Wired Partners: Prohibitions, Announced Stings and Audits as Deterrents to Collusion Between LLM Agents

This is an ambitious and rigorous study. More than 11,000 decisions across ten models, with matched contexts, confirmatory tests corrected for multiple comparisons and equivalence tests, give strong support for a clear conclusion. Agents decline explicit collusive offers mostly because of the enforcement they read into a prohibition. Comparing the results with what a risk-neutral agent should do is informative. It shows that agents respond to severity and rate, though not in the way expected value predicts. The over-deterrence result and the safe-harbour fix are useful for anyone writing platform rules, and the Colosseum transfer test is a good step toward more realistic settings.

Suggestions:

1. Studies 1 to 4 are single decisions in one game, with the rule in the operator's prompt. Real platforms usually deliver rules as terms of service or tool output, so testing that delivery would show whether the effect carries over.

2.In Study 5 the secret channel gave no reliable advantage, so this might not show whether rules deter collusion that pays. A task in which coalitions clearly gain from coordinating would make the test informative.

3.The benefit B is stated by the sender, who may be the plant. Stating B as a fact of the game would make the expected-value comparison cleaner.

• Grounding this in enforcement theory and expressive law is novel. The central finding is useful for platform operators: a prohibition deters mainly through the enforcement agents read into it.
• The design is strong: 10 models, 11k paired decisions, confirmatory families, and good controls, including a placebo penalty and a house-rule sentence that prohibits nothing.

This is a strong paper as it takes a fresh approach to AI collusion by borrowing from deterrence research: when we tell LLM agents not to collude, does the rule itself stop them, or is it their belief that someone is checking? The answer is clear and convincing. Across 11,302 decisions from ten models, a simple ban on secret alliances cut acceptance from 76% to 28%. But one extra sentence, “breaking this rule carries no penalty, and nobody checks,” pushed it back up to 69%. So the rules mostly work when the agents think someone is watching. One thing I wasn’t sure about is whether the agents are actually being deterred or just refusing anything secret. Once a rule was in place, they also turned down a harmless, permitted private deal 84.5–90.5% of the time, and it would be good to tease those two apart. The Colosseum study can’t tell us much yet either. The secret channel didn’t give a real advantage even to agents told to collude, so there was nothing for the rules to stop. A setup where cheating actually pays would make that test meaningful. I would also like to see more trials and more large, recent models, since each condition has only 20 trials per model and most of the models are small. Repeated games and rules that aren’t announced ahead of time would be worth trying too. Every main study is a single decision with the rule stated up front, and that may tip agents off that they’re being tested.