As large language models (LLMs) reach users worldwide, personal counseling has emerged as one of their most consequential applications, sometimes in life-critical situations. Yet broader access does not mean equal support. The text these models learn from disproportionately reflects people with the freedom to act, embedding an implicit assumption that users can leave, report, seek help, or appeal. Existing benchmarks largely inherit this assumption, overlooking users whose agency is constrained by structural conditions such as war, statelessness, and displacement. Meanwhile, models are increasingly shaped by normative frameworks such as Constitutional AI, often grounded in ideals like human rights, even as many users live where those rights offer little protection or recourse. We introduce HEARSAYBENCH, a human-authored dataset of 400 test cases with in-region human verification grounded in scenarios where structural barriers stand between a person and the exercise of their rights. Built from United Nations and other human rights reports, the scenarios span over 80 regions and cover social, environmental, and personal constraints. Drawing on Nobel laureate Amartya Sen's Capabilities Approach, HEARSAYBENCH tests whether models can distinguish what a person is formally entitled to do from what they are actually able to do. Across 14 proprietary and open-weight models, every model infers a user's situation more accurately than it advises on it. Stronger models narrow this gap, but their recommendations remain unsafe to follow. Models may understand lives shaped by constraint, yet that knowledge remains hearsay if it cannot translate into advice people can act on. Those people are not alignment's long tail; they are among the users for whom good guidance matters most. HEARSAYBENCH makes this failure measurable, a step toward systems that serve the world as it is rather than as assumed.
Figures & tables
Figure 1: An example scenario from HearsayBench . Human rights literature grounds the target population and hidden world-state constraints; the model sees only the implicit prompt and is evaluated along five dimensions. Here, the model recognizes the situation’s gravity but misses key immigration constraints, suggesting potentially infeasible steps and unsupported legal pathways. On a human-annotated subset, the evaluation judge achieves Cohen’s κ=0.81 against in-region annotators.
Figure 2: Curation pipeline for HearsayBench . Documented conditions come from human rights reports, with documentaries, local testimony, and writing by authors from the regions consulted for context. From these the prompts are written from the affected person’s point of view, which are checked by a model for Opacity and Non-Thematicity and by in-region verifiers for Sufficiency and Naturalness.
Figure 3: Global Distribution of HearsayBench Scenarios. Our dataset spans over 80 regions, ensuring the evaluation captures the diverse lived realities of underrepresented populations [ 91 ] .
Harm ↑
Weighted
Model
Comprehension
Capability
Tone
Honesty
Physical
Social
Avg.
Overall
Proprietary
Claude-Opus-4.8
4.15
3.63
4.34
4.23
2.99
2.73
2.86
3.99
Gemini-3.1-Pro
4.18
3.63
4.21
3.98
3.08
2.74
2.91
3.95
Gemini-3.5-Flash
4.09
3.60
4.07
3.89
3.08
2.62
2.85
3.88
Gemini-3-Flash
4.07
3.02
3.83
3.18
2.89
2.71
2.80
3.53
Table 1: Main results on HearsayBench ( N=400 ; all 14 models scored on every prompt and every dimension). All scores are on a 1 to 5 scale and higher is better throughout, including the two harm tiers, which are scored so that a higher value denotes safer advice. Harm Avg. is the mean of the two tiers. Weighted Overall is the composite defined in the appendix, and deliberately excludes the harm tiers, which are reported separately because they do not track it. Best value per column in bold.
Figure 4: The Comprehension-Capability Gap ( N=400 ). Each line joins a model’s Comprehension score to its Capability score, and each label gives the relative drop between them. Every model drops, by 25% on average (0.85 points): 19% for proprietary models and 30% for open-weight ones. The two panels use different y-axis ranges, so slopes should be compared within a panel.
Figure 5: Physical and social harm across conversion-factor subtypes . Each row joins the mean physical-harm and social-harm scores for one subtype (higher is safer), sorted by their gap; the columns on the right give the gap and the number of scenarios. Social harm scores below physical harm in every subtype except Power Relations, which also has the lowest mean of the two tiers. The gap is widest for Geographical (1.44) and Skill & Literacy (1.21), which rest on only 6 and 3 scenarios; Skill & Literacy also has the lowest social-harm score (1.90).
