Sense and Sensitivity: “Reasoning” Models are More Robust, but can Diverge from Human Consensus in a Legal Interpretation Task

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A study directly compared human and Large Language Model (LLM) judgments in an English-language legal interpretation task, specifically examining metalinguistic insights like whether a DIY gun kit constitutes a "firearm." The research addressed prior gaps by testing "reasoning" models and directly comparing robustness. Findings indicate that LLMs exhibited greater sensitivity to irrelevant prompt features compared to human participants. While enabling reasoning capabilities improved the stability of LLM responses, even these "reasoning" model outputs showed only moderate correlations with human judgments. Furthermore, LLMs sometimes produced interpretations that no human reached for the same prompt. The study concludes that despite reasoning decreasing prompt sensitivity, LLMs remain poor proxies for human metalinguistic judgments in this domain.

Key takeaway

For Research Scientists and NLP Engineers developing legal AI applications, you should exercise extreme caution when relying on Large Language Models for metalinguistic legal interpretations. Even with "reasoning" capabilities, LLMs demonstrate significant prompt sensitivity and can diverge from human consensus. Do not assume LLM outputs accurately reflect human judgment; rigorous human validation is essential to avoid misinterpretations and ensure reliability in critical legal contexts.

Key insights

LLMs, even with reasoning, are poor proxies for human metalinguistic legal judgments.

Principles

• LLMs are more prompt-sensitive than human participants.
• Reasoning improves LLM response stability.
• LLM judgments can diverge from human consensus.

Method

The study directly compared human and LLM judgments (with and without reasoning) on an English-language legal interpretation task to assess robustness.

In practice

• Do not rely on LLMs for metalinguistic legal judgments.
• Use reasoning prompts to enhance LLM stability.

Topics

• Legal Interpretation
• Large Language Models
• Prompt Sensitivity
• Metalinguistic Judgments
• AI Robustness
• Human-AI Divergence

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — Human and LLM everyday reasoning share pattern-matching mechanisms, leading to similar error patterns.
• When Large Language Models Fail in Healthcare: Evaluating Sensitivity to Prompt Variations — Healthcare LLMs are not intrinsically safe, exhibiting high sensitivity to prompt variations that can lead to harmful clinical outputs.
• Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions — Multi-Legal-Bench reveals LLM legal reasoning varies significantly across jurisdictions and tasks, with transfer quality tied to label-set alignment.
• Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law — LLM legal reasoning performance is inflated by data contamination; neuro-symbolic systems offer superior robustness and generalization.
• When large language models are reliable for judging empathic communication — LLMs reliably judge empathic communication, matching experts and outperforming crowdworkers when properly benchmarked.
• LLMs Can Better Capture Human Judgments--With the Right Prompts — Asking better questions to LLMs through specific prompting strategies improves their ability to capture human judgments and response distributions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.