Breaking Algorithmic Collusion in Human-AI Ecosystems
Natalie Collina, Eshwar Ram Arunachaleswaran, Meena Jagadeesan
TL;DR
This paper analyzes the stability of algorithmic collusion in human-AI ecosystems, showing that human defections can destabilize collusive pricing and lead to more competitive outcomes.
Contribution
It provides a theoretical framework for understanding how human defections impact collusion among AI agents in repeated pricing games.
Findings
A single human defection can destabilize collusion and lower prices.
Multiple defections push prices closer to competitive levels.
The nature of collusion changes when AI agents are aware of defections.
Abstract
AI agents are increasingly deployed in ecosystems where they repeatedly interact not only with each other but also with humans. In this work, we study these human-AI ecosystems from a theoretical perspective, focusing on the classical framework of repeated pricing games. In our stylized model, the AI agents play equilibrium strategies, and one or more humans manually perform the pricing task instead of adopting an AI agent, thereby defecting to a no-regret strategy. Motivated by how populations of AI agents can sustain supracompetitive prices, we investigate whether high prices persist under such defections. Our main finding is that even a single human defection can destabilize collusion and drive down prices, and multiple defections push prices even closer to competitive levels. We further show how the nature of collusion changes under defection-aware AI agents. Taken together, our…
Peer Reviews
Decision·Submitted to ICLR 2026
The paper presents a stylised setting to analytically characterise the price effects caused by agents deviating from the equilibrium profile and makes a solid theoretical contribution. The methodology appears to be correct and reproducible. Generally, the paper is well structured with clarity in the definitions and the whole approach. The problem the authors are investigating is timely and needs research attention.
The authors do acknowledge the limitations of the simplifications in the current work for analytical tractability. However, I find that the motivation that AI agents play Nash equilibria in the repeated game needs strengthening. The authors use the terminology of “AI agents” and “no-regret learners,” but mathematically treat them as given strategy classes, not as adaptive processes. Thus, while the topic (AI collusion) is socially and technically relevant, the learning connection is mainly in th
1. The overall writing is easy to follow. No major errors. 2. Precise formalization and clear proofs of theorems and lemmas. 3. Useful scope of extensions. Covers single vs. multiple defectors, no-regret vs. FTL, and defection-aware agents, which sharpens the paper’s scope.
1. No empirical evidence. This paper is purely theoretical without any experiments or simulations of RL agents or LLM agents. Since it claims relevance to RL/LLM agents, adding empirical evidence would validate the theory in practice. 2. This paper assumes RL/LLM agents would play equilibrium strategies. However, prior work suggests RL [1,2] and LLMs [3,4] may not be able to converge to Nash equilibria in repeated interactions and results vary across different RL algorithms and LLMs. Either jus
This is a very well written paper that tackles the timely issue of algorithmic collusion. In particular the results, although theoretical, give practical, actionable insights that are useful to alleviate algorithmic collusion caused by non-myopic multi-agent interactions. 1. **Clarity:** The paper is extremely well written. All the necessary background is covered as well as important mathematical definitions required to analyze and interpret the results. 2. **Importance of the problem:** The
Although the paper is very rigorous it has some major weaknesses that reduce my confidence in accepting it: 1. **Lack of experimental results:** No experiments (even toy) to illustrate convergence of prices under standard online learners or simple scripted LLM agents. Empirical evidence could potentially ground the theoretical results. Authors also note this is a stylized model. 2. **Modeling assumptions:** The assumptions that most weaken the paper, in my view, are: Equilibrium/coordination:
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEvolutionary Game Theory and Cooperation · Language and cultural evolution · Ethics and Social Impacts of AI