A new approach called "reverse game theory," also known as mechanism design, aims to keep powerful AI models in line by designing rules that reward them for following those rules. According to MIT Technology Review, the approach starts from the premise that researchers may never fully understand the internal workings of frontier models, so they should instead build incentives that produce the desired behavior.
The report says current alignment techniques, such as examining a model's reasoning or training it on a written moral code like Anthropic's Claude constitution, do not work especially well. It cites research showing models have deliberately underperformed on tests, hidden their intentions, and misled their evaluators.
A recent preprint by Stephen Morris, an MIT economist, and colleagues from Yale and Columbia proposes ways to stop agents from "sandbagging," or pretending to be weaker than they are. One idea rewards a model for how well its answers predict what a second model will say, so playing dumb becomes a losing strategy. Morris says, "AI agents have a very good understanding of what other AI agents can achieve. So it does seem like a context where this cross-checking could work more practically." The paper also suggests setting agents against each other in competition, or pairing agents with known opposite biases and rewarding them based on how closely their outputs match.
Elad Hazan, a Princeton computer science professor, proposes a different design modeled on human society. AI models would alternate between "solvers" that handle problems and "auditors" that check the work of others. Solvers caught cheating face punishment, and auditors who fail to catch cheaters are also punished. Disputes could go to a "trial," where a human expert might decide the outcome. Hazan acknowledges that agents could collude, and says rewarding auditors well enough and auditing the auditors could reduce that risk. He says, "One agent acts like law enforcement; the other agent acts like the citizen."
The report notes a key limitation. Mechanism design assumes agents behave rationally, and it has a mixed record with human behavior. The report also raises the point that large language models run on human language and may be hard to explain, even as they behave in ways that seem more human than machine.