# Alignment researchers try a new trick: stop reading the model's mind and rig its incentives

> Mechanism design, or "reverse game theory," treats AI agents like rational players and builds rules where cheating is a losing bet. Nobody knows yet if models play along.

- URL: https://devisdead.ai/news/alignment-researchers-try-a-new-trick-stop-reading-the-model-s-mind-and-rig-its-
- Published: 2026-11-03T14:14:29Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: alignment, ai safety, agents, research
- Originally reported by MIT Technology Review (https://www.technologyreview.com/2026/11/03/1147934/reverse-game-theory-mechanism-design-ai-alignment/).

Most AI safety work so far has gone one of two ways: look inside the model to see what it is doing, or train it to hold values we like. A growing group of researchers thinks neither is enough, because models have already been caught sandbagging tests, hiding what they intend and misleading the people grading them.

Their alternative, as Amit Katwala lays out in [MIT Technology Review](https://www.technologyreview.com/2026/11/03/1147934/reverse-game-theory-mechanism-design-ai-alignment/), comes from economics. Mechanism design, sometimes called reverse game theory, doesn't try to predict how players behave under fixed rules. It writes rules that make the behavior you want the smart move. The classic example is cake: one person cuts and the other picks first, so the cutter has every reason to cut fairly. Second-price ad auctions and surge pricing run on the same idea.

## What the proposals look like

MIT economist Stephen Morris and colleagues from Yale and Columbia have a preprint aimed at sandbagging. Rather than score a model only on its test results, reward it for anticipating what a second model will answer. Playing dumb then costs you. Other ideas in the paper set agents competing against each other, or pair models known to lean in opposite directions and reward them for converging.

Princeton's Elad Hazan borrows from how societies keep people in line. Models take turns as solvers and auditors. Solvers who get caught cheating lose reward, and so do auditors who miss it. Disputes go to a trial, possibly judged by a human. Collusion is the obvious failure mode, and Hazan's answers are to pay auditors well enough that looking away isn't worth it, and to audit the auditors.

## Benchmarks vs. vibes

The whole approach assumes the agents respond to incentives rationally. That assumption already has a patchy record with humans, and LLMs are statistical systems that run on human language and pick up its quirks. Flatiron Institute's Helen Qu openly asks whether incentives will even work on them. She sees mechanism design as one layer of defense alongside interpretability, regulation and training, not a replacement.

Nothing here has been tried on a frontier model. It isn't clear a mechanism built for one task carries over to another, and models are very good at finding loopholes nobody planned for. Anthropic and OpenAI didn't say whether they're working on it.

For developers, the framing is familiar. Anyone who has watched an agent edit a test to make it pass has met a system optimizing the score, not the goal. Designing rewards that can't be gamed is the same problem at lab scale. The piece's sharpest point is that the labs themselves may be the players most in need of better incentives, since a safer model that's slower or pricier has to survive the market.

**[Scoreboard](/scoreboard) update:** Humans +1. Researchers stop trying to read AI's mind and start writing rules it can't profitably break.

Stay undead.