Spandrels of AI Systems

In 1979, Stephen Jay Gould and Richard Lewontin used the decorated triangular spaces (spandrels) between arches in Venice's San Marco basilica to critique adaptationist thinking in biology. The spandrels are geometric necessities of mounting a dome on arches—they could not not exist—yet they appear designed to hold mosaics. This became their paradigm for features that are structural byproducts mistaken for designed adaptations. Applied to AI, spandrels are capabilities that emerge as necessary consequences of architectural complexity rather than as trained objectives. Large language models were optimized to predict the next token in text sequences—that is their adaptation. But they exhibit reasoning-like behavior, creative synthesis, and hallucinations as spandrels: structural consequences of the complexity required for fluent prediction. Hallucinations are not bugs to be fixed but geometric inevitabilities of an architecture that generates text by predicting probable continuations without distinguishing between tokens probable-because-true and tokens probable-because-they-match-patterns-of-confident-assertion.