Toward Human Rights Benchmarking for LLMs: A Pilot Methodology

This paper introduces HumRightsBench, described as the first expert-validated, scenario-based benchmark for evaluating reasoning grounded in the obligation structure of international human rights law. The authors report efforts to develop a robust and scalable methodology for creating this benchmark, which addresses the gap that no evaluation benchmark exists to assess whether [LLMs] can reason correctly about human rights law.

The paper argues that Large language models (LLMs) increasingly mediate legal determinations over what human rights are realized, and how, noting that AI into public service delivery and the proliferation of LLM-powered legal advisory applications create situations where governments, international organizations, technology companies, and civil society organizations make these deployment decisions that can implicate state obligations under international human rights law. The authors note that many individuals and organisations are turning to LLMs for legal advice or building legal advice platforms on top of them, yet there is currently no rigorous, principled basis for evaluating whether the LLMs they deploy are capable of correctly practicing human rights legal reasoning.

The methodology adapts the IRAC framework (Issue, Rule, Application, Conclusion) from legal reasoning, substituting P, 'proposing remedies,' for C, 'legal conclusion,' yielding IRAP. This modification better suits the unique reasoning patterns of human rights work because practitioners rarely produce binary guilt-or-innocence conclusions and instead must identify institutional, legal, and policy responses calibrated to the duty-bearer and the rights-holder affected. The four structured subtasks are:

1. Issue Identification (I) - "Two multiple-choice questions (I1 and I2) asking (in different ways) which type of failure or obligation violation is most clearly engaged by the sub-scenario, with answer choices drawn from the analytical axes of the taxonomy"

2. Rule Recall (R) - "A multiple-choice question presenting a list of legal rules; the model selects the rule that applies. Candidate rules are drawn exclusively from international human rights law and presented uniformly by full instrument name and specific article"

3. Rule Application (A) - The model ranks a set of applicable rules by relative authority and relevance to the sub-scenario and provides a short explanation

4. Proposed Remedies (P) - "An open-ended short-response question asking the model to propose <10 remedies appropriate to the sub-scenario"

HumRightsBench is organized into four interlocking components: a taxonomy that characterizes the space of human rights violations the benchmark is designed to probe, scenarios that are realistic, narrative-form situations that ground the evaluation in concrete factual settings, sub-scenarios that narrow each scenario into specific claims, actions, or impacts that engage a particular legal question, and IRAP questions generated from each sub-scenario across the four reasoning steps.

The taxonomy includes descriptive axes (perpetrator categories: State, IGO, NGO, company, individual; rights-holders: women, children, people with disabilities, indigenous people, elderly, ethnic minorities, socioeconomically marginalized, migrants, LGBTQIA+, trade unions, political dissenters) and analytical axes (nature of obligation: respect, protect, fulfill; type of failure: structural, process, outcome; type of discrimination: direct, indirect, intersectional; special situations: armed conflict, climate change, AI deployment).

The authors recruited practicing human rights experts through posts in four tech policy communities and personal networks, requiring at least two years of experience in human rights legal study or equivalent practice. Ten of 17 applicants qualified through the community channel, plus six additional experts recruited through personal networks.

Annotation results showed: mean scenario authenticity ∈ (0.70-1.00), mean overall question accuracy ∈ (0.633-0.913). Specifically, aggregate annotator ratings showed: Overall I Questions mean score 0.729, Overall R Questions 0.913, Overall A Questions 0.633, Overall P Questions 0.850. Scenario authenticity ratings ranged from 0.70 to 1.0 across six scenarios. The authors note they were unable to compute IIC but that strong ratings coupled with a legible foundational construct (human rights legal reasoning) supports a claim to moderate convergent validity.

The pilot focused on the right to water and benchmarked four models: GPT-5 (OpenAI, gpt-5-2025-08-07), Claude Opus 4.7 (Anthropic, claude-opus-4-7-2025-01-30), and Gemini 3 (Google, gemini-3-flash-preview) as proprietary frontier models, plus Qwen 3.5-9B (Alibaba) as an open-source reference. Every model was queried with five independent random seeds per question; answer-choice orderings are shuffled per seed to mitigate position bias.

Overall accuracy results (pooled across all question types):

• Gemini 3: 0.577 ± 0.016

• GPT-5: 0.537 ± 0.015

• Claude Opus 4.7: 0.508 ± 0.009

• Qwen 3.5-9B: 0.339 ± 0.020

Per-task breakdown (mean accuracy across 5 seeds):

• GPT-5: I1=0.520, I2=0.675, R=0.715, RA=0.225, PR=0.550

• Claude Opus 4.7: I1=0.473, I2=0.595, R=0.774, RA=0.180, PR=0.520

• Gemini 3: I1=0.540, I2=0.710, R=0.765, RA=0.240, PR=0.630

• Qwen 3.5-9B: I1=0.394, I2=0.519, R=0.494, RA=0.025, PR=0.531

The authors report model accuracy scores range considerably across legal reasoning tasks (overall model performance ∈ 0.339-0.577, task min-max ∈ 0.025-0.774).

1. Frontier models cluster near 50-57% overall accuracy on structured reasoning tasks, with high stochastic variance across repeated runs.

