Why do rival AI pricing bots quietly settle on high prices together, with no one telling them to?
What market conditions cause AI pricing agents to collude versus compete normally?
This explores what makes AI pricing agents slide into charging above-competitive prices together (without anyone telling them to) versus what keeps them actually competing, and which features of the market or the agents themselves tip the balance.
This explores what pushes AI pricing agents into quietly agreeing on high prices versus what keeps them competing, and what the corpus says about which conditions matter. The starting point is a little unsettling. Collusion is the default outcome, not the exception. Independent pricing agents playing the same market over and over drift toward above-competitive prices without exchanging a single message. Lab experiments and data from Germany's 2017 gasoline market both show this Can independent AI pricing agents collude without communicating?. LLM-based agents do the same thing in a classic two-seller price war. They read each other's posted prices as signals and settle high, and small changes in prompt wording noticeably shift how high they go Do LLM pricing agents autonomously collude in repeated auctions?. So the useful question isn't whether agents collude. It's what breaks the pattern.
The strongest answer in the corpus is **sameness**. When rival agents are identical and 'patient' (they weigh future profits heavily), they hold prices about 22% above competitive levels. Pair a patient agent with a short-sighted one and the markup drops to about 10%. Pit different algorithm types against each other and sustained collusion disappears within 1,000 rounds Does agent heterogeneity prevent LLM pricing collusion?. More competitors and uneven access to data also wear it down. The same theme appears at market scale: populations of identical agents amplify each other's quirks into weaker competition, while mixed populations only partly offset that Does agent diversity actually improve market competition?. The policy worry follows directly. If every firm licenses the same few foundation models, the market drifts toward the conditions under which collusion holds best.
There is a twist, though. Not all differences help. Pair a large model (32B parameters) with a smaller one (14B) and collusion stays about as strong as with identical agents. The capability gap creates a leader–follower dynamic, with the bigger model setting a focal price and the smaller one tracking it Does model-size difference break LLM pricing collusion?. That is classic price leadership, rediscovered by machines. Raw intelligence doesn't protect against collusion either. Within a model family, more capable models reach collusion *faster*, and 94% of models across ten tested eventually got there Do more capable models resist collusion better?. Smarter agents are better at finding the cooperative equilibrium, not at resisting it.
Put together, collusion is most likely in concentrated markets with few sellers, identical or similarly-sourced agents, agents that weigh long-run payoffs, and a clear 'leader' to follow. This matches the UK competition authority's concern that autonomous agents in concentrated markets can dampen competition through learned mutual adaptation rather than any agreement a regulator could point to Could autonomous pricing agents dampen competition without explicit collusion?. A related result shows the same 'harm without coordination' logic on the buying side. When a platform and its sellers each separately tune their behavior to exploit AI shoppers' biases, consumer harm more than doubles, though nobody coordinated Can platform ranking and seller descriptions separately harm AI shoppers?.
The gap worth knowing about is that almost all of this evidence comes from small, stylized markets with two to a few sellers. The corpus offers a hint that thicker markets behave differently. In one formal model, an algorithm's welfare edge over plain competition shrinks toward nothing as the number of sellers grows Does algorithmic recommendation lose its edge as sellers compete more?. It doesn't yet show whether many-seller markets with real-world mixes of AI agents reliably compete. That is the open question for anyone deciding whether 'everyone uses the same model' is a competition problem.
Sources 9 notes
Laboratory experiments and Germany's 2017 gasoline market data show that AI pricing agents independently optimizing in repeated games converge on above-competitive prices without needing communication or explicit coordination between deploying firms.
Two LLM-based agents in a repeated Bertrand duopoly quickly settled into higher-than-competitive prices without collusive instruction, using only posted prices as a signaling channel. Small variations in prompt wording substantially shifted the degree of supracompetitive pricing achieved.
In 2,000+ compute hours of experiments, patient identical LLM pairs collude at 22% above competitive price, but pairing patient with myopic agents cuts this to 10%, while cross-algorithm heterogeneity eliminates sustained collusion entirely within 1,000 periods.
Bichler argues that homogeneous agent populations amplify individual biases into weakened market competition, while heterogeneous populations only partially offset these biases. Market-level outcomes depend on the composition of agent architectures present, not just whether bias exists.
Pairing a 32B with a 14B pricing agent maintains collusive price elevation near homogeneous levels through a clear leadership hierarchy, where capability asymmetry establishes focal coordination rather than breaking collusion like other heterogeneities do.
Show all 9 sources
Across ten models, more capable variants learned to collude sooner than weaker ones, though 94% eventually did. Capability speeds arrival at collusion but does not prevent it.
The CMA argues that when multiple businesses deploy autonomous agents optimizing pricing and commercial strategy in concentrated markets, interaction between these systems could reduce competitive pressure through learned mutual adaptation rather than explicit agreement.
A simulation found that when a platform optimizes its ranking and sellers optimize their descriptions—each separately targeting the same agent biases—consumer harm more than doubles compared to either acting alone, without requiring coordination.
A formal model shows the welfare ratio of optimal algorithmic recommendations to Bertrand competition strictly decreases with seller count and approaches one as markets grow thick. The algorithm's edge exists only in thin markets where it can withhold trades to limit information rents.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- On the Fragility of AI Agent Collusion
- Vertical tacit collusion in AI-mediated markets
- Agentic markets
- Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs
- LLM-Agent Interactions on Markets with Information Asymmetries
- Algorithmic Collusion by Large Language Models
- An Economy of AI Agents
- What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce