💬 NLP

Can LLMs Reason in a Legally Meaningful Manner? A Small-scale Study on European Court of Human Rights Cases

This study evaluates a top-tier LLM's ability to perform legally meaningful reasoning on European Court of Human Rights cases, finding that while the model produces structurally complete but substantively shallow analyses and automated evaluators fail to align with human judgment, expert prompting improves reasoning depth without enhancing prediction accuracy.

Original authors: Amogh Raina, Ilias Chalkidis, Daniel Hershcovich, Henrik Palmer Olsen

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amogh Raina, Ilias Chalkidis, Daniel Hershcovich, Henrik Palmer Olsen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of law, a judge's decision is never just a guess; it is the result of a careful, step-by-step process of reasoning. A judge must look at the facts of a case, check if the law was followed, decide if the government had a good reason for its actions, and finally weigh whether those actions were necessary in a free society. This process is what gives legal rulings their authority and allows people to understand why a decision was made. Today, powerful computer programs known as large language models can read vast amounts of text and answer complex questions, leading many to wonder if these machines can also perform this kind of deep legal reasoning. The question is not just whether a computer can predict the outcome of a court case, but whether it can explain its thinking in a way that a human lawyer would recognize as sound and meaningful.

A team of researchers set out to test this idea using the European Court of Human Rights, a real court that hears cases about whether countries have violated the rights of their citizens. They focused on a specific group of cases involving freedom of expression, where the court must decide if a government's restriction on speech was justified. The researchers took thirty recent cases from the court's official records and asked a top-tier computer model to act as a judge. They gave the model the facts of each case and asked it to produce a legal assessment, explaining its reasoning step-by-step before predicting the final verdict. To see how well the model did, they compared its explanations against the actual reasoning used by the real human judges in those cases. They also tested the model under different conditions: once with no special instructions, once with a detailed guide written by a legal expert, and once with the official guidebook for that specific law, asking the model to figure out the reasoning steps on its own.

The results revealed a significant gap between what the computer could do and what is required for true legal reasoning. The model was surprisingly good at following the basic structure of a legal argument. In almost every case, it produced an answer that looked like a legal judgment, with distinct paragraphs for each step of the analysis. It correctly identified that a restriction on speech had occurred, checked if there was a law behind it, and considered if the government had a valid goal. However, when the researchers looked closer at the quality of the thinking, they found the model's reasoning was often shallow. While the structure was there, the substance was missing. The model frequently offered vague conclusions instead of firm decisions, relied too heavily on findings from lower courts that had been overturned, and failed to grasp the subtle, practical balancing acts that real judges perform. For instance, in one case involving a prisoner's right to receive mail, the real court understood that reviewing thousands of photocopies for security risks was an unreasonable burden on prison staff, a nuance the model missed entirely.

Perhaps the most surprising finding was that the quality of the model's reasoning had almost nothing to do with whether it guessed the correct outcome. The model predicted the right answer about eighty percent of the time, but this accuracy was largely because it simply guessed that a violation had occurred, which happened to be the most common result in the cases they studied. When the model got the answer right, it was often just as shallow in its reasoning as when it got the answer wrong. This suggests that the model is not truly "thinking" through the legal problems; it is recognizing patterns and guessing the most likely result based on the majority of past cases. Even when the researchers gave the model a detailed, expert-written guide on how to reason, the model produced more comprehensive explanations, but this did not make its predictions any more accurate.

The study also examined how to measure the quality of these computer-generated legal arguments. The researchers had human law students and other computer models act as judges to grade the reasoning. They found that while the computer judges were consistent with each other, they did not agree well with the human experts. The computer judges tended to be more lenient and missed the deeper flaws in the reasoning that the humans caught. This indicates that using another computer program to grade the work of a first computer program is not a reliable way to ensure quality. The human experts found that the model's reasoning, while structurally complete, lacked the depth and certainty of a real legal judgment. The model would often hedge its bets or fail to commit to a clear conclusion, whereas a real judge must be decisive.

Ultimately, the research suggests that while these advanced computer models can mimic the outward appearance of legal reasoning, they do not yet possess the ability to reason in a legally meaningful way. They can produce text that looks like a court opinion, but the logic inside is often superficial and disconnected from the complex realities of the law. The study warns that relying on these models for legal decisions, or trusting automated systems to evaluate their own work, is dangerous. The accuracy of a prediction should not be mistaken for the quality of the reasoning behind it. Until these systems can demonstrate a genuine understanding of legal principles rather than just pattern matching, they should remain tools to assist human lawyers, not replace the human judgment that is essential to the justice system.

Technical Summary: Can LLMs Reason in a Legally Meaningful Manner?

Problem Statement

While Large Language Models (LLMs) have demonstrated strong performance in reasoning tasks across mathematics, coding, and logic, their application to complex, legally oriented tasks remains underexplored. Specifically, the quality of reasoning in legal case forecasting—where reasoning serves as an explainability factor beyond mere pattern matching—has not been rigorously benchmarked. The authors investigate whether recent top-tier LLMs can reason in a "legally meaningful manner" within the context of the European Court of Human Rights (ECtHR), focusing on Article 10 (Freedom of Expression) cases. The study posits that in legal domains, appropriate reasoning is a "virtue" that legitimizes judicial bodies, distinct from and potentially more important than predictive accuracy.

Methodology

Dataset and Task

The study utilizes a small-scale dataset of 30 recent ECtHR cases (published from April 2025 onwards) related to Article 10 of the European Convention on Human Rights (ECHR). The dataset was curated from the HUDOC database to minimize overlap with the model's training data, though the authors acknowledge a residual risk of data contamination given the model's knowledge cutoff (August 31, 2025).

