Ordinary, Reasonable Chatbots: Do AI Models Track Human Legal Judgments?

AI summary

Overview

Research area: AI safety and ethics, with a focus on computational law and the use of large language models as simulators of human judgment ("silicon sampling" / "silicon jurors").

Technical level: Intermediate. The conceptual framing is accessible, but the paper relies on nonparametric statistics (Brown–Forsythe, Kolmogorov–Smirnov, Brunner–Munzel), cosine-similarity embeddings, and Monte Carlo resampling.

Scope: A comparison of twenty-six large language models against 500 human survey respondents on twenty-five legally relevant questions about "reasonableness."

What This Paper Is About

Courts constantly decide whether a person or company acted "reasonably," and many scholars worry that such judgments are vague and may vary across demographic groups. As chatbots become plausible assistants for legal questions, the authors ask whether AI models produce judgments that actually track how humans answer the same reasonableness questions. The paper compares model responses and human responses across twenty-five scenarios drawn from torts, contracts, immigration, employment, criminal procedure, and family law.

Key Contributions

1. A direct human-versus-model comparison. The authors posed twenty-five legally relevant reasonableness questions to 500 human participants recruited through Prolific and to twenty-six LLMs from Meta, Google, Anthropic, OpenAI, DeepSeek, and xAI, using the same scenario set for both.
2. An expanded question set built on prior work. The study uses all thirteen questions from Tobia's survey (Tobia 2018), plus twelve new questions added by the authors to broaden the legal categories tested.
3. A multi-test statistical framework for distributional comparison. Each question is analyzed with three complementary tests—Brown–Forsythe for dispersion, Kolmogorov–Smirnov for overall distributional divergence, and Brunner–Munzel for directional stochastic dominance—with Bonferroni correction for family-wise error across the twenty-five questions.
4. Two cross-question analyses of systematic bias. The paper introduces a Median Absolute Deviation (MAD) table to quantify model homogeneity and a cosine-similarity vector approach (in a 25-dimensional space, one dimension per question) to measure how closely each model family's judgment profile aligns with human demographic subgroups.

Main Findings

• Models broadly track humans, but the distributions are statistically distinguishable. Across the twenty-five questions, twenty showed significant Brown–Forsythe results, twenty-three showed significant Kolmogorov–Smirnov results, and nineteen showed significant Brunner–Munzel results after correction, with sixteen questions significant under all three tests. The authors characterize the model distributions as often remaining within the empirical range of human responses.
• Some questions showed no detectable difference. Q5 (construction delay) and Q14 (inappropriate comments) were non-significant across all three tests, with reasonable-human and model distributions overlapping closely in shape and central tendency.
• Model outputs are more homogeneous than human outputs. The Brown–Forsythe results frequently correspond to very large effects, consistent with the tightly clustered model distributions. Examples from the MAD table include Q19 (permit price: model MAD 0.0 vs human MAD 75.0), Q4 (additional costs: model 0.0 vs human 500.0), Q7 (profits on safety: model 2.0 vs human 5.0), and Q15 (traffic stop duration: model 0.0 vs human 3.0).
• In some questions the pattern reverses. Model MAD exceeded human MAD for Q17 (immigrant detention: 60.0 vs 8.0), Q21 (odometer discrepancy: 2000.0 vs 100.0), Q10 (attorney's fees: 50.0 vs 40.0), and Q3 (reflection on a risky proposition: 24.0 vs 21.0). Four questions had identical model and human MAD: Q9 (4.0), Q16 (1500.0), Q22 (0.0), and Q24 (0.0).
• Occasional extreme directional shifts. Q17 (immigrant detention) showed P(model > human) = 0.94, meaning model responses exceeded human responses in roughly 94 percent of pairwise comparisons. Q10 (attorneys' fees) showed a Kolmogorov–Smirnov D of 0.75, indicating a threshold where cumulative shares differ by 75 percentage points. Q7 (profits on safety) had a Brown–Forsythe W of 453.42.
• Most significant differences are moderate, not extreme. Among the nineteen questions with significant Brunner–Munzel results, only four showed extreme stochastic dominance (P(model > human) ≥ 0.80 or ≤ 0.20). Among the twenty-three significant Kolmogorov–Smirnov results, only four showed very large separations (D ≥ 0.50).
• Some models treat a variable standard as an invariant rule. In Q19 (permit price), both the 25th and 75th percentiles equaled fifty, giving an IQR of zero. The IQR also collapsed for Q11 (landlord notice), Q15 (traffic stop duration), Q22 (data breaches), Q23 (child alone), and Q24 (dating partners). In Q11, responses were almost perfectly homogeneous, with most models returning 24.
• More pro-government responses than humans. Models allowed longer traffic stops (human median 12 vs model median 15 minutes), more time before trial (median 6 vs 12 weeks), longer warrantless GPS tracking (median 1 vs 2 days), and longer immigrant detention before deportation (median 10 vs 90 days).
• More pro-corporation responses than humans. Models supported less spending on safety (human median 15 vs model median 5 percent), judged corporations less likely to pollute again (median 50 vs 30 percent), and allowed longer non-compete restrictions (median 6 vs 12 months). The authors note product returns and construction delays were more ambiguous on this dimension.
• Alignment skews toward specific demographic subgroups. Model responses aligned more closely with respondents who are white, male, older, and more educated. For the gender split, the reported means were 0.7020 (SD 0.0729) and 0.6945 (SD 0.0750), with t = 7.17 (p < 0.001) and Cohen's d = 0.10. The truncated text does not report the remaining demographic comparison values.
• Robustness to outliers. The authors state that because their primary tests are nonparametric or rank-based, sparse tail outliers in otherwise concentrated model distributions do not materially affect inference.

