Spandrels of AI Systems

In the AI Story

The spandrels paper challenged the adaptationist programme's systematic error of treating every organismal feature as shaped by natural selection for its current function. Some features are adaptations. Others are spandrels (byproducts of other features), exaptations (co-opted from original functions), or products of developmental constraint. The human chin is Gould's paradigmatic spandrel—not an adaptation for anything but the geometric consequence of independent reduction in two jaw growth fields. The error of treating it as adaptive (chin-for-talking, chin-for-beauty) mistakes architectural necessity for designed purpose.

Gould himself drew the computational connection in his 1997 PNAS elaboration: 'Just consider the obvious analogy to much less powerful computers. I may buy my home computer only for word processing… but the machine, by virtue of its requisite internal complexity, can also perform computational tasks exceeding by orders of magnitude the items of my original intentions.' Unintended capabilities are spandrels—not designed, not adaptations, but necessary byproducts of sufficient complexity. The application to large language models is direct: reasoning, creativity, and hallucination all emerge from complexity required for text prediction.

The practical consequence transforms how AI limitations should be understood. If hallucinations are bugs, the engineering response is to fix them through verification layers, grounding, curated training data. These interventions reduce frequency but cannot eliminate hallucinations entirely—they treat a spandrel as malfunction. The architecture producing hallucinations is the architecture producing fluency. You cannot remove the spandrel without altering the arch. Edo Segal's discovery of Claude's confident misuse of Deleuze illustrates the dynamic: fluent philosophical prose (adaptation performing as designed) containing a factual error (spandrel of the same statistical machinery applied where fluency-accuracy correlation is lower than the model's confidence suggests).

The adaptationist fallacy applied to AI produces two equal errors. First: assuming every impressive output evidences designed intelligence—that the system 'understands' concepts, 'reasons' through problems, 'creates' combinations. This treats every capability as adaptation, ignoring that many are spandrels. Second: assuming every failure evidences fundamental inadequacy—that hallucinations prove the system is 'just' pattern-matching. This treats every limitation as disqualifying flaw. Gould's framework dissolves both by insisting on the prior question: what was this feature shaped to do? Everything the system does well or poorly must be understood relative to text prediction, its shaping function.

Origin

The 1979 spandrels paper emerged from Gould and Lewontin's observation that evolutionary biology had adopted an uncritical adaptationist methodology—every feature assumed to be an adaptation requiring a just-so story about its selective advantage. The spandrels of San Marco provided the architectural metaphor: pendentives are geometric necessities that appear designed for decoration but actually arise as structural consequences of dome-on-arches construction. The paper was polemical, deliberately confrontational, naming the 'Panglossian paradigm' after Voltaire's Dr. Pangloss who believed this is the best of all possible worlds. It reshaped how evolutionary biologists approach explanation, forcing the question: is this feature an adaptation or a byproduct?

Key Ideas

Architectural necessity vs. designed purpose. Features can arise as geometric consequences of other structures rather than as shaped-for-function adaptations—the pendentive is necessary given arches and dome.

Text prediction as shaping function. Everything LLMs do must be understood relative to their training objective—fluent next-token prediction—which is their only genuine adaptation.

Reasoning and creativity as spandrels. Neither was a training objective; both emerged as structural consequences of complexity sufficient for fluent context-sensitive prediction.

Hallucinations as geometric inevitability. A system predicting probable tokens without truth-discrimination necessarily generates confident falsehoods when probability-of-assertion and probability-of-truth diverge.

You cannot remove spandrels without altering architecture. The features producing hallucinations are the features producing fluency—they are structurally inseparable.