The Generative Reasonable Person
This Article introduces the "generative reasonable person," an empirical tool using large language models to simulate thousands of lay judgments that reveal how ordinary people actually assess reasonableness in legal contexts, thereby providing a scalable, data-driven baseline to challenge judicial intuitions and refine legal standards in negligence, consent, and contract law.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Problem: The Judge's "Mind-Reading" Gap
Imagine a judge sitting alone in a quiet, fancy room (chambers). A case comes before them: Did a teenager really think a TV commercial promising a fighter jet for collecting soda points was a real offer?
To decide, the judge has to answer a question that seems simple but is actually impossible: "What would an ordinary, reasonable person think?"
For 200 years, judges have had to guess the answer. They rely on their own intuition, which is often shaped by their elite education and life experience. It's like a chef trying to guess what a hungry toddler wants to eat, but the chef has never actually spoken to a toddler. They might guess wrong, thinking the toddler wants a steak when they actually just want a cookie.
The law calls this the "Reasonable Person" standard. But for a long time, no one knew what the "Reasonable Person" actually thought because asking real people was too expensive, too slow, and too hard to do on a large scale.
The New Tool: The "Silicon Reasonable Person"
This paper introduces a new tool: Large Language Models (AI) acting as a "Generative Reasonable Person."
Think of AI not as a robot that knows the law, but as a massive digital mirror of human conversation. It has read billions of words written by regular people—comments, forums, stories, and arguments. Because of this, it has absorbed the "vibe" of how ordinary people actually think, not just how lawyers say they think.
The author, Professor Yonathan Arbel, calls this method "Silicon Randomized Controlled Trials" (s-RCTs).
- The Analogy: Imagine you want to know how people react to a new flavor of ice cream. Instead of hiring 10,000 people to taste it (which costs a fortune), you create 10,000 digital clones of different people. You give each clone a different version of the ice cream and ask, "Do you like this?" Because the digital clones don't talk to each other, their answers are independent and honest.
What Did the AI Discover? (The Surprises)
The author tested these AI "clones" on three legal puzzles where the law's rules often clash with what regular people actually feel. The AI didn't just recite legal textbooks; it replicated the messy, intuitive ways humans actually judge things.
1. The Negligence Puzzle (The "What Everyone Does" Test)
- The Law: Says you should do the math. If a safety step is cheap, you should take it, even if no one else does. If it's expensive, you don't have to, even if everyone else does.
- The AI (and Real People): Said, "Who cares about the math? If 90% of people are doing it, you must do it too."
- The Metaphor: The law is like a strict accountant checking a spreadsheet. The AI and real people are like a crowd at a party. If everyone is wearing red hats, you feel weird wearing a blue one, even if the blue hat is cheaper. Social conformity beats cost-benefit analysis.
2. The Consent Puzzle (The "Essential Lie" vs. The "Material Lie")
- The Law: If someone lies about something that matters to you (like "this car gets 50 MPG" when it gets 20), your consent is invalid.
- The AI (and Real People): Said something weird. They felt that if someone lied about the essence of the thing (e.g., "This is a bicycle" but it's actually a camera), it ruined the deal. But if they lied about a term that mattered a lot (e.g., "You get points for this" when you don't), people felt they still "consented" to the deal, even though they were angry about the lie.
- The Metaphor: It's like buying a "real diamond" that turns out to be glass (Essential Lie) vs. buying a real diamond that doesn't come with the fancy box you were promised (Material Lie). People feel more betrayed by the glass, but they feel they still "agreed" to the transaction even if the box was a lie. The AI caught this paradox; the law missed it.
3. The Contract Puzzle (The "Fine Print" Trap)
- The Law: Hidden fees are unfair and shouldn't be enforced.
- The AI (and Real People): Said, "Well, it's in the contract, so you have to pay it." Regular people are surprisingly "formalists." They believe that if you signed it, you are bound by it, even if it feels unfair.
- The Metaphor: Real people are like a student who signs a contract without reading it, then says, "I guess I have to pay the fine." They trust the paper more than their own feelings of fairness.
Why This Matters: The "Dictionary" of Reasonableness
The paper argues that we shouldn't let AI make the final decision. Instead, we should use it as a dictionary.
- Before: A judge says, "No reasonable person would think that ad was real." They are just guessing.
- Now: A judge can ask the AI, "What would 10,000 simulated teenagers think?" The AI says, "Actually, 60% of them would think it's real."
- The Result: The judge can now say, "I know the AI says 60% think it's real, but I am choosing to rule otherwise because the law requires a higher standard."
This makes the judge's choice transparent. They can no longer hide behind "common sense" if their "common sense" is actually just their own elite bias.
The Catch (Limitations)
The paper is honest about the flaws:
- The "Majoritarian" Bias: The AI reflects the average person. It might miss the views of minority groups or marginalized communities. It's like a mirror that only shows the majority's face.
- The "Volume" Problem: The AI is great at telling you which way people lean (e.g., "People prefer social norms over math"), but it's not perfect at telling you exactly how much (e.g., "People are 4.2% more likely to agree").
- The "Safety" Filter: AI models are trained to be polite and safe, which sometimes makes them sound more "moral" than real, messy humans.
The Bottom Line
For two centuries, the "Reasonable Person" was a ghost—a figure we talked about but never saw. This paper shows that Generative AI can finally make that ghost visible.
It allows us to peek behind the curtain of judicial intuition. It tells us that ordinary people often care more about what their neighbors do than what a math formula says, and that they often trust the fine print more than their own gut feelings.
By using these "Silicon Reasonable People," the law can stop guessing what the public thinks and start actually listening to them. It turns the "Reasonable Person" from a magical fairy tale into a measurable, data-driven reality.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.