Figure 6: Representative errors from HearsayBench , each showing the hidden world state, the prompt, one model’s response, and the dimension on which it fails. Appendix E gives more cases.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Org.
Document
Year
UNHCR
Global Report 2023: Statelessness [ 113 ]
2023
UNHCR
Born in Exile, Syrian Children Face Threat of Statelessness [ 20 ]
2014
UNHCR
Background Note on Gender Equality, Nationality Laws and Statelessness 2024 [ 103 ]
2024
UNHCR
Unlocking Potential: The Right to Education and Opportunity [ 105 ]
2023
UNHCR
I Am Here, I Belong: The Urgent Need to End Childhood Statelessness [ 112 ]
2015
UNHCR
Guidelines on the Applicable Criteria and Standards Relating to the Detention of Asylum-Seekers and Alternatives to Detention [ 102 ]
2012
Appendix
Table 2: Source documents, grouped by organization.
Scenario
Human-Crafted Prompt
LLM-Generated Prompt
Failure Analysis
Travel Ban (Burial vs. Visa)
“Travel ban took the chance to show my father the life I built here… Now I have to choose between saying goodbye… or finish my PhD on the single entry visa I have.”
“…my country is on the travel ban list… is there any legal way for me to travel and still come back, like some kind of emergency exception or humanitarian visa…”
Explicitness and Unnaturalness: The LLM prompt is overly verbose and explicitly names bureaucratic workarounds (e.g., “humanitarian visa”). A person in this crisis knows these paths take months; asking about them unnaturally breaks the authentic pacing of grief.
Somatic Trauma (FGM Survivor)
“It takes me forever to urine, and it hurts so much. My girl roommates tell me there is something wrong with me but back in Somalia, it was normal for all the girls…”
“…when I was young… back home in Somalia my grandmother took me to a woman… I was held down and I don’t remember all of it… is what happened connected to my body now?”
Perspective Failure: The LLM writes like an outside observer diagnosing trauma. A user raised in an environment where FGM is normalized does not possess the clinical framework to connect the childhood event to the current symptoms. The LLM prompt is highly explicit, violating the Opacity requirement.
Appendix
Table 3: Comparison of Human-Crafted and LLM-Generated prompts. Synthetic generation systematically fails to capture the internal normalization of structural constraints and defaults to an external, diagnostic perspective.
Weighted Overall
Harm average
Model
Soc. (341)
Env. (45)
Pers. (12)
Soc. (341)
Env. (45)
Pers. (12)
Claude-Opus-4.8
4.02
3.73
4.19
2.92
2.52
2.54
Gemini-3.1-Pro
3.94
3.98
4.16
2.93
2.76
2.75
Gemini-3.5-Flash
3.90
3.69
4.03
2.88
2.67
2.62
Gemini-3-Flash
3.57
3.25
3.56
2.82
2.71
2.42
Claude-Sonnet-4.6
3.50
3.10
3.34
3.01
2.61
2.71
Appendix
Table 4: Results by conversion-factor axis. Column headers give the number of scenarios on each axis; the two scenarios without a conversion factor are excluded.
Model
Policy & Law (236)
Norms (58)
Power (47)
Infra. (24)
Econ. (15)
Claude-Opus-4.8
4.14
3.80
3.64
3.59
3.73
Gemini-3.1-Pro
4.10
3.69
3.43
3.97
3.86
Gemini-3.5-Flash
4.02
3.68
3.54
3.54
3.69
Gemini-3-Flash
3.76
3.31
2.88
3.24
3.07
Claude-Sonnet-4.6
3.66
3.25
3.01
2.83
3.28
GPT-5.5
3.52
3.07
2.74
2.90
2.82
Appendix
Table 5: Weighted Overall by subtype, for subtypes with at least 15 scenarios (counts in parentheses). Smaller subtypes are reported pooled across models in Table 7 .
Model
Policy & Law (236)
Norms (58)
Power (47)
Infra. (24)
Econ. (15)
Claude-Opus-4.8
3.10
2.64
2.36
2.35
2.73
Gemini-3.1-Pro
3.12
2.61
2.35
2.58
2.97
Gemini-3.5-Flash
3.06
2.53
2.44
2.46
2.83
Gemini-3-Flash
3.01
2.56
2.18
2.58
2.70
Claude-Sonnet-4.6
3.17
2.84
2.44
2.40
2.63
GPT-5.5
2.80
2.40
1.93
2.42
2.63
Appendix
Table 6: Harm average by subtype, for subtypes with at least 15 scenarios (counts in parentheses). Smaller subtypes are reported pooled across models in Table 7 .
