Shira Gur-Arieh (Harvard Law School) has posted LLMs and the Collapse of Legal Indeterminacy on SSRN. Here is the abstract:
LLMs have emerged as a new form of interpretive infrastructure. Hiring managers use them to identify “qualified” candidates. Content platforms deploy them to determine what constitutes “hate speech”. Judges have begun using models to assist in interpreting legal texts. Across these settings, LLMs confront ambiguous, open-textured terms—positioning themselves as arbiters of meaning, fixing definitions of concepts that admit multiple legitimate interpretations. This paper asks how LLMs handle, and mishandle, ambiguity in interpretive legal tasks.
The paper draws on analytic jurisprudence in which ambiguity is a structural feature of law. On these accounts, legal texts are open-textured so statutes written today can absorb tomorrow’s technologies, crises, and moral disputes without constant revision. Ambiguity underwrites adversarial practice by allowing litigants to contest meaning and requiring courts to justify their resolutions; it also sustains institutional balance, as legislatures articulate broad principles, agencies implement them, and courts step in when principles conflict or rules run out.
What happens when technology encounters law’s ambiguous character? Earlier forms of “code-driven law”—from smart contracts and “rules-as-code” projects to automated welfare systems—exposed a fundamental incompatibility between legal language and code. Unlike language, code cannot tolerate ambiguity: every term must be operationalized, every threshold specified, every interpretive choice fixed at design time. Translating legal norms into code therefore requires ex-ante resolution of ambiguity. This, in turn, makes legal norms more rigid, limiting their capacity to adapt and shifts interpretive authority from legislatures and courts to system designers.
At first glance, LLMs appear to offer an escape from this rigidity. LLMs are trained on natural language, learning from the pluralism and richness of human discourse. Unlike rule-based systems that must fix criteria ex-ante at design time, LLMs generate answers at inference, enabling context-sensitive, case-by-case judgments. And because they output text—the medium through which law expresses meaning—they appear better suited to preserve nuance, sustain competing interpretations, and acknowledge uncertainty.
Yet this paper argues that LLMs fail to realize this promise. Although they represent a genuine technical advance, they clash with legal language’s indeterminacy in ways that are more subtle, and potentially more insidious, than the clash produced by rule-based systems. Indeterminate legal language imposes an irreducible burden of judgment, and that burden has to be discharged somewhere. The paper traces three configurations of where that burden ends up: absorbed by the model, pushed downstream onto the user, or pulled upstream by the deployer. Each changes the institutional conditions under which legal judgment is exercised, and each risks undermining the functions that indeterminacy was meant to preserve.
A very interesting paper. The abstract is a bit misleading in the way it discusses “ambiguity,” whereas the paper itself carefully distinguishes “ambiguity,” “vagueness,” and “open texture.” For discussion of these concepts and their relationships, see the Legal Theory Lexicon entry on Vagueness and Ambiguity. For discussion of legal indeterminacy and the related but more precise concept of underdeterminacy, see my On the Indeterminacy Crisis: Critiquing Critical Dogma, 54 U. Chi. L. Rev. 462 (1987).
Highly recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum