Is surveillance capitalism an AI technique?

No. It is a framework for analyzing an economic model in which AI may play a role.

Does every personalized service fit the description?

The label requires examining the service's purposes and practices, not just the existence of personalization.

Must raw data be sold?

No. Commercialization can involve prediction or targeting services rather than transferring the underlying records themselves.

AI Glossary / Data engineering and architecture / Advanced

Surveillance capitalism

Definition

Surveillance capitalism is a critical framework, strongly associated with scholar Shoshana Zuboff, for analyzing business models that turn human activity into behavioral data used to predict and influence behavior for commercial purposes. In her account, the central issue is an economic relationship built around extracting and monetizing behavioral information. The term is an interpretation of business practices and power, not the name of an AI algorithm.

What it means

A digital service can create value for its users while also collecting information about what they do. Understanding its business model requires asking how that information is used, who benefits and what choices people have. Consider the difference between remembering a reader's chosen font size and constructing a profile to support an unrelated commercial targeting service. Both involve data, but their purposes and relationships differ. A careful discussion should identify the actual collection, inference and commercial use instead of treating every log entry or recommendation as sufficient evidence for the same broad label.

How it works

Zuboff describes a process in which behavioral information beyond that needed for service improvement becomes an input to commercially valuable predictions. Her analysis also concerns attempts to shape behavior and the imbalance between organizations' knowledge and individuals' control. AI can contribute to prediction and targeting, but the framework concerns the economic purpose directing those tools. Assessing a particular service therefore requires evidence about its data practices and incentives; the presence of machine learning alone does not establish that the framework applies.

Three practical examples

A product review traces whether activity collected to operate a service is also used for a separate commercial profiling purpose.
A media-literacy exercise compares advertising selected from the current page's topic with advertising selected using an accumulated behavioral profile.
An AI governance discussion examines who benefits from a prediction system, what information it uses and whether affected people can understand or contest the arrangement.

Think of it this way

Imagine a shop assistant who remembers which shelf you asked about. Now imagine a separate arrangement that records your movements across many shops and sells predictions about what you might do next. The distinction is not simply that both assistants observed something. It concerns purpose, scope, commercial relationships and control. This fictional analogy helps frame questions; it does not establish that a particular observation or business practice is unlawful.

Worked example

Consider an original hypothetical service with 1,000 user sessions. It records 10 activity events per session, producing 10,000 event records. Suppose an internal review identifies 4 event types per session used for the stated service function and 6 used only by a separate commercial targeting process. Under those simplifying assumptions, 4,000 records support the stated function and 6,000 feed the additional purpose. The latter figure is 60 percent of the modeled records. This is an inventory calculation, not a legal threshold or a measure of harm. Ten thousand event records are not ten thousand people, and the classification depends on verifying the actual purposes. The example provides a way to ask concrete questions about a data flow rather than diagnosing a business model from a percentage alone.

Illustrative case study

In a fictional online-learning company, a proposed assistant would use reading activity to recommend the next lesson. A second proposal would reuse the same activity to create audience segments for external promotions. The review separates these purposes, maps who receives which outputs and asks what users have been told and can control. The team evaluates the second proposal as a distinct business decision instead of assuming that approval of lesson recommendations covers every reuse. The scenario illustrates how the framework can prompt scrutiny of incentives and data use. It does not claim that a particular company follows these practices or that all personalization has the same social effects.

Common misconceptions and mistakes

The term is not interchangeable with all analytics, all advertising or all AI. Nor does it necessarily mean that a company sells raw personal records; commercial value can come from predictions, targeting or other services. Another mistake is treating a critical framework as an automatic factual verdict about an organization without examining evidence. Zuboff's argument should be identified as her analysis, especially when discussing its broader claims about autonomy and power. A further error is confusing numerical prediction accuracy with social acceptability: they answer different questions about the same system.

Business use

The concept can help teams examine how a proposed revenue model depends on information about people. Map the stated service purpose, additional uses, recipients and incentives in understandable language. For an AI product, separate the benefit of a recommendation to the user from the benefit of a prediction to another customer. Review what meaningful choices exist and how the service would function with less collection. These questions support a substantive product discussion, but a label cannot replace evidence about the design. Record the actual decisions and data flows so later changes can be assessed against what was originally agreed.

Check your understanding

How many event records did the toy service collect?

Answer: Ten thousand.

How many were assigned to the additional purpose under its assumptions?

Answer: Six thousand.

Why does that not establish a legal violation or a quantified harm?

Answer: The calculation only classifies modeled records by purpose.

What distinction did the fictional learning company examine?

Answer: User-facing lesson recommendations versus a separate use for external promotions.

Related terms

Profiling
Recommender System
Data governance
Behavioral data
targeted advertising
recommender systems
information asymmetry

Learn next

Choose a fictional digital service and draw its data flows from user action to collected records, inferred outputs and recipients. Give each flow an explicit purpose. Then compare a design limited to the core service with one containing an additional commercial use. Explain the differences without assuming that every use has identical benefits, risks or justification.

Sources

Shoshana Zuboff, Surveillance Capitalism and the Challenge of Collective Action, New Labor Forum, 2019: https://www.oru.se/contentassets/981966a3fa6346a8a06b0175b544e494/zuboff-2019.pdf Shoshana Zuboff, author statement and research background: https://shoshanazuboff.com/book/about/ The event-count calculation, shop analogy and learning-company scenario are original educational illustrations. The entry attributes the framework to its proponent and does not use it to make unsupported allegations about named organizations or to determine legal compliance.

Previous: Surrogate model Next: Survey