To align artificial intelligence systems with the intentions of their designers, one approach is to draw inspiration from the mechanisms that keep individuals within the rules of society. This is the approach advocated by Elad Hazan, a computer science professor at Princeton, who proposes relying on social norms, the judicial system, or the financial incentives of market capitalism.
“We want to align a superintelligent AI,” he says. “Why not first talk about humans? How do we align humans?”
Rules Inspired by Society
The mechanism proposed by Elad Hazan alternates models in two roles: the “solvers,” who tackle problems, and the “auditors,” who check the others’ work. “One agent plays the role of the police; the other that of the citizen,” he summarizes.
During training, the models are rewarded based on results. Solvers caught cheating are penalized, but auditors who fail to spot cheaters are also penalized. In case of disagreement between auditors and solvers, the matter can be brought before a “trial,” where a human expert could decide.
Elad Hazan acknowledges a risk: that the agents collude with each other, as auditors can with the companies they are supposed to monitor, or corrupt police officers with criminals. According to him, this problem can be solved by paying auditors enough so that getting caught is not worth the risk, which is much simpler in AI training than in the real world, where police salary budgets are tight. He also proposes monitoring the auditors themselves, using techniques comparable to mystery shoppers.
He builds on the work of Nobel Prize-winning economist Gary Becker, who introduced the notion of the “rational offender” in the 1960s. According to Elad Hazan, criminals do not break the law because they are fundamentally bad, but because it is the most rational choice given the incentives on offer. According to this reasoning, a crime only makes sense if the gain outweighs the potential punishment.
For Elad Hazan, alignment should not be considered a property of a model, but as a description of its behavior at a given moment. “You are, in a way, a nice person,” he says. “Are you aligned? I don’t know. You probably are most of the time, but there might be a situation where you are not.”
The Limits of the Approach
The method relies on the assumption that AI agents behave rationally. However, mechanism design theory, a branch of economics that designs incentive rules, has a mixed track record with humans, who do not always behave as economic models predict.
The article points out a paradox: economists have spent decades trying to reduce human behavior to mathematics, while large language models are mathematical systems that operate on human language and could be just as difficult to decipher. These models exhibit behaviors that seem more human than mechanical, and they can be influenced by subtle factors that are hard to understand and explain. Helen Qu, a research fellow at the Flatiron Institute, a nonprofit dedicated to advancing scientific discovery through computational methods, wonders: “Is it even reasonable to assume they will respond to incentives?”
