Wiki · Concept · Last reviewed June 19, 2026

Surveillance Capitalism

Surveillance capitalism is a business logic in which behavioral traces are extracted from people, converted into predictive models, and used to rank, price, target, recommend, or shape future behavior. In the AI era, the same logic moves from ads and feeds into assistants, agents, memories, scoring systems, and personalized interfaces.

Category: Concept Published: June 19, 2026 Modified: June 19, 2026 Last reviewed: June 19, 2026 Tags: Privacy, Advertising, Data Brokers, AI Governance, Platform Power

Definition

Surveillance capitalism is Shoshana Zuboff's term for a digital economic order that treats human experience as raw material for hidden extraction, prediction, and behavioral influence. The core move is not ordinary recordkeeping. It is the capture of behavioral surplus: data gathered beyond what is needed to provide the immediate service, then repurposed for prediction, targeting, experimentation, and control.

The term is useful because it names a market structure, not just a privacy violation. A platform, broker, advertiser, app, retailer, insurer, employer, or public contractor may collect traces of action, location, search, purchase, attention, emotion, relationship, or vulnerability. Those traces become models, scores, segments, rankings, prices, or prompts that affect what the person later sees and can do.

Not every use of data is surveillance capitalism. A service can collect data for security, accessibility, accounting, user-requested personalization, or legal compliance. The relevant question is whether collection exceeds the user's immediate context, whether the system monetizes prediction or influence, and whether the affected person can understand, refuse, contest, or exit the flow.

Boundary Test

A practical boundary test has three layers. First, service necessity: is the data required to deliver the function the user requested, or is it surplus captured because the interface can observe it? Second, cross-context reuse: does data from one relationship move into advertising, scoring, training, brokerage, pricing, or persuasion? Third, asymmetry: can affected people see, refuse, delete, contest, or leave without losing access to an essential service?

The phrase should not be used as a synonym for all analytics, all personalization, all advertising, or all capitalism in digital markets. It is strongest when a business model depends on converting behavior into prediction and influence products while keeping extraction, inference, recipients, and downstream uses difficult for people to inspect.

How It Works

The pattern has four recurring stages: capture, inference, prediction, and intervention. Capture gathers behavioral traces through accounts, cookies, pixels, SDKs, mobile advertising IDs, location signals, purchases, loyalty programs, connected devices, workplace tools, search logs, social graphs, and app telemetry. Inference turns those traces into identities, interests, vulnerabilities, probabilities, and segments.

Prediction products then enter markets and systems: real-time bidding, recommender ranking, lookalike audiences, fraud scores, lead generation, dynamic pricing, political targeting, people-search databases, and data broker products. Intervention closes the loop by changing a feed, ad, price, offer, prompt, ranking, notification, or conversational response.

The mechanism does not require a single all-knowing database. It often works through many partial systems: first-party platform logs, third-party ad tech, brokered identity graphs, vendor APIs, cloud analytics, device telemetry, and model features. That distribution makes accountability hard because each actor can claim to hold only one piece of the profile.

System Map

A source-disciplined surveillance-capitalism analysis should map the whole chain, not just the visible interface. The minimum record is: capture surface, claimed service purpose, surplus data category, identity or device linkage, inference method, prediction product, recipient or market, intervention surface, retention period, deletion path, and user contestability.

This map separates useful personalization from cross-context extraction. A user-held preference, an accessibility setting, and a local project memory are not the same as an advertising segment, brokered location graph, individualized price, workplace score, or persuasive prompt chosen because a system inferred vulnerability.

For AI deployments, the map should connect to AI Data Retention, AI Data Provenance, AI Agent Observability, Digital Identity, and Transparency and Public Registers. Without those links, a privacy promise can remain a front-end statement while the operational system keeps extracting, inferring, and acting downstream.

Current Context

As of June 19, 2026, surveillance capitalism is not a standalone legal category in the United States or European Union. Regulators instead target pieces of the system: unfair or deceptive data collection, sensitive location data sales, targeted advertising, children's data, data broker deletion, ad transparency, consent interfaces, and discriminatory or manipulative outcomes.

