The scheme-content relation is Goodman's term for the connection between a symbol system's formal structure and the domain of things, properties, and events to which the symbols refer. The scheme is the set of symbols and the syntactic rules governing their combination—the alphabet and grammar of a language, the pitch-classes and rhythmic values of musical notation, the color palette and compositional conventions of painting. The content is the referential field—the objects, events, properties that the symbols pick out, exemplify, or express. The relation between scheme and content is not given by the scheme alone. It is established by practice, convention, and—critically—by the worldmaking intentions of the symbol-user. When Cézanne deployed the schemes of post-Impressionist painting to construct a version of Mont Sainte-Victoire emphasizing geometric structure over fleeting light, the scheme-content relation was established by his specific intention to reveal the mountain's deep formal organization. The same symbolic scheme (fragmented planes, flattened depth, structural emphasis) could be deployed toward different content—a different mountain, a different aspect of the same mountain, an entirely abstract formal exploration—and the difference would be constituted by the worldmaker's intention in deploying the scheme.
The scheme-content distinction is one of the most contested concepts in twentieth-century philosophy, debated across Quine, Davidson, and their successors. Goodman's version is distinctive in its emphasis on the worldmaker's role in establishing the relation. The symbols do not come pre-attached to their referents. The attachment is made by the symbol-user's intentional deployment of the scheme toward specific content, and the deployment is what makes the symbols function referentially rather than merely existing as marks. This is why Goodman rejected the copy theory: there is no pre-given content that the scheme copies. The content is constituted by the scheme, and different schemes constitute different contents—different versions of reality, each organized according to the symbol system's particular resources.
In AI-generated outputs, the scheme-content relation is established by—what? The training data establishes correlations: these symbolic patterns (schemes) tend to co-occur with these referential patterns (contents). The model reproduces the correlations, generating outputs where the scheme-content pairing conforms to statistical regularities. But conforming to statistical regularities is not the same as intentionally establishing a scheme-content relation, because the latter requires a worldmaker who chooses this pairing for reasons grounded in lived experience and evaluated by purposes. The model has no purposes. It has optimization targets (predict the next token, minimize perceptual distance from a target image), but optimization targets are not worldmaking purposes. They are engineering specifications, external to the model, determined by the designers who built the training regime.
What this means is that AI-generated symbols can denote, exemplify, and express—can perform the referential functions Goodman identified—only when embedded in a human worldmaking project that establishes the scheme-content relation. The human provides the purposes that determine why these symbols, referring to this content, configured in this way. The machine provides the symbols. The scheme-content relation is collaborative, and the collaboration succeeds when the human maintains the worldmaking role—the establishment of purposes, the evaluation of fit, the judgment about whether this particular symbolic configuration is right. When the human delegates the worldmaking to the machine—accepts the machine's scheme-content pairing without evaluating whether it serves purposes the human has independently established—the output may be plausible (it conforms to statistical patterns of right pairings) without being right (it does not serve the purposes that would make the pairing more than arbitrary).
The scheme-content terminology appears throughout Goodman's work but is developed most explicitly in Ways of Worldmaking and in his exchanges with W.V. Quine, whose 1969 'Epistemology Naturalized' rejected the scheme-content dualism. Goodman maintained the distinction while insisting it was internal to worldmaking—schemes and contents are both constructed, but they are constructed in relation to each other, and the relation is established by the worldmaker's intentional activity. The concept has proven durable in philosophy of science (where it illuminates theory-change and the incommensurability of paradigms) and now in AI philosophy, where it identifies the structural feature of intentional symbolic deployment that rendering engines cannot replicate.
Symbols do not come pre-attached to referents. The scheme-content relation is established by the worldmaker's intentional deployment of symbols toward specific referential purposes, not by the symbols' intrinsic properties.
Different schemes constitute different contents. The physicist's and painter's versions of the same mountain are different mountains, constituted differently by the symbolic resources each deploys—worlds, not descriptions.
AI correlates, does not establish. Models reproduce statistical correlations between schemes and contents from training data but do not establish scheme-content relations through purposeful worldmaking intentions.
Relation requires a worldmaker. The pairing of symbols with referents achieves rightness only when grounded in purposes a worldmaker holds—purposes the rendering engine assists but cannot supply.