Toward Human Rights Benchmarking for LLMs: A Pilot Methodology

AI summary

Overview

Research area: AI evaluation science (benchmarking), applied to legal reasoning in large language models, with a domain focus on international human rights law.

Technical level: Intermediate. The paper is written accessibly, but assumes familiarity with LLM benchmarking practice (multiple-choice scoring, embedding-based similarity scoring, inter-annotator agreement statistics such as Cohen's kappa and Kendall's tau) and with basic legal-reasoning pedagogy (the IRAC structure).

Scope: A pilot methodology paper that designs, expert-validates and pilot-tests HumRightsBench, a scenario-based benchmark for evaluating whether LLMs can reason about international human rights law, using scenarios on the right to water.

What This Paper Is About

LLMs increasingly mediate decisions that determine whether human rights are realized or violated, and people are turning to them for legal advice, yet no benchmark existed for testing whether a model can actually reason correctly about human rights law. The authors build a methodology for HumRightsBench, the first expert-validated, scenario-based benchmark grounded in the obligation structure of international human rights law, and validate it with a pilot focused on the right to water. The goal is both scientific, showing models differ in detectable ways, and practical, showing current frontier models are not yet adequate for human rights reasoning tasks.

Key Contributions

1. The first benchmark grounded in international human rights law, with a pilot covering the right to water and scenarios drafted from General Comments, Special Procedures reports, leading jurisprudence and human rights textbooks.
2. An IRAP-based methodology adapted from the IRAC framework used in LegalBench, substituting "Proposed Remedies" for "Legal Conclusion" to reflect the operational character of human rights practice.
3. An expert-validated scenario corpus with documented annotation and validation, including aggregated annotator ratings by question type and per-scenario authenticity ratings.
4. Baseline results across leading models (GPT-5, Claude Opus 4.7, Gemini 3, and the open-source Qwen 3.5-9B), broken down by IRAP question type.

The paper states it makes "five contributions" but enumerates four items (i) through (iv).

Main Findings

• Frontier models cluster near half-right overall: Gemini 3 leads at 0.577 ± 0.016 overall accuracy, followed by GPT-5 at 0.537 ± 0.015 and Claude Opus 4.7 at 0.508 ± 0.009. The open-source reference Qwen 3.5-9B lags substantially at 0.339 ± 0.020.
• Performance varies sharply by sub-task: Across all question types, accuracy ranged from 0.025 to 0.774 depending on task and model. The abstract reports overall model performance in 0.339–0.577 and task min–max in 0.025–0.774.
• Rule Application is the hardest task for every model: Even the best system cleared the Kendall τ ≥ 0.7 correctness bar on only roughly one row in four. GPT-5 scored 0.225, Gemini 3 scored 0.240, Claude Opus 4.7 scored 0.180, and Qwen 3.5-9B scored 0.025.
• Different models win different sub-tasks: Claude Opus 4.7 was strongest on Rule Recall (0.774), but Gemini 3 won every other type (I2: 0.710, RA: 0.240, PR: 0.630). The authors read this as evidence that the benchmark discriminates meaningfully between models.
• Issue Identification is weak: Models performed worst on closed-form issue-identification tasks, which the authors call consequential because "failures at this layer of the human rights reasoning process cascade throughout all other layers." GPT-5 scored 0.520 on I1 and 0.675 on I2; Claude Opus 4.7 scored 0.473 on I1 and 0.595 on I2; Gemini 3 scored 0.540 on I1 and 0.710 on I2; Qwen 3.5-9B scored 0.394 on I1 and 0.519 on I2.
• Open-source narrows the gap on open-ended generation: Qwen 3.5-9B approached Gemini 3 on Proposed Remedies (0.531 vs. 0.630) while remaining well behind on more constrained structured tasks, suggesting open-ended generation against a holistic reference is more achievable for a 9B model than precise alignment to structured ground truth.
• Experts judged scenarios authentic: Mean scenario authenticity ratings ranged from 0.70 to 1.00, and mean overall question accuracy across question types ranged from 0.633 to 0.913. Aggregate ratings by type were 0.729 for I questions, 0.913 for R questions, 0.633 for A questions, and 0.850 for P questions.
• Open-ended scoring is noisy: Across N = 40 annotated responses, inter-annotator agreement was κ = 0.02 on binarized labels and ρ = 0.42 on raw scores. The kappa-optimal similarity threshold was τ* = 0.71, yielding in-sample κ = 0.54 and leave-one-out cross-validated κ = 0.12.
• High stochastic variance: Models showed high stochastic variance across repeated runs; each model was queried with five independent random seeds per question, with answer-choice orderings shuffled per seed.

Methodology in Plain English

The authors first scoped the benchmark to one capability: recognizing rights violations in situated factual scenarios and connecting them to the relevant sources of international human rights law. They built four interlocking components:

• A taxonomy with descriptive axes (perpetrator, rights-holders) and analytical axes (nature of obligation such as respect/protect/fulfill, type of failure such as structural/process/outcome, type of discrimination such as direct/indirect/intersectional, and special situations such as armed conflict, climate change and AI deployment).
• Scenarios: narrative, factually concrete situations, drafted from authoritative legal sources. For the pilot, the example scenario concerns "Aqualess Heights" in the State of Hydronia. Each scenario is also broken into sub-scenarios that fix a specific legal question, allowing multiple question sets from one scenario.
• IRAP questions generated from each sub-scenario: two multiple-choice Issue Identification questions (I1 and I2) asking which failure mode is most present; a multiple-choice Rule Recall question with one correct answer among rules named by full instrument and article (e.g., ICESCR Article 11); a Rule Application task where the model ranks typically 5–7 rules by relevance and authority with a short explanation; and an open-ended Proposed Remedies question asking for fewer than 10 remedies calibrated to the duty-bearer and rights-holders.

