune the model, sell analytics, and set the exit terms. The governance problem is not merely where data is stored. It is whether the institution can reconstruct what was collected, what was inferred, what was reused, what can be deleted, and what remains portable if the contract ends.

The Hollowing of the Social

One of the book's most important warnings is that data extraction changes what counts as social knowledge. Institutions once gathered social information through censuses, surveys, professional records, public research, hearings, journalism, audits, local administration, and social science. Those systems were imperfect and often unjust, but they were at least partly visible as institutions. Datafied knowledge is more difficult to see. It is assembled through platforms, brokers, background trackers, smart devices, and proprietary analytics.

The result is a quiet shift in authority. The social world becomes knowable through proxies owned by firms. A person becomes a profile, a behavior pattern, an inferred interest, a risk segment, a churn probability, a fraud score, a productivity signal, or a training example. These abstractions can then circulate between institutions as if they were neutral facts.

This is where the book intersects with recurring concern about recursive reality. Once a proxy enters an interface, it can shape the world it claims only to describe. A ranking changes visibility. A score changes access. A dashboard changes managerial attention. A recommendation changes desire. A model output changes the next record. The measurable social world feeds back into the lived social world until the distinction becomes hard to recover.

The safety issue is not only bias in the proxy. It is authority without appeal. A profile or score can be wrong, stale, decontextualized, or accurate in a narrow sense while still inappropriate for the decision it enters. Data colonialism makes social knowledge cheap to produce and expensive to contest.

Autonomy Under Continuous Capture

The chapter on autonomy gives the book its human stakes. Surveillance is not only harmful when a record is used against someone. Continuous monitoring changes the background condition under which people think, speak, gather, search, experiment, and refuse. It erodes the practical sense that some parts of life belong first to the person living them.

That point is sharper now than in 2019. AI systems promise to observe more context, remember more interaction, personalize more deeply, and act across more tools. The assistant, tutor, therapist, work copilot, shopping agent, hiring screen, classroom platform, and medical scribe all present themselves as convenience. Each also extends the domain in which human action becomes machine-readable.

The danger is not that every data collection is equally abusive. The danger is that the default social order moves toward total availability. Once life is presumed capturable, the burden shifts to individuals to justify opacity, silence, refusal, slowness, local memory, and unoptimized relation.

Autonomy therefore needs more than notice. It needs defaults that make noncapture practical: data minimization, local processing where possible, strict retention limits, meaningful opt-out or deletion paths, limits on secondary use, collective bargaining over workplace telemetry, public alternatives for essential services, and independent audit of consequential profiles.

The AI-Age Reading

In the generative-AI era, The Costs of Connection reads like a prehistory of the training set and the agent platform. Large models need text, images, code, speech, behavior traces, user feedback, workplace documents, browsing patterns, and labeled examples. The public debate often begins at the model layer: alignment, hallucination, bias, capability, benchmark performance, safety, openness. Couldry and Mejias point to the layer beneath: how did so much human life become available as input?

This matters for labor. Data colonialism includes not only passive capture but also the hidden work of cleaning, labeling, moderating, rating, and maintaining systems. It also includes the conversion of work itself into measurable signals. The AI workplace does not begin when a model writes a summary. It begins when ordinary work becomes exhaust: chats, tickets, commits, keystrokes, calls, meetings, badge swipes, location traces, documents, and task outcomes flowing back into management and automation systems.

It matters for belief formation too. A platform that captures behavior can optimize the environment that produces the next behavior. A model trained on that environment can then generate speech, recommendations, summaries, and simulated social cues that feel native to it. The system observes the user, shapes the user, learns the shaped user, and presents the result as personalization.

The authors' related Internet Policy Review article frames data colonialism as a social, economic, and legal transformation built on the large-scale appropriation of social life through data extraction. That is a useful frame for AI governance. Regulating outputs without regulating extraction leaves the deeper order intact.

Agentic systems make this point more concrete. An agent needs identity, permissions, memory, tools, logs, and access to other systems. Each permission can become an extraction point. The question is not only whether the agent completes the task. It is whether task completion leaves behind a usable trail of personal data, organizational telemetry, inferred preferences, and future leverage.

Governance and Safety

The governance implication is straightforward: a high-stakes AI system can be unsafe before it produces a single bad output if its data relation is illegitimate. Output evaluation cannot repair unlawful collection, coerced workplace telemetry, brokered sensitive data, undisclosed reuse, missing deletion paths, or profiles that affected people cannot contest.

