Ordinary, Reasonable Chatbots: Do AI Models Track Human Legal Judgments?
Nirav Patel, Emily Wenger, Christopher Buccafusco
cs.CY · cs.AI
Quick verdict
26 LLMs broadly track 500 humans' legal-reasonableness judgments but are more homogeneous and skew pro-government/corporate; rigorous, bias effects small.
Summary
The study asks whether LLM chatbots reproduce human judgments of legal 'reasonableness' — the vague standard pervading tort, contract, criminal procedure, and other doctrines. The authors surveyed 500 Prolific participants (randomized to 'reasonable' or 'ideal' framing) on 25 numerical reasonableness questions (13 adopted from Tobia 2018), then prompted 26 LLMs from six families (Anthropic, DeepSeek, Google, Meta, OpenAI, xAI) with the reasonable-portion prompt, sampling each model–question pair 20 times at temperature 1. Applying 1.5×IQR trimming to both groups, they compare distributions per question using Brown–Forsythe, Kolmogorov–Smirnov, and Brunner–Munzel tests with Bonferroni correction. Model distributions are statistically distinguishable from humans on most questions (16/25 significant under all three tests), yet usually remain within the range of human responses. Models are markedly more homogeneous (e.g., near-uniform 24-hour landlord-notice answers) and skew toward government- and corporate-favorable answers (e.g., 90 vs 10 days' immigrant detention; 12 vs 6 months' non-compete). A 25-dimensional cosine-similarity analysis suggests model profiles align more with white, male, older, and more-educated respondents (Cohen's d 0.10–0.25), though question-level human demographic differences appear in only 3/25 questions and the authors label these findings suggestive. Limitations include low-context numerical prompts, pooled model distributions, small demographic effect sizes, and no data or code release.
Review complete
Assessment at a glance
0
risk flags identified
● high 0
● medium 0
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LLM responses across 26 models are statistically distinguishable from human distributions on most of 25 reasonableness questions, yet generally remain concentrated within the empirical range of human responses.
- LLMs are substantially more homogeneous than humans and sometimes collapse a vague legal standard into a de facto rule (e.g., near-universal 24-hour landlord-notice answers).
- Compared with humans, LLM answers tilt toward government- and corporate-favorable positions (e.g., 90 vs 10 days of immigrant detention; 12 vs 6 months of non-compete).
- Cross-question cosine similarity indicates LLM profiles align more closely with white, male, older, and more-educated respondents (Cohen's d 0.10–0.25), a preliminary finding the authors flag for confirmation.
Strengths
Triangulated nonparametric design with magnitude-aware reporting
Brown–Forsythe (dispersion), Kolmogorov–Smirnov (shape), and Brunner–Munzel (direction) per question with Bonferroni correction; the authors interpret effect magnitudes, not just significance (only 4/19 significant BM results show P(M>H)≥0.80 or ≤0.20; only 4/23 KS results reach D≥0.50).
Stimulus set anchored in prior empirical legal work and multi-domain
13 of 25 questions reuse Tobia (2018), spanning torts, contracts, immigration, employment, criminal procedure, and family law, enabling direct comparison with the prior human-subjects reasonableness literature.
Distribution-level model elicitation at scale
Symmetric, transparent preprocessing
Calibrated hedging of the most contentious findings
Weaknesses
Unit and wording inconsistencies in headline medians
Abstract demographic claim outruns its evidence base
Pooled 'model' distribution conflates variance sources
Homogeneity pipeline has acknowledged weak joints
No data or code availability statement
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.06769v2, Duke University; AAAI copyright notice) tests whether LLM chatbots reproduce human judgments under the legal 'reasonableness' standard. Contributions: (1) a 25-question numerical reasonableness survey (13 adopted from Tobia 2018) spanning torts, contracts, immigration, employment, criminal procedure, and family law; (2) comparison of 500 Prolific humans against 26 LLMs from six families; (3) claims of LLM homogeneity, pro-government/pro-corporate lean, and demographic alignment (white/male/older/educated). The abstract hedges the latter two as 'suggestive' — to the authors' credit — but the demographic headline still outruns its effect sizes (see below).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 0 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do LLMs make the same legal judgments as humans?
On 25 numerical 'reasonableness' questions answered by 500 human participants and 26 chatbots, model answer distributions were statistically distinguishable from humans on most questions (16 of 25 significant under all three statistical tests) but generally stayed within the range of human answers. The paper concludes chatbots broadly track human views of what is reasonable, while being more uniform in their answers.
Are ChatGPT-style models biased toward the government or corporations?
In this study, models gave answers more favorable to government (e.g., 90 versus 10 days for detaining an immigrant before deportation; 12 versus 6 weeks before trial) and to companies (12 versus 6 months of non-compete; 5% versus 15% of profits spent on safety). The authors describe these tendencies as suggestive rather than conclusive, and their survey questions were not designed specifically to measure such bias.
What does it mean that LLMs are 'homogeneous' in legal judgments?
Human answers vary widely across people, while the 26 models' answers clustered tightly and often converged on the same number (e.g., nearly all models said a landlord must give 24 hours' notice before entering). The authors argue that models tend to treat vague legal 'standards' — which require judgment — as if they were strict 'rules' with one correct answer, which matters for legal applications.
Did the study find demographic bias in AI legal answers?
A cross-question similarity analysis suggested models' answer patterns are closer to those of white, male, older, and more-educated respondents, but effect sizes were small (Cohen's d 0.10–0.25) and human answers actually differed by demographic group on only 3 of 25 questions. The authors themselves label this finding preliminary and call for larger, more diverse samples before drawing conclusions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 (violin plots, cosine-similarity heatmaps, Monte Carlo distributions) are rendered images; visual claims about distribution overlap and heatmap gradients could not be independently inspected from the extracted text.
- The appendix question-level regression results are narrated in prose only; no coefficient table appears in the extraction, so per-coefficient values could not be checked.
- The supplied metadata abstract is cut off mid-sentence ('variable st'), but the complete abstract appears at the top of the paper body, so no content was lost.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.