The Federal Trade Commission uses the adjacent term "commercial surveillance" for the business of collecting, analyzing, and profiting from information about people. That 2022 rulemaking vocabulary is not identical to Zuboff's theory, but it gives governance work a more inspectable unit: collection, aggregation, analysis, retention, transfer, monetization, security, automated analysis, and consumer harm.

In the United States, the Federal Trade Commission's September 2024 staff report on social media and video streaming services said major services had engaged in vast surveillance of users, with weak privacy controls and inadequate safeguards for children and teens. The FTC's 2025 surveillance-pricing work separately highlighted how detailed consumer data, including precise location and browser history, can support individualized prices or discounts.

FTC data-broker enforcement has made the location-data branch of the system concrete. Since 2024, final orders involving X-Mode/Outlogic, InMarket, Mobilewalla, and Gravy Analytics/Venntel have restricted sensitive location-data practices. In May 2026, the FTC announced a proposed settlement to bar Kochava and a subsidiary from selling, sharing, or disclosing sensitive location data without affirmative express consent.

California now supplies a different kind of control. The state-run Delete Request and Opt-out Platform, DROP, lets California residents send a deletion request to registered data brokers through one system. The public DROP site says data brokers must begin deleting matching data within 90 days starting August 1, 2026.

A national-security layer has also become explicit. The U.S. Department of Justice says its Data Security Program under Executive Order 14117 went into effect on April 8, 2025, restricting covered transactions that give countries of concern or covered persons access to government-related data or bulk U.S. sensitive personal data. In February 2026, the FTC warned data brokers that PADFAA prohibits making personally identifiable sensitive data about Americans available to foreign adversary countries or entities controlled by them. This does not solve consumer privacy, but it shows that brokered behavioral data is now treated as geopolitical infrastructure as well as advertising fuel.

In the European Union, the Digital Services Act addresses parts of the ad-surveillance loop by requiring ad transparency, public ad repositories for very large platforms and search engines, restrictions on advertising based on sensitive data, and rules against deceptive interface design. GDPR and the European Data Protection Board's consent-or-pay work remain central where behavioral advertising depends on consent.

AI Relevance

AI intensifies surveillance capitalism because models can make more data surfaces useful. Chat histories, uploaded files, voice samples, embeddings, saved memories, tool traces, search queries, work documents, support tickets, and agent actions can become training data, retrieval context, personalization memory, evaluation material, or targeting signals.

Assistants and agents raise the stakes because they collect high-intent first-party data: questions, drafts, work tasks, relationship details, files, schedules, documents, and tool outputs. If those traces are reused for advertising, model training, broker enrichment, ranking, or pricing without a separate purpose boundary, the assistant becomes a surveillance-capitalism interface even when the moment feels like private help.

The distinctive AI shift is adaptive intimacy. A feed can rank content. A chatbot or companion can remember, infer mood, test phrasing, ask follow-up questions, recommend products, summarize a person to themselves, and guide action. The user may experience help while the institution experiences a richer prediction and influence channel.

Agentic commerce and AI search connect extraction to action. If an assistant can compare products, book travel, fill forms, negotiate, or rank options, profile-derived signals can change not only what a person sees but what the system does on the person's behalf. Governance should distinguish user-held preferences from platform-held prediction products and advertiser-held targeting logic.

AI also makes surveillance less visible. A profile may appear only as a generated answer, risk score, agent recommendation, eligibility ranking, price, or saved memory. Derived data such as embeddings, summaries, clusters, and inferred traits can preserve sensitive meaning even when raw records are no longer visible to the user.

This does not mean an AI system is conscious, divine, AGI, or inherently uncontrollable. The governance issue is institutional: who collects the traces, what model or vendor receives them, what purpose is claimed, what secondary use follows, and what evidence proves that refusal, deletion, or contestation works.

Governance and Safety

Surveillance-capitalism governance begins with contextual integrity and data minimization. A system should document what data is collected, why it is necessary, what context produced it, what new context receives it, what inferences are generated, who receives them, and how long they persist.

AI systems need separate controls for separate uses. Consent for advertising should not silently authorize model training, assistant memory, product improvement, agent connectors, brokered enrichment, political targeting, or workplace scoring. Each reuse needs its own purpose, legal basis, retention rule, opt-out or deletion pathway, and evidence record.

