Data Doubles

Haggerty and Ericson's data double described the abstraction of persons into flows of information that could be recombined, classified, and acted upon. Web 2.0 greatly expanded the resolution of these representations. Search histories, clicks, pauses, locations, social relationships, images, facial features, and other behavioural traces became inputs for systems promising increasingly personalized services. Convenience and surveillance developed from much of the same machinery. Better recommendations required finer distinctions among users; engagement optimization demanded closer observation of attention; personalization depended on increasingly elaborate inferences about what a person might want or do next. Zuboff's surveillance capitalism captured the economic importance of converting behavioural traces into predictions, but the systems built around that logic have since acquired purposes and capacities extending well beyond targeted advertising.

Generative AI intensifies this trajectory. Users increasingly disclose personal histories, relationships, ambitions, anxieties, preferences, and unfinished thoughts directly to commercial machine systems, while personalized conversational agents can maintain context and develop machine-generated identities to which users may form substantial attachments. The resulting representation is more dynamic and potentially more intimate than the consumer profiles associated with earlier Web 2.0 services. A data double is neither a portrait nor a clone. It is a model-dependent and changing representation assembled from observations, derived features, inferred characteristics, and predictions, partly by locating an individual among patterns learned from other people. The same traces can produce different doubles under different models, and each new interaction can alter what the system subsequently infers.

BDSL studies these systems historically as well as computationally. App decompilation and reverse engineering allow successive releases of software to be read as evidence of changing categories, features, measurements, and client-side operations. We examine this material alongside interfaces, version histories, algorithm papers published by Web 2.0 companies, technical documentation, and legal proceedings to reconstruct how systems for recommendation, engagement, facial modification, personalization, and user inference have changed over time. The distinction among observation, derivation, inference, prediction, and intervention remains essential. What a system records is different from what a model infers from it, and both are different from the action taken on the resulting prediction.

Personalization also creates a recursive problem. A system observes behaviour, constructs an inference, and uses that inference to rank a feed, recommend content, alter an image, generate a response, or otherwise modify what the user encounters. That intervention changes the environment in which subsequent behaviour occurs, producing new observations from which the representation is revised. Data Doubles investigates this history of increasingly personalized computation as a feedback relationship between people and systems, asking how machines construct representations of their users and how those representations, once acted upon, begin to shape the conditions under which users encounter the world.