TLDR: A new study demonstrates that large language models (LLMs) can accurately predict how ordinary people judge reasonableness in legal contexts. Using ‘silicon randomized controlled trials’ with AI models assigned diverse personas, the research found that LLMs replicate human intuitions regarding negligence, contract interpretation, and consent, often prioritizing social norms over economic efficiency, even when this contradicts traditional legal theories. This suggests AI can serve as a valuable, cost-effective tool to understand public opinion in legal decision-making, though with noted limitations in response magnitude and potential biases.
In our daily lives, we constantly make judgments about what is “reasonable”—whether it’s how fast to drive, what an advertisement truly means, or if someone genuinely consented to a gesture. For the legal system, predicting these everyday judgments is incredibly difficult. Judges are often accused of relying on their own elite perspectives, juries can be influenced by biases, and traditional surveys are too expensive for most.
This challenge arises because human judgments of reasonableness are based on intuitive mental frameworks that are hard to explain, following complex statistical patterns. A new research paper, “The Silicon Reasonable Person: Can AI Predict How Ordinary People Judge Reasonableness?” by Yonathan A. Arbel, explores whether large language models (LLMs)—the powerful AI systems trained on vast amounts of text—can learn to identify these subtle patterns that guide human reasonableness judgments. You can read the full paper here: The Silicon Reasonable Person.
AI as a Tool for Understanding Human Reasonableness
The paper suggests that LLMs, which are essentially advanced pattern detectors, are surprisingly well-suited for this task. While it might seem counterintuitive for AI to grasp deeply intuitive human decisions, the difficulty in defining general rules for reasonableness stems from these judgments following complex patterns influenced by many situational and contextual variables. LLMs excel at picking up such patterns, much like they learn to understand poetry or coding.
The research emphasizes that LLMs are not meant to replace human judgment entirely. Instead, they could offer a reliable improvement over current methods, providing a “dictionary of reasonableness” that judges, lawmakers, and litigants could consult. For example, a judge in a case like Leonard v. Pepsico (where a teenager tried to redeem Pepsi points for a fighter jet) could check their intuitions against an AI model simulating teenage perspectives to get a more realistic understanding of common perceptions.
How the Research Was Conducted
To investigate this, the study used a novel approach called “silicon randomized controlled trials” (S-RCTs). This method is similar to the “gold standard” in experimental research, randomized controlled trials, but it uses language models as independent “subjects.” This ensures that each AI model’s response is fresh and not influenced by previous interactions, mimicking how new human participants would be recruited for each experimental condition. Researchers could then present thousands of variations of scenarios to the models to see how specific variables affected their judgments.
A key aspect of this methodology was assigning detailed “personas” to the AI models. These personas included demographic information (age, gender, race, income, education) and personality descriptions, allowing the AI to role-play as a diverse range of individuals matching US demographic patterns. This helped the models provide responses that aligned with broad patterns of reasoning among ordinary people, shifting from a default “expert” mode to a more realistic “layperson” mode.
Key Findings from the Studies
The research involved three different studies, encompassing hundreds of human respondents and over 10,000 simulated judgments by AI models:
The first study replicated research on how people judge negligence. It found that both humans and AI models were significantly influenced by social norms (whether a precaution was common practice) but gave little weight to economic factors (the cost of taking the precaution). This contradicts traditional economic theories, which often prioritize cost-benefit analysis. While AI models showed the same pattern as humans, the magnitude of their response to social norms was slightly less pronounced than human subjects.
The second study explored perceptions of consent, fairness, and enforceability in contract scenarios, specifically regarding hidden fees. It compared AI models’ judgments with those of both lay individuals and legal professionals. The study found that certain AI models closely mirrored lay individuals’ tendency to be “contract formalists,” meaning they perceived hidden charges as more binding than legal professionals might, believing strongly in the power of contracts. This again showed AI aligning with common, non-expert understanding.
The third study delved into subtle aspects of consent under deception. It presented scenarios where a person was misled about either an “essential” feature of a purchase (e.g., buying a bicycle but getting a camera) or a “material” feature (e.g., a purchase not being eligible for reward points as promised). Paradoxically, human subjects perceived more consent given to “material” lies than to “essential” lies, even though the deceit mattered more to them in the material lie scenario. Strikingly, some AI models successfully replicated this counterintuitive pattern, suggesting a sophisticated understanding of human intuition regarding consent.
Also Read:
- AI Judging AI: A New Era for Language Model Evaluation
- Assessing LLM Capabilities in Explaining Well-Being Concepts: A Deep Dive into AI’s Readiness
Implications and Future Directions
These findings suggest that “silicon reasonable people” could have immediate practical applications. They could help lawmakers understand how policies might resonate with the public, offer judges a way to calibrate their intuitions against broader societal patterns, and provide resource-constrained litigants with an affordable alternative to expensive mock juries. Compliance departments in companies could also use them to test whether their procedures or advertisements would be considered reasonable by the average consumer.
However, the research also highlights important limitations. AI models sometimes differ from humans in the magnitude of their responses, meaning they might identify the right factors but not weigh them exactly as humans do. There’s also the issue of “value drift,” where models trained at a specific time might not reflect evolving social norms. Furthermore, performance varies significantly across different AI models, emphasizing the need for careful selection and validation.
Ultimately, the paper advocates for using silicon reasonable people as an “adjunct” tool, not a replacement for human judgment. Transparency, addressing potential biases (especially the “majoritarian bent” that might overlook minority perspectives), and continuous validation against real-world data are crucial for their responsible integration into legal practice. This new frontier in AI and law promises to make the concept of the “reasonable person” more transparent, accessible, and democratically accountable.
