Silicone Reasonable People: AI Generated Ordinary Judgments in Legal Contexts
- Date: Dec 9, 2024
- Time: 04:00 PM (Local Time Germany)
- Speaker: Yonathan Arbel (University of Alabama)
- Room: Hybrid: Zoom and basement
Reasonableness, fundamental to tort, criminal, contract, employment, and constitutional law, relies on as-sessing how ordinary people would judge actions in context. Unfortunately, current methods for eliciting these lay judgments are limited, biased, and resource-intensive. To assist with such determinations, this Article introduces "silicone reasonable people": AI simulations of ordinary judgments in legal contexts. This method offers us a way to study common views that is far more rapid, cost-effective, and controllable than traditional methods.
To study whether silicone reasonable people are a valid solution, more than superficial testing is required. Models can easily parrot back persuasive definitions of reasonableness without having internalized the concept. This Article instead evaluates the deeper, latent representation of reasonableness embedded in A1 models. It uses a novel randomized controlled trial methodology, analyzing over
10,000 AI—simulated responses across twelve different language models, to see how humans and models judge reasonableness in context.
The results show statistically significant alignment across experiments between the factors that matter for human schema and AI judgment. While the factors are similar, their assessment levels do not always match, suggesting limits on the models’ abilities to fully capture lay judgments. Nonetheless, the evidence suggests models can in fact learn the deeper structure of reasonableness necessary to supplant human juries.
Understood within their actual capabilities, silicone reasonable people can enhance our understanding of legal decision—making and assist in evaluating reasonableness standards for legal actors. They are not, and cannot be, a total replacement to human judgment, for reasons of legitimacy, reliability, and explaina-bility. But they are a powerful complement to human judgment in legal analysis, and they open new avenues for incorporating broader perspectives into legal reasoning, potentially amplifying underrepresented voices by simulating diverse viewpoints.
To study whether silicone reasonable people are a valid solution, more than superficial testing is required. Models can easily parrot back persuasive definitions of reasonableness without having internalized the concept. This Article instead evaluates the deeper, latent representation of reasonableness embedded in A1 models. It uses a novel randomized controlled trial methodology, analyzing over
10,000 AI—simulated responses across twelve different language models, to see how humans and models judge reasonableness in context.
The results show statistically significant alignment across experiments between the factors that matter for human schema and AI judgment. While the factors are similar, their assessment levels do not always match, suggesting limits on the models’ abilities to fully capture lay judgments. Nonetheless, the evidence suggests models can in fact learn the deeper structure of reasonableness necessary to supplant human juries.
Understood within their actual capabilities, silicone reasonable people can enhance our understanding of legal decision—making and assist in evaluating reasonableness standards for legal actors. They are not, and cannot be, a total replacement to human judgment, for reasons of legitimacy, reliability, and explaina-bility. But they are a powerful complement to human judgment in legal analysis, and they open new avenues for incorporating broader perspectives into legal reasoning, potentially amplifying underrepresented voices by simulating diverse viewpoints.