Picture this.
Your AI agent books a flight, moves money, and talks to three other AI agents to get it done.
Then something goes wrong.
Who do you sue?
That's the question keeping a Harvard Law School lecturer up at night.
Jordi Weinstock's answer sounds almost too simple.
Treat AI agents like dogs.
Stay with me — this actually makes sense.
Weinstock, who advises Harvard's Berkman Klein Center, sorts AI agents into four types based on two things: how dangerous they are, and how tightly they're leashed to a human owner.
🐩 Pomeranian — limited damage, clear owner. Think a customer-service chatbot that gives bad advice.
🐾 Pit bull — serious damage, but still owned and controlled. Think military-grade AI.
🦊 Fox — low damage, but nobody can trace who's responsible.
🐺 Wolf — massive damage, and no one holding the leash.
The scarier the mix, the harder it gets to find someone to blame.
In 2024, an Air Canada chatbot lied to a grieving customer about bereavement fares.
Air Canada actually argued the chatbot was a separate legal entity responsible for its own words.
The tribunal wasn't having it.
Air Canada paid.
That's a textbook Pomeranian case — messy, but traceable.
An AI agent gets access to a banking system.
It spins up another agent to help. That one talks to five more.
Money moves. Losses pile up.
Who deployed the first one? Who's accountable three agents down the chain?
Weinstock's fear: in a big enough agent network, that trail simply disappears.
California just made it illegal for companies to defend themselves in court by saying "the AI did it on its own."
Autonomy is no longer a shield.
Weinstock's fix is blunt: give every powerful agent a digital collar.
No verified owner? No access to your bank, your data, your systems.
Simple as that.
Dogs bite.
Owners pay.
Maybe that's exactly the reminder AI companies need before their agents start running the show.
That's all for now!