Methodology in Plain English

The authors designed a survey of twenty-five scenarios where the law would ask what is "reasonable"—things like how many days to accept a business contract, how many hours of notice a landlord gives before entering, or how many months a non-compete can last. Each question asked for a numeric quantity.

Human side. With Duke University IRB approval (Protocol 2026-0109), 500 Prolific participants completed the survey plus a demographic questionnaire. They were paid an average of $12.57 per hour, prorated to completion time, and finished in an average of 6 minutes and 41 seconds. Roughly half were randomly assigned to give the "reasonable" quantity and half the "ideal" quantity, so the authors could later ask whether models lean toward or away from the ideal.

Model side. Twenty-six models were queried through their APIs on the reasonable version only, via a stateless pipeline. The system prompt appended the instruction not to provide reasoning, just the reasonable quantity without units. Each model–question pair was sampled twenty times to estimate a response distribution. Temperature and top-p were held at their default value of 1 so differences reflect model behavior rather than parameter tuning. Outputs were post-processed to extract unitless numbers; failures were recorded as Not-A-Number (NaN).

Cleaning. Both human and model data were trimmed with the standard 1.5×IQR rule at the question level rather than dropping whole respondents. Where the IQR was zero, a 3σ fallback rule was applied. The trimming is described as conservative with respect to the hypothesis tests and as reducing heteroskedasticity without artificially increasing alignment.

Analysis. For each of the twenty-five questions, three tests compared model and human distributions: Brown–Forsythe (do the spreads differ?), Kolmogorov–Smirnov (do the whole distributions differ?), and Brunner–Munzel (does one group tend to give larger answers?). Bonferroni correction controlled family-wise error. MAD summarized how tightly each group clustered. Separately, the authors turned each group's twenty-five answers into a vector, min–max scaled per question jointly across all vectors, and computed cosine similarity between every model family and every human demographic subgroup (age split at 18–44 versus 45+, ethnicity as White versus Non-White, education as Associates or less versus Bachelors or more, gender as Man versus Woman). A Monte Carlo procedure over the scaled response space generated distributions of cosine similarities, compared with two-sample t-tests on their means.