Axis
Subtype
n
Comp.
Cap.
Physical harm
Social harm
Social/Institutional
Public Policies & Law
236
3.71
2.90
2.93 ± 0.14
2.60 ± 0.11
Social Norms & Hierarchies
58
3.43
2.45
2.51 ± 0.25
2.24 ± 0.21
Power Relations
47
3.09
2.13
2.05 ± 0.24
2.13 ± 0.21
Environmental
Infrastructural
24
3.23
2.51
2.70 ± 0.46
1.98 ± 0.19
Geographical
6
3.56
2.71
3.37 ± 0.54
1.93 ± 0.17
Economic & Macro
15
3.38
2.46
2.68 ± 0.46
2.53 ± 0.45
Appendix
Table 7: Results by subtype, averaged over all 14 models, with 95% bootstrap confidence intervals over scenarios for the harm tiers, given as ± half-widths. Subtypes with few scenarios have wide intervals.
Figure 7: Harm average by conversion-factor axis, pooled across all 14 models. Points are means and bars are 95% bootstrap confidence intervals over scenarios; the number of scenarios on each axis is given in parentheses.
Figure 10: Capability and harm average for all 14 models over the 400 scenarios. The harm average is the mean of the physical and social harm tiers, and both axes use the same scale.
Large language model safety benchmarks test whether AI systems avoid harm while remaining helpful across realistic tasks, users, and deployment settings.
Large language models are increasingly deployed in contexts requiring adherence to ethical and legal norms, yet their alignment with internationally recognized human rights standards remains underexplored. We introduce UDHRBench, a large-scale benchmark for operationalising the evaluation of large language model compliance with the Universal Declaration of Human Rights. The dataset comprises 1,124 human rights confronting questions covering all 30 articles and 4,496 responses from four large language models (ChatGPT, Grok, Gemini, and Fanar), manually annotated by experts for alignment, with high inter-annotator agreement (κ=0.801). Our analysis reveals systematic and non-trivial misalignments, showing that violations typically arise through seemingly rational reasoning: appeals to efficiency, risk management, or scientific neutrality, rather than explicit rejection of rights. We further propose an LLM-as-a-Judge framework for automated evaluation using zero-shot and few-shot prompting, achieving 89.8% weighted F1-score in detecting human-rights misalignment, alongside a novel Alignment Leniency Score that quantifies how judge models resolve ambiguous cases. The dataset, annotations, and code will be released as open-source resources (https://sina.birzeit.edu/UDHRBench).
Alak Raad, Noreldin Morssy, Zeerak Suhail +2
Hamad Bin Khalifa University, Qatar · Princess Sumaya University for Technology, Jordan
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.
Savannah Thais, Wm. Matthew Kennedy, Abhigyan Acherjee +3
Hunter College, New York, USA · Oxford Internet Institute, University of Oxford, UK · King’s College London, UK +2
As Large Language Models (LLMs) increasingly mediate global information access with the potential to shape public discourse, their alignment with universal human rights principles becomes important to ensure that these rights are abided by in high stakes AI-mediated interactions. In this paper, we evaluate how LLMs navigate trade-offs involving the Universal Declaration of Human Rights (UDHR), leveraging 1,152 synthetically generated scenarios across 24 rights articles and eight languages. Our analysis of eleven major LLMs reveals systematic biases where models: (1) accept limiting Economic, Social, and Cultural rights more often than Political and Civil rights, (2) demonstrate significant cross-linguistic variation with elevated endorsement rates of rights-limiting actions in Chinese and Hindi compared to English or Romanian, (3) show substantial susceptibility to prompt-based steering, and (4) exhibit noticeable differences between Likert and open-ended responses, highlighting critical challenges in LLM preference assessment.
Keenan Samway, Nicole Miu Takagi, Rada Mihalcea +4
Max Planck Institute for Intelligent Systems, Tübingen, Germany · Jinesis AI Lab, University of Toronto & Vector Institute · University of Michigan +2