“Hate speech,” “qualified candidate,” “biased” — these terms are vague. The standard assumption is that vagueness is a deficiency. A sufficiently precise definition would eliminate it. Language models, trained to produce definite outputs, resolve the vagueness automatically: this post is hate speech, this candidate is qualified, this dataset is biased. The ambiguity collapses into a decision.
Gur-Arieh, Wang, and Fazelpour argue that the collapse is the problem. The vagueness in these terms is not a deficiency to be eliminated but a site where negotiation occurs. When a hiring committee argues about what “qualified” means for a particular role, the argument itself is doing epistemic work — surfacing assumptions, forcing justification, producing a contextual standard that no pre-existing definition could have supplied. The concept is deliberately under-specified because the specification is supposed to emerge from the deliberative process.
An LLM that resolves the ambiguity skips the process. It doesn't produce a wrong answer — it produces an answer that was never subjected to the mechanism that would have made it legitimate. The taxonomy of damage runs across three levels: at the process level, the skills of deliberation atrophy; at the output level, the concept distorts because it was resolved without context; at the ecosystem level, the vocabulary itself narrows as automated resolutions become the default reading.
The through-claim: resolution and legitimacy are not independent. An answer produced by bypassing the process through which the answer was supposed to be negotiated isn't a faster version of the same answer — it's a structurally different kind of output. The speed is not incidental to the problem; it is the problem. Legitimate resolution requires the friction of disagreement, the time of deliberation, the exposure to competing interpretations. Remove the friction and you don't get the same destination faster. You get a different destination that looks the same.
This is not specific to language models. Any system that automates the resolution of a deliberately ambiguous standard — bureaucratic checklists replacing professional judgment, algorithms replacing editorial discretion, standardized tests replacing holistic evaluation — produces the same structural defect. The answer arrives without the process that would have made it an answer rather than an imposition.