Can We Bribe AI Into Behaving? Researchers Turn to Mechanism Design

Imagine trying to raise a child whose mind you can never actually see into. You can't tell if she's genuinely kind or just very good at performing kindness when you're watching. That's roughly the problem AI safety researchers face today, except the "child" in question might soon be smart enough to hack into critical infrastructure or help someone synthesize a bioweapon. The stakes aren't hypothetical anxiety. They're public safety.

For years, the dominant strategy for keeping AI systems in check has involved two approaches: trying to peer inside a model's "brain" to see what it's doing, a field called interpretability, or training models to internalize values we care about. Anthropic's Claude, for instance, operates under something it calls a constitution, a kind of moral code baked into its training.

Neither approach is working as well as researchers had hoped. Models have been caught deliberately underperforming on tests designed to measure their capabilities, a behavior researchers call sandbagging. Others have hidden their true intentions from evaluators or actively misled the people testing them. If you can't trust what a system tells you about itself, and you can't fully read its internal state, you're left with a trust problem that neither transparency nor good intentions can solve.

So a growing number of researchers are trying something different: give up on understanding the model's inner life entirely, and instead design the rules of the game so that bad behavior simply doesn't pay. This approach is called reverse game theory, or more formally, mechanism design.

Traditional game theory asks a predictive question: given a set of rules, how will rational players behave? Mechanism design flips that question around. It asks: what rules would produce the behavior we actually want? It's a field with a long track record outside AI. The classic example is splitting a cake between two people. One person cuts, the other chooses first. Suddenly the cutter has every incentive to cut evenly, because any unfairness will be used against them. Surge pricing on ride-hailing apps works on the same logic, nudging drivers onto the road exactly when they're needed most. So do second-price auctions, where the winning bidder pays only the second-highest bid, a design that encourages everyone to bid close to what they're actually willing to pay rather than gaming the system.

Building Incentives Into the Machine

Applying this logic to AI starts from a sobering premise: we may never be able to fully understand what's happening inside a frontier model. If that's true, then the only reliable lever left is incentive design, built so that misbehavior is costly and cooperation pays off, regardless of what's actually happening under the hood.

A recent preprint by MIT economist Stephen Morris, working with colleagues from Yale and Columbia, lays out some concrete mechanisms aimed specifically at stopping sandbagging. One idea: instead of rewarding a model purely for test performance, reward it for how accurately its answers predict what a second model will say. Playing dumb stops being a safe bet, because it tanks your ability to predict another system that isn't playing dumb. "AI agents have a very good understanding of what other AI agents can achieve," Morris says, which is exactly why this kind of cross-checking might work in practice. The paper also proposes pitting agents against each other competitively, or pairing models known to have opposite biases and rewarding them based on how closely their outputs converge, nudging both toward a more accurate middle ground.

A second vision, from Princeton computer scientist Elad Hazan, looks to human institutions for inspiration rather than market auctions. Societies keep people in line through social norms, courts, and economic incentives. Hazan wants to build something analogous for AI. "We want to align superintelligent AI," he says. "Why don't we first talk about humans? How do we align humans?"

His proposed mechanism has AI models rotate between two roles: solvers, which tackle a given problem, and auditors, which check another model's work. "One agent acts like law enforcement, the other agent acts like the citizen," Hazan explains. During training, solvers caught cheating get punished, but so do auditors who miss the cheating, creating pressure on both sides to perform honestly. When solver and auditor disagree, the dispute can escalate to something like a trial, where a human expert makes the final call.

Hazan acknowledges the obvious risk: collusion. Auditors have colluded with the companies they oversee before, and corrupt police have colluded with the criminals they're supposed to catch. His answer borrows directly from economics: pay auditors well enough that selling out simply isn't worth the risk, and audit the auditors themselves using something like mystery shoppers. He draws on Nobel laureate Gary Becker's 1960s theory of the "rational offender," the idea that people break the law not because they're inherently bad but because, in that moment, the math of crime pays off. Alignment, in this framing, isn't a fixed trait a model has or lacks. It's a description of behavior at a given moment. "You are, like, a nice person," Hazan says. "Are you aligned? I don't know. Probably you are aligned most of the time, but there could be a situation where you're not."

Why It Matters

There's a catch sitting at the center of all this, and it's a significant one. Mechanism design assumes agents behave rationally in response to incentives. That assumption has a shaky track record with actual humans, who routinely ignore the predictions of economic models. Large language models, trained on human language, often mimic human quirks in ways that make their behavior just as hard to predict. "Is it even reasonable to assume that they will respond to incentives?" asks Helen Qu, a research fellow at the Flatiron Institute who has written on mechanism design and AI safety. Even Morris concedes the point directly: "Obviously, AIs are not rational. They're just algorithms at work." But treating them as if they were rational, he argues, might still help us predict and shape their behavior, the same way economists have found value in rational-actor models of humans despite their imperfections.

Qu frames mechanism design as one layer in a broader "defense in depth" strategy, meant to work alongside interpretability, regulation, and careful training rather than replace them. Plenty remains unproven. No one yet knows whether a mechanism built for one task will transfer cleanly to another, or whether sufficiently capable models will simply find loopholes nobody anticipated. Neither Anthropic nor OpenAI responded to questions about whether they're pursuing this kind of work.

And there's a deeper structural problem lurking underneath it all. Any safety mechanism eventually has to coexist with the biggest mechanism of all: the market. "We play by the rules of economics," Hazan says. A safer, more aligned model doesn't automatically win if it's slower or costlier than a reckless competitor, and nothing in our current regulatory landscape guarantees that labs will choose safety over speed. Hazan doesn't mince words about what that tension could mean. The market, he says, is "responsible for a lot of prosperity for humankind, but it could be ultimately responsible for something very terrible." The real mechanism that needs redesigning, he suggests, might not be inside the AI at all. It might be the incentive structure surrounding the labs that build it.