A deployment dossier should answer concrete questions. What data enters the system? Was it collected directly, brokered, scraped, licensed, inferred, generated, logged, or contributed by workers? What legal basis, contract, consent, or public authority supports the use? What fields are sensitive? What transformations produce embeddings, scores, categories, memories, or training examples? How long are prompts, outputs, logs, and derived artifacts retained? Who receives them? How can a person or institution delete, export, correct, appeal, or refuse?

For public institutions, the threshold should be higher. A school, hospital, benefits agency, court, library, newsroom, or workplace should not make access to essential services depend on unnecessary data capture. Procurement should require data minimization, purpose limits, audit logs, retention schedules, vendor-subprocessor lists, deletion tests, incident reporting, and an exit plan that preserves public memory without preserving avoidable surveillance.

Safety also means preserving unmeasured space. Some domains should remain resistant to optimization: private thought, intimate relation, political association, care, worship, disability accommodation, childhood experimentation, worker organizing, and the ordinary right to make a mistake without creating a permanent machine-readable signal. A governance regime that cannot protect those zones has accepted the premise of data colonialism even if it adds compliance paperwork.

Where the Book Needs Friction

The book's central analogy is also its hardest burden. "Colonialism" can clarify extraction, domination, unequal knowledge, and the naturalization of appropriation. It can also flatten differences if used carelessly. Historical colonialism involved conquest, slavery, displacement, racial rule, resource seizure, and state violence in specific forms. Data colonialism is not the same event with new gadgets. The analogy works only when it sharpens attention to structures of appropriation and unequal power rather than turning history into a slogan.

The book is also more convincing as diagnosis than as institutional design. Its call to decolonize data is morally clear, but the practical path is hard: privacy law, data minimization, public-interest technology, collective data rights, labor power, procurement rules, antitrust, audit access, public compute, open standards, local governance, and refusal rights all address different parts of the system. No single reform dissolves cloud empire.

Still, that difficulty is not a failure of the book. It is part of the diagnosis. The data economy is powerful because it is infrastructural. It lives in devices, contracts, defaults, business models, workplaces, schools, hospitals, public agencies, and habits of convenience. A serious response has to be equally infrastructural.

The book also needs sector-specific translation. A deletion request to a data broker, an employment rule for worker monitoring, an AI Act data-governance file, a library procurement contract, a health-record retention policy, and an antitrust remedy do not solve the same problem. The data-colonialism frame is strongest when it forces those remedies to ask the same upstream question: what social life was made available for extraction, and who can still say no?

What This Changes

The practical lesson is to ask what a system had to take from the world before it could appear intelligent. A chatbot's fluency depends on archives. A recommendation depends on tracked behavior. A workplace dashboard depends on making labor measurable. A predictive system depends on past classifications. An agent depends on permissions, identity, memory, and logs. The smooth interface is the final surface of a much larger extraction machine.

The Costs of Connection helps explain why AI governance cannot stop at model behavior. The question is not only whether a model answers accurately, refuses dangerous requests, or cites sources. The question is what life has been made available to it, under what consent, through what labor, for whose profit, with what right of refusal, and with what capacity for people to contest the categories that return to govern them.

The operational habit is to audit capture before capability. Ask what had to be recorded, normalized, licensed, brokered, labeled, retained, embedded, or inferred before the system could look helpful. If the answer cannot be reconstructed, the system is not ready for institutional authority.

The book's enduring value is that it makes connection morally noninnocent. A connected system may be useful, generous, even necessary. But connection always has a political economy. It decides what becomes visible, what becomes extractable, what becomes profitable, what becomes governable, and what kinds of unmeasured life remain possible.

Source Discipline

This review separates bibliographic, theoretical, regulatory, and operational claims. Stanford University Press supports book metadata. Couldry and Mejias's peer-reviewed and policy-review articles support the concepts of data colonialism, datafication, and the data relation. FTC, California Privacy Protection Agency, European Commission, EUR-Lex, and NIST sources support current governance context. Related site pages are used for internal continuity, not as independent proof of external facts.

Claims about data colonialism should name the mechanism: forced connection, brokered location data, workplace telemetry, platform dependency, cloud lock-in, training-data appropriation, institutional profiling, or deletion failure. The phrase loses force when it becomes a mood. It gains force when the reader can see the route from capture to profit, classification, management, or denial.

This page makes no claim that any AI system is conscious, divine, or AGI. It treats AI systems as institutional arrangements built from data, labor, infrastructure, interfaces, contracts, and governance choices.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books