Imagination or Reason? A Spinozist Criterion for Genuine Analogical Reasoning in Large Language Models

When Webb, Holyoak, and Lu reported that large language models (LLMs) solve a wide range of analogy problems zero-shot, they reopened an old philosophical question: do such systems reason relationally, or do they exploit surface regularities absorbed from training? Critics show that performance degrades on counterfactual variants that leave human performance intact; defenders answer that these failures reflect peripheral deficits. I argue that the debate lacks a principled criterion for distinguishing genuine relational reasoning from associative pattern-completion, and that Spinoza’s epistemology supplies one. Spinoza separates imagination—general ideas formed through the contingent order in which the body is affected—from reason, the adequate cognition of what is genuinely common through common notions; his own illustration (Ethics 2p40s2) is a proportional analogy that turns on the difference between applying a learned rule and grasping a common form. Reconstructed functionally, the distinction yields a diagnostic heuristic: invariance of competence under transformations that preserve relational structure while destroying surface similarity counts as defeasible evidence of relational grounding. This separates the following two questions the literature conflates: whether a system produces correct analogies, and whether its competence is relationally grounded. I give the criterion a measurement protocol and apply it symmetrically to humans and machines.