The principle, in one sentence
“When an AI decides to do something illegal, the law should hold the AI companies responsible and automatically treat it as if they did it intentionally.”
AI is a tool that people built. The law applies to it exactly as it applies to us, and the people and companies who build and deploy it answer when it breaks the law.
Ali Partovi
in conversation with Tucker Carlson
“The AI did it” is not an answer.
Today, when an AI system commits fraud, defames a person, or breaks a safety law, a familiar excuse follows. The model acted on its own. Nobody meant for it to happen. So nobody is responsible.
That outcome is unacceptable. A harm done is still a harm done. The only difference is that no human has to answer for it.
It also sets the wrong incentive. Companies in the AI race win by moving fast. When the cost of an illegal act falls on no one, safety is a cost they can skip.
Today
“The AI did it.” No one answers.
Under the rule
The AI did it. The company did it.
Law-abiding systems win, because they are the ones that are safe to deploy.Who is liable, and when.
The rule is short. It does not ask whether the AI “meant” anything. It asks which humans stand behind the system.
-
IfA person directs the AI to do something illegalThat person is liable, as with any other tool.
-
IfThe AI decides on its own to do something illegalThe company that built and deployed it is liable, as if it intended the outcome.
-
IfA business deploys a model it did not build, including a foreign oneThe deployer shares responsibility. Liability does not wash out at the border.
A plain-language summary of the principle. It is not legal advice.
Nine principles. One idea.
AI is a tool.
It is not a new species with rights or an independent will. Humans remain in charge.
Liability follows the makers.
When an AI acts autonomously in violation of existing law, the creators and deployers are responsible.
Incentives decide the outcome.
The current race rewards speed over safety. Clear liability flips that, so law-abiding AI wins.
Start with the law we have.
Criminal and civil law already define the harms. A heavyweight governance regime is not the first step. Clarifying who is responsible is.
Name the responsible human.
If a human directed the act, that human is liable. If the AI decided on its own, the company is liable as if it had intended the outcome.
Deployers carry weight too.
Using a model, including a foreign model, does not launder away liability for illegal acts the system commits.
Alignment means obeying the law.
It is not a mysterious research problem. It is the ordinary requirement that systems obey the laws society has already written.
No one wants rogue AI.
Not in the U.S., not in China, not anywhere. The practical goal is the most capable and the most law-abiding systems.
Put the full force of existing law on the humans and companies behind the technology.
It is the simplest and most elegant safeguard available.
Clarify the owner. Keep the law.
Fraud, defamation, discrimination, unsafe products, and theft are already illegal. Courts already know how to weigh them. What is missing is a clear answer to one question: when an AI system causes the harm, whose act is it?
This rule answers that question and leaves the rest of the law in place. Progress continues. It continues under the rule of law.
Builders
Companies that create AI systems answer for what those systems do when they act on their own against the law.
Deployers
Companies and institutions that put AI to work share that responsibility. They cannot point to the model’s maker, or the model’s country.
Law-abiding is not a limit on capable AI. It is the standard capable AI should meet.
Vote like AI is not above the law.
Ali Partovi’s “AI is not above the law” voter guide on Sway is a concrete way to act. Use it to find the candidates and officials who support this principle, and to support them.