Input: Curated case facts and the relevant legal framework as presented in the court's decision.
Task: The model must assess the facts, reason about potential violations of Article 10, and predict the outcome (violation vs. no violation).
Ground Truth: The ECtHR's official assessment (reasoning) and decision.

Model and Prompting Strategies

The study evaluates OpenAI GPT-5.4 (Knowledge cutoff: Aug 31, 2025) under three distinct prompting settings:

Non-Curated (A): The model is instructed to assess the case and format the output as a list of titled paragraphs but receives no explicit guidance on the specific reasoning strategy for Article 10.
Curated (Expert) (B): The prompt includes a detailed, step-by-step description of the ideal reasoning strategy (based on ECtHR jurisprudence) curated by a legal expert. The model is instructed to follow this specific structure.
Curated (Guide) (C): The prompt provides the official ECtHR Case-Law Guide for Article 10. The model is first instructed to infer the reasoning strategy from this guide and then apply that inferred strategy to the case.

Evaluation Protocol

The study prioritizes the quality of the reasoning (assessment) over the prediction (decision).

Human Evaluation: Three senior law students, supervised by an expert, rated the model's outputs based on:
- Step Occurrence: Binary check if the four doctrinal steps (Interference, Lawfulness, Legitimate Aim, Necessity) are present.
- Step Comprehensiveness: A Likert scale (1–5) rating the depth of each step compared to the court's assessment.
- Overall Conciseness: A Likert scale (1–5) rating the absence of irrelevant information.
LLM-as-a-Judge: Three top-tier models (GPT-5.5, DeepSeek V4 Pro, Claude Opus 4.7) were used to evaluate the same outputs using the same criteria.
Metrics: In addition to reasoning quality, the study measured predictive accuracy, factual reference overlap (F1-score), and internal token usage (billed reasoning tokens vs. output tokens).

Key Contributions

Dataset Release: A small dataset of 30 recent ECtHR Article 10 cases, including human and LLM evaluation results, designed to assess legal judgment forecasting and reasoning capabilities.
Prompting Analysis: An exploration of how different prompting strategies (zero-shot, expert-curated, and guide-inferred) affect the reasoning performance of a state-of-the-art LLM.
Decoupling Accuracy and Reasoning: A demonstration that predictive accuracy does not correlate with reasoning quality in legal tasks.
Evaluation Validity: A critical assessment of "LLM-as-a-Judge" reliability, finding that while LLM evaluators are internally consistent, they align only weakly with human legal experts.

Results

Reasoning Quality vs. Predictive Accuracy

Accuracy: All prompting settings achieved predictive accuracy between 77% and 82%. However, this performance is largely driven by the class imbalance of the dataset (77% of cases involve a violation). A trivial "always-violation" baseline achieved 77% accuracy, matching settings B and C.
Reasoning Quality: There was a significant divergence between accuracy and reasoning quality. The Curated (Expert) setting (B) produced the most comprehensive reasoning (highest human scores) but did not yield higher predictive accuracy than the Non-Curated setting.
Correlation: The study found essentially no correlation ( $r = 0.08$ ) between reasoning comprehensiveness and prediction correctness. Correct predictions were often made with shallow reasoning, and incorrect predictions were sometimes accompanied by comprehensive reasoning.

Step-wise Analysis

Step Occurrence: The model performed the required doctrinal steps in nearly all cases (94–100%), even without explicit prompting (Setting A), suggesting the model has internalized the standard Article 10 structure.
Step Comprehensiveness: The Curated (Expert) setting yielded the highest comprehensiveness scores, followed by Curated (Guide) and Non-Curated. However, even the best setting scored an average of ~4.4/5, indicating reasoning that is "far from ideal" compared to the court's assessment.
Weaknesses: The model frequently struggled with Step 2 (Lawfulness) and Step 4 (Necessity/Proportionality). Qualitative analysis revealed recurring failure modes: vague conclusions where the court is firm, over-reliance on overturned domestic court findings, shallow proportionality analysis that misses practical balancing (e.g., prison security vs. individual rights), and a bias toward predicting violations.

Evaluation Reliability and Validity

Human Agreement: The three human annotators showed high raw agreement (97% exact on step occurrence) but low chance-corrected coefficients ( $\kappa \approx 0.14$ ), attributed to the restricted range of scores (most ratings were high).
LLM vs. Human: The LLM judges were highly consistent with each other (Judge-Judge $\alpha = 0.41$ ) but showed weak alignment with human consensus ( $\rho = 0.16–0.33$ ). The authors conclude that LLM judges are reliable but not valid substitutes for human evaluation in this domain, as they may reinforce shared model biases rather than human legal judgment.

Internal Reasoning

The model utilized significant internal reasoning tokens (1.7k–3.5k billed tokens) which were largely invisible in the output. The emitted "reasoning summary" was often empty, indicating that the visible output is a compressed, potentially unfaithful digest of the internal computation.

Significance and Claims

The paper concludes that while top-tier LLMs can structurally reproduce the doctrinal framework of legal reasoning, their substantive reasoning remains shallow. The study challenges the community to:

Avoid relying solely on automated LLM-based evaluation for legal reasoning tasks, as LLM judges do not align well with human experts.
Stop using task accuracy as a proxy for reasoning quality, as accurate predictions can be achieved through majority-class bias or pattern matching without genuine legal understanding.
Exercise caution in deploying LLMs in high-stakes legal settings, as the "fluent" assessment provided by the model may not reflect the actual reasoning process or the necessary nuance of legal judgment.

The authors emphasize that their findings are limited to a single model (GPT-5.4) and a specific legal domain (ECtHR Article 10), but the decoupling of accuracy and reasoning quality suggests a broader limitation in current LLM capabilities for complex legal reasoning.