Yonathan Arbel

Black and blue butterflies made out of paper structured as if they are flying away from pink AI letters made out of paper. It is on a white background.

How to Count AIs: Individuation and Liability for #AIAgents

This Article is the first to diagnose the legal problem of identifying #AIs.

Authors: Yonathan A. Arbel, Simon Goldstein, Peter Salib

Read More: spkl.io/63329Avgjv
@arbel.bsky.social @petersalib.bsky.social

The result: emergent, persistent, legally legible AI entities that respond rationally to tort, regulatory, and economic incentives.
Download before the singularity!

papers.ssrn.com/sol3/papers....

Because AIs are resource-constrained (compute above all), the AIs controlling an A-corp’s assets have powerful incentives to delegate authority only to sub-agents that share their goals.
Internally misaligned A-corps are selected out by resource exhaustion and competition.

Our solution: the Algorithmic Corporation (“A-corp”) 🎩

A-corps are legal persons: owned by identifiable humans, able to hold property, contract, and be sued, but governed by AIs via cryptographically secure keys and fine-grained, revocable permissions.

Thick identity: individuating stable, goal-coherent AI entities that law can directly govern when human oversight inevitably falls short.

Thin identity is necessary but insufficient. Principal–agent problems explode with truly autonomous systems.

We distinguish two problems:
Thin identity: tying every action back to a responsible human principal.

That’s the sort of problem legal scholarship started thinking about when they argue over holding users or developers liable.

But they haven’t solved the problem of how to exactly identify agents.

Law needed some principle to attach liability, mens rea, and consequences to a loose group of — you get it — agents.

AI agents present this problem but much more sharply. They have no bodies. They fork, swarm, merge, and vanish in milliseconds.

When someone does something wrong, how do we know who did it?

When only neighbors were involved, that was hard but solvable.

When individuals started cooperating over large distances, or in large groups that shifted members, that became much harder.

In How to Count AIs: Individuation and Liability for AI Agents, with @psalib
and philosopher Simon Goldstein, we apply an old legal dilemma to a new problem and offer a time-tested solution.

Paper (just posted): papers.ssrn.com/sol3/papers.cf
m?abstract_id=6273198

🚨Very soon, billions of AI agents will swarm the economy—copying, splitting, merging at will.

Just as soon, someone will get hurt.

Is law ready for this moment?

We think not quite. But we have a solution.

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)

🚨🚨🚨Can judges really know what “reasonable people” think? Can AI help bridge that gap—and can AI agents themselves behave “reasonably” in the wild?

A new draft, The Silicon Reasonable Person, asks these questions.

tl;dr: early signs point to yes, with careful limits.

While students hit the books, AI professors head to Alabama. Great start to the AI Safety Law Roundtable.

It's a real treat to collaborate with @arbel.bsky.social & @petersalib.bsky.social to encourage more profs to evaluate the promise & peril of AI innovation. Stay tuned for more!

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

Oh wow