High prices do not prove AI agents formed a cartel—what does?
A causal audit asked whether rival agents punish a competitor for briefly cutting price. In stylised pricing and trading games, cartel-like outcomes often reflected rigid rules or anchors, while credible punishment was rare and fragile.
The short version
- What happened
- The audit copies a market state, forces one seller to undercut in one copy and then compares rivals' actions with an unperturbed copy under shared random seeds.
- Who it affects
- Competition authorities and businesses could use intervention-based audits to challenge suspicious pricing agents, but a failed toy-model test cannot clear a real system and a high-price observation alone cannot prove collusion.
- What remains unproven
- The unreviewed experiments use stylised games, small trained models and prompt-sensitive agent setups rather than live markets.
Keep in mind
The unreviewed experiments use stylised games, small trained models and prompt-sensitive agent setups rather than live markets.
The four-of-nine seed result is imprecise, and the paper reports few cheating experiences in several training runs.
OpenAI API calls could not perfectly share random seeds across counterfactual copies, weakening causal matching for those models.
The authors report $245 in OpenAI API costs and AI assistance with code and writing, but no separate funding or competing-interest statement.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The test looks for punishment, not just expensive outcomes
Parallel high prices can arise without an agreement: agents may copy a prompt, anchor on a currency value or independently learn a stable strategy. The researchers therefore cloned the same market state twice with shared randomness. In one copy, they forced a firm to undercut for a single round. If rivals subsequently lowered prices enough to erase the deviator's short-term profit, that would be causal evidence of an enforcement mechanism capable of sustaining a cartel.
They applied the audit to tabular Q-learning, small Qwen models trained with reinforcement learning and 13 prompted language models in two toy settings: logit-Bertrand pricing and Kyle-style trading. Q-learning produced enforcement only in the pricing game. A 0.6-billion-parameter Qwen model trained against a cheating rival showed thin punishment in four of nine seeds, but the cuts were too small to remove the benefit of deviation.[1]
The systems rarely learned a deterrent price war
No language model left to learn unaided developed a punishment strong enough to deter cheating. When researchers added a scripted price war during training, models could reproduce it, but mostly in the state where it had been taught. Among prompted models, punishment appeared in one of nine pricing conditions; many high-price outcomes were better explained by rules that ignored the market or fixed a nominal price.
The audit is a useful diagnostic, not a real-market finding. The games simplify demand, information and competitors; some learning seeds experienced only five to 23 cheating events. An exploratory GPT-5.2 exercise used eight sessions and selected two cartel-like outcomes after inspection. Shared random seeds were only approximate through the OpenAI API, and event-level intervals combine dependent observations. Field validation, richer markets and pre-registered thresholds would change the assessment.[1]
The distinction that matters
A counterfactual audit for cartel enforcement
The test asks what rivals do after the same market is altered in only one place.
Copy
Duplicate one market state with shared randomness
Creates a comparable control path
Deviate
Force one firm to undercut for one round
Creates a profitable challenge
Compare
Measure rivals' response and deviator profit
Separates punishment from naturally high prices