It's just the first study. Much more work is needed to confirm these results and also evaluate consistency and reliability. It's not clear what aspects of model training contribute the most (RLHF?)

This, alongside other results, is consistent with LLMs having picked up something deep about reasonableness. If true, that would allow tool building for legal and agent applications, with important AI safety implications. but...

Findings: humans judge harshly people who didn't take the socially common level of precautions, even though textbook says that shouldn't matter.

Textbook does say they should care about costs, but humans don't care much.

LLM repeat this exact pattern in RCTs. (but not all)

To test whether LLMs may have picked up the deeper structure of reasonableness judgment (not just parroting), I offer a new methodology: Silicon RCTs (S-RCTs).

Randomize a relevant fact, run session-isolated “participants,” & compare LLM deltas to human deltas.

{A note for CS friends: a large share of law isn’t if-then rules; it’s open-textured standards like “reasonable,” “ordinary meaning,” and “undue burden”
But our instruments are thin. Expert intuitions (judges) are criticized as elite, out-of-touch, and crypto-political, juries are noisy&biased}

Two hard problems meet here:

(a) courts must infer what ordinary people will reasonably think (e.g., would a teenager read the Pepsi jet as an offer?);

(b) we want AI agents to follow open-textured norms (“keep a reasonable distance”)

Reasonableness quietly structures daily life: following distance in traffic, how loud a house party can be, what “up to 50% longer” implies, or whether a Pepsi ad offering a jet is a joke.

These judgments are fast, intuitive, and hard to explain. In other words: they are "system-1"

Read (papers.ssrn.com/sol3/papers...., arxiv.org/abs/2508.02766)

or listen (auto generated)

I ended up using Walt. I attach the semester plan. I was underwhelmed but it was serviceable