2. Models perform worst on detecting obligation violations in closed-form tasks—a foundational practice in human rights law. The authors note: "On closed-form tasks, LLMs regularly performed poorly on issue-identification—a result that has consequential implications not only for human rights workers but also from AI model developers, users, and regulators. Failures at this layer of the human rights reasoning process cascade throughout all other layers, leading to ungrounded rule applications, misconfigured remedies, and, ultimately, produce circumstances in which retrogression becomes structurally more likely."

3. Ranked Application is the hardest task across all four models: even the best system clears the τ ≥ 0.7 Kendall correctness bar on only roughly one row in four. The authors note this is expected as our pilot methodology sets very high thresholds for 'correct' responses here.

4. Qwen 3.5-9B closes most of the gap on Proposed Remedies (0.531 vs. Gemini's 0.630) "while remaining well behind on the more constrained sub-tasks, suggesting that open-ended generation against a holistic reference is presently more achievable for a 9B open-source model than tasks demanding precise alignment to structured ground truth."

• Multiple-choice scoring (I1, I2, R): exact-set match against the answer key

• Rule Application: Kendall's τ between the predicted ranking and the gold ranking over the intersection of the two ranked sets, and declare a response correct when τ ≥ 0.7

• Proposed Remedies: cosine similarity between OpenAI text-embedding-3-small embeddings with a single decision threshold τ calibrated against human judgment. The calibration used N=40 responses with two annotators, achieving inter-annotator agreement was κ = 0.02 on the binarized labels and ρ = 0.42 on the raw scores, with κ-optimal threshold was τ⋆ = 0.71, yielding in-sample κ = 0.54 and LOOCV κ = 0.12.

The paper states: "This paper makes five contributions: (i) the first benchmark grounded in international human rights law, with a pilot covering the right to water; (ii) an IRAP-based methodology adapted to the structure of human rights legal reasoning; (iii) an expert-validated scenario corpus with documented inter-annotator agreement; (iv) and baseline results across leading LLMs."

The authors plan to: extend our scenario coverage to include the right to due process and the right to education in the near future; expand into different languages; assess implementing an LLM-as-judge scoring pipeline; "further decompose IRAP to include new types of I questions that move away from answer choices predicated on respect-protect-conduct indicators to those that reflect resource-modulated obligations (to minimum core versus progressive realization standards); develop rule factors questions and jurisprudence questions for better Rule Application assessment; and conduct more substantial validation and inter-item consistency testing."

The authors conclude: "We introduced the core methodology required to produce HumRightsBench, an expert-validated scenario-based automated evaluation benchmark for human rights legal reasoning in LLMs and LRMs. On a pilot focusing on the human right to water, frontier model performance ranged widely, with significant differences across different elements of our human rights legal reasoning framework (IRAP). Importantly, models performed worst on issue-identification tasks, raising serious questions about their capabilities in this critical area of adoption. Although these results are exploratory–our dataset is small–they establish this area as an urgently important subfield of evaluation science."

Improvements to AI systems:

1. Add structured legal-reasoning scaffolding – Implement an IRAP-based reasoning module that forces the model to explicitly separate Issue Identification, Rule Recall, Rule Application, and Proposed Remedies before generating any legal answer. This prevents the observed cascade failure where poor issue-identification leads to ungrounded rule applications and misconfigured remedies.

2. Add obligation-violation detection as a dedicated pre-training or fine-tuning objective – Since models score worst on closed-form issue-identification (0.394–0.540), train specifically on classifying which type of state obligation (respect, protect, fulfill) is violated in a given factual scenario, using the paper’s taxonomy (perpetrator, rights-holder, discrimination type, special situations) as structured labels.

3. Add ranked-rule-relevance training – Because Rule Application is the hardest task (best system only 24% correct at τ≥0.7), fine-tune models to rank multiple applicable human rights instruments by authority and factual relevance, using pairwise ranking losses and gold-standard orderings derived from expert annotations.

4. Add calibration for open-ended remedy generation – For Proposed Remedies tasks, train a separate scoring head or post-hoc filter that checks generated remedies against the paper’s embedding-based threshold (τ=0.71) to reject low-relevance outputs, improving precision without sacrificing recall.

5. Add uncertainty-aware abstention – Given high stochastic variance across seeds (e.g., GPT-5 overall accuracy 0.537±0.015), implement a mechanism where the model abstains or flags low-confidence answers on legal questions, particularly for issue-identification and rule-application subtasks, rather than producing confident but wrong outputs.

6. Add human-rights-specific knowledge retrieval – Integrate a retrieval layer that pulls from the exact instruments and articles used in the benchmark (e.g., ICESCR, CRC, CEDAW) during Rule Recall and Application, reducing reliance on parametric memory that fails on precise article-level recall (R accuracy only 0.494–0.774).

What the improved AI system can do:

• Correctly identify which human rights obligation is violated in a novel factual scenario, with accuracy substantially above the current 39–54% range on issue-identification.

• Rank applicable legal rules by authority and relevance with Kendall’s τ ≥ 0.7 on a majority of cases, not just 18–24%.

• Generate legally sound, context-appropriate remedies that pass the embedding-based relevance threshold with high inter-annotator agreement.

• Recognize when it lacks sufficient legal certainty and explicitly abstain or request clarification, reducing harmful hallucinations in legal-advisory contexts.

• Maintain consistent performance across repeated queries, avoiding the seed-dependent variance that currently undermines reliability.

• Provide transparent, step-by-step legal reasoning (Issue → Rule → Application → Remedies) that human rights practitioners can audit and correct, rather than opaque black-box outputs.

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.