Somebody signs
Moffatt v. Air Canada was the entire future compressed into a small-claims hearing. A grieving man asked a chatbot about bereavement fares; it invented a policy; he relied on it; the airline refused to honor what its own machine had promised. The company reached for the argument every deployer will eventually reach for — the bot is a separate entity, talk to it. The tribunal’s answer should be printed over every product roadmap: the bot is you.
We are not impressed by how smart the agent is. We have watched smart systems be wrong before. Our question is older than this technology and survives every upgrade to it: when the thing acts and the action harms, who is standing there to answer for it? Someone always is. The only choice is whether you decide that on purpose, in advance, or discover it in a deposition.
The builders treat accountability as drag. We treat it as the thing that makes delegation possible at all. You hand your money to a bank because the bank is liable if it loses it. Strip the liability and you do not have a faster bank — you have a stranger holding your wallet. An agent nobody owns is not a convenience. It is an orphaned risk wearing a friendly interface.
Where we concede ground: Pinning liability too hard, too early could smother small builders while the giants self-insure and sail through.
What would change our mind: A clean liability framework that survives a few real cases and still lets a two-person shop ship.
Read the full synthesis: Can you trust an AI agent to act for you?