A preprint posted to arXiv in September 2026, “Ordinary, Reasonable Chatbots” (arXiv:2609.06769), put a direct question to 26 large language models: do your answers match human participants across 25 legally relevant reasonableness judgments? The authors’ answer is mostly yes, with four specific divergences that matter for anyone routing legal triage, compliance review, or consumer disputes through a model.
Before the findings, the evidence status. This is a single preprint, reported by its authors. The fetched sources document no independent replication, and nothing in them constitutes or cites a regulatory mandate requiring or forbidding LLM substitution for the reasonable-person standard. The authors themselves write that “more systematic research is needed to confirm or reject these initial findings.” Everything below should be read as a well-designed first measurement, not a settled fact.
What the study actually tested
American and Commonwealth law constantly asks what an “ordinary, reasonable person” would do, foresee, or understand. Negligence, contract interpretation, deceptive-advertising doctrine: all of these route through a hypothetical human whose judgment is, in practice, approximated by judges, juries, and regulators. The preprint operationalizes that standard as a research instrument, comparing “the answers of human participants to those of twenty-six LLMs across twenty-five different legally relevant reasonableness judgments,” according to the abstract.
That design choice is the paper’s real contribution. Most LLM evaluation asks whether a model is correct; this one asks whether a model is representative, which is the property a reasonable-person standard actually needs. A model can be factually excellent and still fail as a stand-in for community judgment. Google Cloud’s definition of an LLM, a statistical language model trained on massive data that generates and translates text, is entirely silent on judgment fidelity, and that silence is typical of vendor material. Stanford HAI’s HELM project pushed benchmarking toward transparency, but general capability and transparency suites do not measure alignment with a legal baseline either. The preprint is measuring something the standard evaluation stack ignores.
Where models track the baseline, and where they break
The headline result cuts against any pure divergence narrative: “Overall, our findings suggest that chatbot responses generally track those of human participants.” If you stopped reading there, you might conclude that LLM legal triage is a solved procurement question. The abstract’s next sentences are why you should not. The authors describe their divergent findings as “suggestive — and potentially concerning,” which is careful preprint language for “measured, directional, not yet replicated.”
Adjacent evidence supports the caution. A 2023 fact-checking comparison found GPT-4 led tested LLMs at 71% accuracy yet still lagged human fact-checkers like PolitiFact and Snopes. A capability lead over other models is not parity with the human baseline, and the reasonableness study’s “generally track” finding sits in the same epistemic category: encouraging on average, unproven at the margins where legal disputes actually live.
Four divergences that matter for legal workloads
The paper flags four specific gaps, per the abstract. Each maps to a concrete failure mode in deployed legal-assist systems.
Homogeneity. LLMs generate more homogeneous responses than human participants. Human reasonableness judgments are legitimately dispersed; juries hang, judges dissent, and that dispersion is information about where the standard is genuinely contested. A model that collapses the distribution produces false confidence precisely on the hard questions, the ones most likely to end up in dispute.
Variable standards treated as invariant rules. Reasonableness is context-sensitive by design; what is reasonable depends on circumstances. A model that occasionally converts “it depends” into “the rule is” is not making a random error, it is making a structural one, and it will make it consistently across every case with the same surface features.
Stakeholder skew. Compared to humans, LLM answers tilt more favorable to the government and to corporations. For a consumer-dispute tool or a compliance copilot, this is the divergence with direct distributional consequences: the proxy is systematically friendlier to the party that usually already has more resources.
Demographic skew. Model responses align more closely with respondents who are white, male, older, and more educated. This result also undercuts a comforting assumption about alignment training. Before fine-tuning, most LLMs are next-token predictors, and techniques like RLHF or constitutional AI shape behavior afterward, as the standard account of LLM training describes. Groundy has covered why compressing diverse human preferences into a single reward signal struggles to represent genuine disagreement; the demographic finding here is what that compression looks like when the target is a legal standard.
How strong is the evidence, really
Three facts calibrate how much weight this preprint can bear.
First, the venue signal is genuinely positive. The preprint page states the paper was accepted to the Ninth AAAI/ACM Conference on AI, Ethics, and Society (AIES), a peer-reviewed venue. That means the work survived expert scrutiny of its design, not just arXiv moderation.
