Three Configurations of Human-AI Creation
Moles distinguishes three configurations of the human-AI compound channel, each with distinct information-theoretic properties. In the first configuration, the human provides the semantic message and the AI provides the aesthetic elaboration; the information content of the output is bounded by the semantic information supplied by the human. In the second configuration, the AI generates candidate aesthetic messages and the human selects; here, the AI provides entropy and the human provides the filter that transforms entropy into information. In the third configuration, human and AI engage in iterative exchange, each modifying the other's output; the information content may exceed what either channel could produce independently, because the interaction itself generates new information through the collision of incompatible coding systems. Only the third reliably produces the supersignal.
In the AI Story
The typology is diagnostic. A user can observe their own practice and identify which configuration dominates. A user who describes what she wants in detail and accepts the AI's substantial output without iteration is operating in configuration one. A user who generates multiple candidates and selects is operating in configuration two. A user who engages in sustained back-and-forth, modifying the AI's suggestions and having the AI modify hers, is in configuration three.
Each configuration has appropriate uses. Configuration one is efficient for well-specified tasks where the human's semantic input fully determines the desired output — routine code generation, formulaic documents, competent first drafts of standard genres. Configuration two is effective when the human has clear selection criteria but cannot easily generate candidates — visual design, naming, divergent brainstorming. Configuration three is required when the desired output exceeds what either party could produce alone — genuine creative collaboration, novel problem-solving, the production of aesthetic information.
The danger is that the configurations are not always visible to the user, and configuration drift is easy. A session that begins in configuration three can drift into configuration one as the human tires and accepts AI output without modification. A session intended as configuration two can collapse into passive acceptance of whatever the AI produces. The drift is toward reduced human engagement, because reduced engagement is easier and the tool will fill the space the human vacates.
The practical consequence is that supersignal emergence — the outcome that justifies the collaboration on creative rather than productivity grounds — requires sustained attention to the configuration the collaboration is in. It is work the tool will not do. It is the work that distinguishes genuine creative partnership from the productive addiction that masquerades as it.
Origin
The three-configuration typology extends Moles's analyses of different coupling arrangements in mass communication, where similar distinctions apply between producer, editor, and audience roles. The AI context maps these distinctions onto the compressed time-scale of a single creative session.
Key Ideas
Configuration one: semantic-plus-elaboration. Human specifies, AI fills in; output bounded by human's semantic input.
Configuration two: generate-and-filter. AI produces candidates, human selects; entropy meets filtering.
Configuration three: iterative exchange. Both parties modify each other's output; supersignal emergence becomes possible.
Drift is toward lower engagement. Sessions tend to slide from three toward one as fatigue sets in.
Only three reliably produces genuine novelty. The other configurations are efficient for bounded tasks; only iterative exchange generates the information that justifies partnership.
Appears in the Orange Pill Cycle
Further reading
- Abraham Moles, Théorie structurale de la communication et société (Masson, 1988)
- Edwin Hutchins, Cognition in the Wild (MIT Press, 1995)
- Lucy Suchman, Human-Machine Reconfigurations (Cambridge University Press, 2007)