The Fluency Illusion: Why Perfect Outputs Don't Prove AI Understands
Quine's indeterminacy problem shows LLMs can produce outputs consistent with multiple incompatible interpretations, making true comprehension fundamentally unverifiable.
Drawing on Quine's philosophy of language, this piece argues that LLMs inherit an irreducible epistemological problem: their confident outputs can correspond to multiple incompatible internal representations, leaving no way to verify genuine comprehension. While alignment techniques mitigate surface errors, they cannot bridge the gap between statistical token prediction and understanding human intent. The most dangerous failures—outputs that appear correct yet fundamentally miss the point—may be epistemically invisible.
In 1960, Quine demonstrated that a field linguist hearing “gavagai” could never determine whether the word referred to rabbit, undetached rabbit parts, or a temporal rabbit-slice—because all three interpretations are perfectly consistent with the same observable behavior. Every LLM prompt reproduces this indeterminacy: the model’s fluent output can align with multiple, incompatible internal representations, and you have no way to verify which one it selected. Alignment techniques reduce surface errors but cannot dissolve the fundamental gap between token statistics and shared human intent. The failure mode isn’t that the model might misunderstand you—it’s that an output that looks correct is epistemically indistinguishable from one that has catastrophically misinterpreted your meaning.