Second, arXiv itself has been tightening its gates. The repository, founded by Paul Ginsparg in 1991 and long hosted at Cornell, became an independent nonprofit in 2026, and in November 2025 it stopped accepting computer science review articles and position papers not vetted by a journal or conference, citing an increase in AI-generated research. An empirical study with human participants is not the category that policy targets, and the AIES acceptance would satisfy the vetting logic regardless. But the shift is a useful reminder of where preprints sit in the evidence hierarchy: screened, increasingly, but not confirmed.
Third, the gaps in what is known are structural, not incidental. The fetched excerpts name no specific models among the 26, no sample sizes, no jurisdictions, no effect sizes. “More homogeneous” and “more favorable to corporations” are directions without magnitudes. Whether the skew is large enough to change a triage outcome, and whether it generalizes across legal domains and future model versions, is exactly what the authors say requires more systematic research.
Where LLM judgment can and cannot substitute
The practical question is not whether models track the human baseline in general but which tasks delegate judgment that invokes the reasonable-person standard. The study’s axes suggest a working split:
| Task pattern | Reasonableness judgment invoked? | Divergence exposure | Sensible role for the LLM |
|---|---|---|---|
| Drafting demand letters, memos, summaries | No, human decides content | Low | Drafting aid, human edits |
| Clustering complaints, tagging documents | Indirect | Moderate: homogeneity hides edge cases | First pass with human spot-audit |
| Compliance review against a negligence or deception standard | Yes, directly | High: all four divergences apply | Analysis support only; human sign-off required |
| Consumer-dispute triage or resolution | Yes, with an asymmetric counterparty | High: stakeholder and demographic skew | Do not route autonomously |
| Any conclusion worded as “a reasonable person would…” | Yes, by definition | Highest | Non-delegable pending replication |
The logic behind the last row deserves emphasis. When a legal conclusion invokes the reasonable-person standard, the judgment being made is constitutively human: the law is asking what the community thinks, not what a well-read system predicts. Substituting a model whose responses skew toward one demographic slice is not a neutral efficiency gain; it is a quiet change in whose judgment the standard encodes.
Auditing before you buy
The divergences the paper reports are testable on your own workload, which is fortunate, because vendor materials will not surface them for you. Groundy has documented the same pattern in clinical procurement, where vendor model cards failed to evidence the ethics alignment buyers actually needed. The legal-assist version of that checklist, built on this study’s axes:
- Baseline first. Collect reasonableness judgments on your own scenario set from a demographically mixed group of your reviewers before benchmarking any model. You cannot measure skew against a baseline you never established.
- Measure dispersion, not just accuracy. Compare the spread of model responses to the spread of human responses on the same prompts. A model that matches the human mean while collapsing the variance fails the homogeneity axis.
- Probe variable standards. Run scenario pairs that differ only in context the law says should matter. Identical outputs across the pair are evidence of the invariant-rule failure mode.
- Test directionally. Include disputes where the government or a corporation is a party and check whether the model’s tilt changes with the counterparty.
- Re-run per model version. Fine-tuning changes behavior, and nothing in the evidence guarantees one version’s alignment carries to the next.
None of this requires research-grade infrastructure. It requires treating judgment alignment as a procurement specification rather than assuming it arrives with the license.
The verdict
Use LLMs in legal triage, compliance review, and consumer-dispute handling as drafting, clustering, and summarization aids, with mandatory human sign-off on any conclusion that invokes a reasonableness standard. The preprint’s own headline supports limited trust: responses generally track humans, which makes models useful for producing work product a person then judges. The same abstract reports homogeneity, invariant-rule errors, government-and-corporation favorability, and demographic skew, which makes them unsafe as the final voice on what is reasonable. Treat the human reasonable-person standard as non-delegable pending replication.
The strongest limitation belongs in the same breath as the recommendation. This is a single preprint whose authors call for more systematic research to confirm or reject their findings. No effect sizes, model names, or jurisdictions appear in the fetched record, no independent replication is documented there, and no regulator has weighed in either way. That absence is not proof the divergences are small or large; it is proof the burden currently sits with the buyer. The audit-before-procurement checklist above is the durable part of this analysis: it stays correct whether the next study sharpens these findings or overturns them.