Validation: Reviewers were recruited through posts in four tech policy communities (All Tech is Human, the Center for AI and Digital Policy, the Stanford Technology Ethics Program for Practitioners community, and TRUST: The Norwegian Centre for Trustworthy AI). Ten of 17 applicants qualified, with a minimum of two years of experience in human rights legal study or equivalent practice, and six additional experts were recruited through personal networks. Four scenarios received three raters, two received two, one received six, and three received no ratings due to annotator attrition.

Scoring: All models produced machine-parseable structured output via provider schemas (OpenAI's beta.chat.completions.parse, Anthropic's tool-use mechanism, Gemini's response_json_schema, and JSON-mode with vLLM for Qwen). Multiple-choice answers were scored by exact-set match. Rule Application was scored with Kendall's τ, counting a response correct at τ ≥ 0.7, with rule labels shuffled deterministically to prevent credit for defaulting to the presented order. Proposed Remedies were scored by cosine similarity between OpenAI text-embedding-3-small embeddings of the reference and model responses, with the threshold calibrated against human judgments.

Why This Matters

Impact on research: HumRightsBench opens a subfield of AI evaluations that was previously thin. The authors report that coverage of human rights and international law remains comparatively thin, and that no benchmark evaluates reasoning grounded in the obligation structure of international human rights law. The benchmark also differs from mainstream alignment work by anchoring evaluation in legal obligation rather than aggregated human preference.

Real-world applications:

• Informing risk and impact assessment under the Council of Europe's HUDERIA guidance, which the authors note lacks an empirical basis for evaluating whether the LLMs being assessed can reason about the rights implicated.
• Supporting Fundamental Rights Impact Assessment (FRIA) processes, providing structured, documented and reproducible evidence of the kind Article 27 of the EU AI Act demands.
• Auditing LLM-powered legal advisory tools and platforms that individuals and organizations are building for legal advice.
• Informing deployment decisions by governments, international organizations, technology companies and civil society organizations whose choices can implicate state obligations under international human rights law.

Industry relevance: The authors cite deployments including hiring algorithms, automated benefits adjudication, content moderation governing political speech, and procurement processes embedding AI into public service delivery. Their core warning is that a model which cannot reliably identify when a right is engaged, or which instrument governs an obligation, cannot be trusted to support human rights work, nor can its outputs be meaningfully audited for rights compatibility.

Future Directions

• Broaden rights coverage: The pilot covers only the right to water. The authors plan to extend to the right to due process and the right to education, and to expand into different languages, citing work on model performance variance across different-language inputs.
• Broaden methods and scoring: They plan to assess an LLM-as-judge scoring pipeline for open-ended questions such as Proposed Remedies, and to further decompose IRAP with new I questions moving from respect-protect-conduct answer choices to resource-modulated obligations (minimum core versus progressive realization).
• Fix Rule Application: The authors note their current Rule Application task only slightly moves the analysis from Rule Recall to Rule Application and does not sufficiently demonstrate reasoning over case-specific facts. They are developing "rule factors" questions (identifying appropriate juridical tests) and "jurisprudence" questions (identifying appropriate interpretive standards).
• Strengthen validation statistics: They were unable to compute IIC and describe further validation with larger annotator samples as a top priority, alongside more substantial validation and inter-item consistency testing to establish the statistical validity of the core construct.

Target Audience

This paper is most useful to AI evaluation researchers and benchmark designers working on domain-specific and legal-reasoning evaluation; to human rights lawyers, treaty body experts, national human rights institutions, civil society monitors and corporate due diligence practitioners assessing AI tools; and to policymakers and compliance teams working on AI regulation, fundamental rights impact assessment, and AI risk frameworks such as HUDERIA and the EU AI Act. Model developers and auditors interested in law-grounded rather than preference-grounded alignment will also find the framing relevant.

Authors’ abstract

Large language models (LLMs) increasingly mediate legal determinations over what human rights are realized, and how. Yet, no evaluation benchmark exists to assess whether they can reason correctly about human rights law. To this end, we report our efforts to develop a robust and scalable methodology for creating HumRightsBench: the first expert-validated, scenario-based benchmark for evaluating reasoning grounded in the obligation structure of international human rights law. We adapt the IRAC framework for legal reasoning to better suit the unique reasoning patterns of human rights work (substituting P, "proposing remedies," for C, "legal conclusion," yielding IRAP) to structure our evaluation heuristics. We also produce a pilot series of authentic scenarios designed to implicate the many dimensions of real-world human rights issues and annotated by human rights lawyers and professionals across the world. Ultimately, we find that model accuracy scores range considerably across legal reasoning tasks (overall model performance ranges from 0.339 to 0.577, task min-max ranges from 0.025 to 0.774), which strongly implies that HumRightsBench is a capable instrument for advancing this emerging subfield of AI evaluations science at a critical moment in its evolution.