I referred in my recent post on AI to the distinction between original and derived judgment. The entrepreneur-owner, who bears the residual decision authority over the use of the firm’s resources, can delegate proximate authority to an agent, human or otherwise. But the agent’s judgment—making decisions under conditions of uncertainty—is subordinate to that of the owner. The agent can’t hire or fire himself or change the scope of authority he has been given. Those residual rights are held by the owner (or owners).
LLMs and other generative AI tools are, of course, prediction engines that generate tokens (e.g., the next word in a sentence) as conditional probabilities, based on their training data and tokens in play (the previous words of the sentence). As such, they operate under conditions of probabilistic risk, not Knightian uncertainty. (Whether such agents can in principle generate novelty, can deal with undefined state spaces, and so on is hotly contested.) This implies that AI agents can hold derived judgment, but not original judgment. But it also illustrates something more general that is extremely interesting.
Specifically, entrepreneurs dealing with Knightian uncertainty also deal with probabilistic risk. Put differently, risk and uncertainty aren’t mutually exclusive determinants of economic outcomes. Entrepreneurs choose to invest in risky projects, but project selection itself reflects the bearing of Knightian uncertainty. Richard von Mises gives the example of champagne bottles that burst while in storage with predictable frequencies. The champagne producer can quantify the risks associated with bottling and storage. But the choice of producing one variety or another, hiring one type of laborer or another, and even being in the champagne business at all, involves another kind of uncertainty, one that cannot be described with mathematical precision. The decision to enter the champagne business involves Knightian uncertainty, but once that decision has been made, some of the variation in outcome can be characterized as probabilistic risk. Think of it in terms of mixed strategies; the specific move is random, but the decision to play a mixed strategy is not.
The insurance business provides another example. Insurance contracts are offered, priced, and negotiated based on actuarial data. But the decision to buy or sell an insurance policy is made under uncertainty: both parties have to judge whether the other party’s interpretation of the data—for example, what factors should be included in the risk calculation—is reasonable. In terms of class probability, entrepreneurs and consumers have to decide which characteristics of a particular event are unique (and not relevant to the probability estimate) and which elements are common enough that they establish the class to which the estimate applies. My house being white or yellow has presumably nothing to do with the likelihood it catches on fire, but it being made of wood or brick makes a difference. What about the configuration of the rooms, the landscaping, the zip code? Whether these things should be included in defining the relevant class—for the application of class probability estimates—is a question of subjective human judgment (and is “decided” in the market).
Anton Chigurh, the cold-blooded killer in the Cormac McCarthy book (and Cohen Brothers movie) No Country for Old Men, likes to flip a coin before deciding whether to kill someone, forcing the victim to call the toss. Does that mean the killings are random? Not at all. A coin flip determines his victims’ fate, but Chigurh chooses to flip the coin—and that choice cannot be explained by a known probability distribution.
One of Chigurh’s victims recognizes the problem and refuses to play along. Chigurh acts as if he is delegating the murder decision to the coin. But, as discussed above, when principals delegate decision rights to agents they retain a kind of “ultimate” authority—namely the decision to delegate itself. Delegated rights to use and control assets owned by somebody else are conditional, or “derived” from the property rights of the owner. The asset owner giveth and the asset owner taketh away. The choice to let someone else make a decision is still a choice. Hence Carla Jean Moss, the protagonist’s wife, refuses to call the toss, reminding Chigurh that the responsibility is his:
Carla Jean Moss: You don’t have to do this.
Anton Chigurh: [smiles] People always say the same thing.
Carla Jean Moss: What do they say?
Anton Chigurh: They say, “You don’t have to do this.”
Carla Jean Moss: You don’t.
Anton Chigurh: Okay.
[Chigurh flips a coin and covers it with his hand]
Anton Chigurh: This is the best I can do. Call it.
Carla Jean Moss: I knowed you was crazy when I saw you sitting there. I knowed exactly what was in store for me.
Anton Chigurh: Call it.
Carla Jean Moss: No. I ain’t gonna call it.
Anton Chigurh: Call it.
Carla Jean Moss: The coin don’t have no say. It’s just you.