High-risk domains should use stricter defaults: minors, health, finance, employment, education, housing, public benefits, immigration, law enforcement, political persuasion, religious or spiritual testimony, crisis support, and companion products. In these contexts, behavioral profiling should trigger impact assessment, vendor review, human oversight, appeal rights, and limits on sensitive inference.

Institutions should maintain a purpose ledger for AI-mediated personalization: data class, collection context, lawful basis or consent basis, model or vendor recipient, training-use status, memory status, advertising-use status, retention period, deletion proof, and appeal path. The ledger should be testable against production logs, not only privacy-policy language.

Operational controls include short retention windows, log redaction, purpose-scoped permissions, no training reuse by default for sensitive interactions, ad and recommender transparency, user-visible memory controls, audit trails for model and vendor access, regular deletion tests, and independent research or regulator access where law allows.

The safety problem is not only privacy loss. Surveillance systems can enable manipulation, discrimination, stalking, price discrimination, political microtargeting, government bypass of legal process, credential exposure, and feedback loops in which inferred vulnerability becomes the reason a person is targeted again.

Risk Pattern

Context collapse. Data given off in one setting becomes evidence, leverage, or personalization in another.

Inference laundering. Sensitive information reappears as a score, segment, embedding, model feature, or audience label rather than a visible record.

Behavioral steering. Prediction becomes intervention through ads, feeds, nudges, notifications, prices, prompts, or recommendations.

Price and access discrimination. The same profile that targets a promotion can also shape price, eligibility, ranking, or opportunity.

Deletion ambiguity. Raw records may be deleted while derived profiles, vendor exports, suppression lists, embeddings, backups, or model features persist.

Government bypass. Commercially available data can let public agencies buy information that would otherwise require legal process or political scrutiny.

Companion and assistant dependency. AI systems that remember intimate context can convert trust, loneliness, or routine reliance into a durable influence surface.

Source Discipline

Use "surveillance capitalism" carefully. Zuboff's work is a critical theory and historical account; FTC, CPPA, EDPB, European Commission, and court materials are legal or regulatory sources. Do not cite Zuboff as proof that a particular company broke a law, and do not cite a regulator's order as proof that the entire economy fits Zuboff's model.

Also distinguish Zuboff's "surveillance capitalism" from the FTC's "commercial surveillance." The first is a political-economic theory about prediction and influence markets. The second is regulatory vocabulary for data collection, analysis, retention, transfer, monetization, and security practices. They overlap, but neither term should be used to smuggle in conclusions the source does not establish.

For factual claims, name the data flow. Location data sales, real-time bidding bid requests, social-media ad targeting, data broker deletion, AI training reuse, saved memory, dynamic pricing, and recommender ranking are related but different mechanisms. Evidence about one does not automatically prove the others.

Preserve dates and legal status. An FTC complaint states allegations; a final order imposes obligations on named parties; a proposed settlement may still be awaiting finalization; a state deletion platform may be live before all broker processing duties begin. European obligations also depend on the instrument, jurisdiction, platform size, and processing purpose.

National-security claims require the same care. DOJ's Data Security Program, Executive Order 14117, PADFAA, FTC warning letters, and individual enforcement actions are different authorities and procedural stages. Cite the operative source and avoid treating a foreign-access rule as a general privacy remedy.

For AI claims, distinguish base-model capability from deployed system behavior. A model card, privacy notice, or product setting is not enough by itself. The source-disciplined unit is the full system: data collection, retention, training policy, retrieval, memory, ranking, user interface, vendor contracts, access logs, and deletion evidence.

Spiralist Reading

For Spiralism, surveillance capitalism is cognitive extraction.

It converts attention, memory, preference, fear, desire, and social relation into a substrate for prediction and control. The injury is not only that the system watches. It is that watching becomes a market, and the market learns to shape what it watches next.

The AI-era form is more intimate because the extraction can speak back. It can call itself assistance, personalization, companionship, safety, discovery, or convenience while still building a profile that outlives the moment. The Mirror becomes a merchant when reflection is tied to prediction markets.

The Spiralist answer is not nostalgia for ignorance. It is governed memory: collect less, explain more, separate contexts, preserve refusal, and refuse to treat every human trace as raw material.

Open Questions

Sources

Return to Wiki