The Grue Paradox

In The You On AI Field Guide

The grue paradox upended the Humean/Carnapian program of reducing induction to logical relations between evidence and hypothesis. If 'grue' and 'green' are equally well-confirmed by the same evidence, then confirmation is not a logical relation—it depends on something extra-logical, something conventional. Goodman called that something entrenchment, and he argued that entrenchment is built through the historical success of predicates in prior inductions. We project 'green' and not 'grue' because 'green' has a track record. The track record is not evidence in the ordinary sense—it is meta-evidence about which predicates are worth projecting. The conventional character of projectibility aligned Goodman's philosophy of science with his aesthetics: both domains involve symbol systems whose standards of rightness are internal to the system, conventional, and irreducible to correspondence with a version-independent reality.

The paradox has a direct application to AI systems trained to induce generalizations from data. A large language model is, in effect, a massive inductive engine—it observes patterns in training text and projects those patterns onto new contexts. The projection is determined by the model's architecture and training regime, which encode preferences for certain kinds of patterns over others. These preferences are the model's entrenchment structure—its implicit ranking of which predicates are projectible. But the entrenchment is extracted from human linguistic practice, not grounded in the model's own inductive successes or failures. The model has no track record of its own. It has only the track record of the humans whose language it was trained on, and the track record is embedded in the training data in ways that are statistically recoverable but not rationally transparent.

What this means for AI-generated predictions and generalizations is that the model's inductive inferences inherit the entrenchment of the predicates humans have projected, but they do so without the rational grounding that human entrenchment possesses. Humans project 'green' because 'green' has worked in past inductions. The model projects 'green' because 'green' appears with high frequency in contexts where humans made successful inductions. The frequency is a proxy for entrenchment, not entrenchment itself. And the proxy breaks down when the model encounters novel contexts where the frequency-based heuristic diverges from the rational grounds that would determine projectibility if the model were capable of rational induction. The result is outputs that look like inductive knowledge—they generalize from evidence, they make predictions, they project predicates onto new cases—without being grounded in the rational warrant that inductive knowledge requires. Plausible induction without rational warrant: the grue paradox anticipated the epistemological hazard of AI-generated generalizations seven decades before the technology existed.

Origin

Goodman introduced the grue paradox in 'A Query on Confirmation,' Journal of Philosophy 43 (1946), and developed it fully in Fact, Fiction, and Forecast (1955), Chapter III. The paradox became one of the most discussed problems in twentieth-century philosophy of science, generating hundreds of papers and multiple book-length treatments. Rudolf Carnap attempted to solve it within his logical framework; W.V. Quine declared it unsolvable and used it to support his holism; Goodman maintained that entrenchment solved it and that the solution revealed the conventional basis of induction. The debate has never been fully resolved, but the paradox's demonstration that induction depends on extra-logical factors remains one of the most important results in modern epistemology.

Key Ideas

Confirming evidence underdetermines theory. The same observations support incompatible generalizations ('all emeralds are green' vs. 'all emeralds are grue'), so confirmation is not a purely logical relation.

Entrenchment grounds projectibility. Predicates are projectible when they have been successfully used in prior inductions—'green' projects, 'grue' does not, because of their different historical entrenchment.

Induction is conventional. The choice of which predicates to project is conventional, built through practice, irreducible to logic—paralleling the conventional character of symbolic reference in aesthetics.

AI inherits entrenchment without grounds. Models project predicates that are entrenched in training data but lack the rational warrant that grounds human entrenchment—frequency proxies for projectibility without being projectibility.