The Prompted Imagination

In the AI Story

Every software environment contracts the user's sense of the possible to the space the software services—the spreadsheet user thinks in rows and columns, the presentation user thinks in slides. The contraction is normally visible in the medium's formal constraints. AI tools conceal the contraction because they operate in natural language: the interaction feels unconstrained. The builder describes problems in ordinary speech, receives solutions in code or prose, experiences no visible limitation. But the model's outputs cluster around its training distribution. Statistically likely solutions appear; statistically unlikely solutions are systematically suppressed not through censorship but through probability. Over time, through habitual interaction, the builder's imagination is shaped by this statistical landscape—learning to navigate the territory the model covers while losing visibility of what lies beyond.

Segal's laparoscopic surgery example (the analogy Claude generated connecting ascending friction to medical technique) illustrates both the value and the limitation. The connection was genuine—Segal had not seen it, Claude made it visible, the insight strengthened the argument. But what connections did Claude not make? What examples fell outside its training data? What disciplines, traditions, bodies of knowledge were absent from the associative range the model could traverse? Segal cannot answer—the absence is invisible. And through thousands of such interactions, the builder's imagination is tuned to the model's range. Not deliberately. Through habituation—the automatic, unreflective shaping of expectation by experience.

The risk is particularly acute for the generation entering professional practice with AI tools as their primary collaborative medium. These builders have no pre-AI baseline against which to measure the contraction. Their entire professional imagination has been formed in dialogue with a model whose characteristic outputs defined, from the start, what seemed possible. The older generation can remember what it felt like to imagine without prompting; the younger generation has no such memory. Whether this produces a categorically different kind of imagination—narrower in some dimensions, broader in others—or simply a differently structured one is an empirical question the next decade will answer.

Origin

Chun's concept builds on her career-long investigation of how media shape cognition. Programmed Visions (2011) established that software structures perception by organizing memory—determining what is stored, retrieved, and presented as relevant. The prompted imagination extends this into generative systems: AI does not merely curate existing information (Google search) or display organized data (databases). It produces information that reflects its training distribution. The builder's imagination is shaped not by what they see displayed but by what they see generated—and what gets generated is determined by statistical regularities learned from the corpus.

The concept has intellectual debts to Benjamin Lee Whorf's linguistic relativity (the language you speak shapes what you can think), Thomas Kuhn's paradigms (the framework you adopt determines what problems you can see), and Marshall McLuhan's medium theory (the medium is the message; the tools reshape the user). Chun's synthesis specifies the mechanism: the prompted imagination is shaped through habitual interaction that consolidates statistical tendencies into perceptual defaults. The builder does not choose the contraction; it is deposited through repetition, below awareness, as the automatic by-product of sustained engagement.

Key Ideas

Imagination shaped by outputs received. The builder's sense of what is possible contracts over time to match the range of outputs the AI model characteristically produces—learning the model's territory while losing visibility of what lies beyond it.

Absence is invisible. The builder cannot see solutions never generated, connections never made, approaches outside the training distribution—the constraint operates through what does not appear, leaving no trace to examine.

Natural language conceals formal constraint. Because the interface is ordinary speech, the interaction feels unconstrained; the actual constraint is in the model's statistical architecture, invisible to the user engaged in apparently free conversation.

Homophily contracts diversity. Like attracts like: the model learns patterns from training data, generates outputs conforming to those patterns, and over time trains the user to expect and imagine within the pattern-space rather than beyond it.

The youngest generation has no baseline. Builders entering practice with AI as their primary collaborative medium have no pre-AI imagination against which to measure what habituation has shaped—their sense of the possible was formed from the start within the model's range.