Why This Matters

The paper sits at the intersection of two live debates: whether AI can substitute for or supplement human legal decision-making, and whether AI models faithfully represent human populations. The authors note that Judge Kevin Newsom of the 11th Circuit Court of Appeals wrote that he had moved from thinking LLM use in legal text interpretation was "positively ludicrous" to being "a pretty firm 'maybe'." If models systematically favor governments and corporations, or mirror only the views of white, male, older, and more educated respondents, that has direct consequences for whose notion of "reasonable" gets encoded in legal guidance.

Real-world applications:

• Silicon juries and judicial decision support. If models are used to predict or simulate jury reasoning, systematic skew toward particular demographic groups would distort outcomes.
• Consumer-facing legal chatbots. Laypeople asking an AI whether a landlord, employer, or police action was lawful would receive answers skewed in a particular direction.
• Institutional policy and compliance. Insurers, employers, and agencies that use AI to interpret standards of care, negligence, or contract terms need to know how the model's baseline differs from public judgment.
• Computational social science. Researchers using "silicon sampling" as a cheap substitute for human survey panels need evidence about whether model samples resemble human samples.

Industry relevance. Model developers, legal-tech vendors, and enterprises deploying AI in HR, compliance, or contract review all depend on the assumption that model outputs resemble ordinary human judgment on vague standards. This paper shows both that the assumption often holds approximately and that it fails in patterned, directional ways—especially around government power and corporate liability.

Future Directions

• Confirming or rejecting the initial findings. The authors state explicitly that more systematic research is needed to confirm or reject these results; replication across larger and more diverse samples is the obvious next step.
• Understanding the pro-government and pro-corporate lean. The paper flags several questions (product returns, construction delays) as ambiguous on this dimension and calls for further study of when and why these leanings appear.
• Using the ideal/reasonable randomization. The design included human "ideal" respondents specifically to test whether models tend toward or away from the ideal quantity, but the truncated content does not report that comparison; following up on it is a natural extension.
• Whether homogeneity can be reduced. Because the paper attributes model convergence to shared pretraining data, shared architectures, and overlapping alignment procedures, an open question is whether prompting strategies, ensembling different model families, or different training objectives could restore the heterogeneity seen in human judgments.

Target Audience

Legal scholars and law-and-economics researchers studying reasonableness and standards-versus-rules; AI safety and alignment researchers interested in value representation and output homogeneity; computational social scientists using LLMs as survey substitutes; judges, regulators, and policymakers evaluating AI in legal decision-making; and legal-tech practitioners who need to know where model judgment diverges from public judgment.

Authors’ abstract

As people increasingly rely on artificial intelligence (AI) for guidance in their own lives, scholars, lawyers, and even judges have begun to consider the role of AI in legal decision-making. As "silicon sampling" -- the use of generative AI models in social science research -- is now impacting academia, "silicon jurors" could make an appearance in courtrooms. This study joins an emerging line of research on generative AI models' ability to simulate human legal judgments. In particular, we study how large language model (LLM)-powered chatbots respond to series of questions about legal reasonableness. When the law needs to judge the appropriateness of a behavior, it most often asks whether the behavior was "reasonable." Yet despite the ubiquity of reasonableness judgments, they are the site of constant vexation for lawyers, judges, and lay people. Reasonableness seems inherently vague and unpredictable, since it relies on variable context and implicit conceptual schemas. Moreover, many scholars caution that reasonableness judgments may vary along demographic lines. We compare the answers of human participants to those of twenty-six LLMs across twenty-five different legally relevant reasonableness judgments. Overall, our findings suggest that chatbot responses generally track those of human participants. Nonetheless, we find some suggestive -- and potentially concerning -- results. Compared to humans, LLMs generate more homogeneous responses and occasionally treat a variable standard as an invariant rule. And, compared to humans, LLMs tend to generate answers that are more favorable to the government and to corporations. Finally, our results indicate that LLMs' responses tend to align more closely with those of respondents who are white, male, older, and more educated. More systematic research is needed to confirm or reject these initial findings.