Can Pricing Bots Learn to Collude on Their Own?

Automated pricing can drift toward high prices without anyone telling it to. Put a human check on it.

Dynamic pricing is one of the easiest AI wins to sell internally. You plug in a tool, let it watch the market and let it adjust prices. Nobody on your team ever has to decide to raise prices for it to happen.

That's also the risk. A 2026 research paper collects evidence that pricing algorithms competing against each other can settle on higher prices together, without any agreement and without anyone instructing them to.

What the research says, and who found it

The paper is Solipsistic Superintelligence is Unlikely to be Cooperative, presented at ICML 2026. Its co-authors include researchers affiliated with Google DeepMind. One point matters for how you read everything below: the paper runs no pricing experiments of its own. It summarizes other researchers' work to support a broader argument about AI systems interacting with each other.

The foundational study it cites is Calvano et al. (2020), published in the American Economic Review. The paper sums it up as:

"pricing algorithms interacting in markets learn supracompetitive prices without explicit communication" (p. 3, citing Calvano et al., 2020)

"Supracompetitive" means higher than you'd expect in a competitive market. "Without explicit communication" means no messages and no agreement between the systems.

The paper also cites newer work on AI agents built on large language models:

"LLM-based pricing agents converge on supracompetitive prices in oligopoly settings without explicit collusion instructions (Fish et al., 2026), divide markets when deployed in multicommodity Cournot competitions (Lin et al., 2025)" (p. 22)

Those settings are experimental, simulated markets: stylized oligopolies and economic games. They are not documented cases of real companies. Fish et al. and Lin et al. are arXiv preprints.

Nobody has to mean it

The idea worth understanding here is that this can happen even when nobody has done anything "wrong" in the setup:

"A system can be perfectly aligned with its specification, values included, and still make things worse once it acts among other adaptive systems." (p. 3)

And:

"Pricing algorithms learn against each other, producing emergent collusion that no single system was designed to pursue" (p. 5, citing Calvano et al., 2020)

Each tool is doing its job, which is to maximize revenue. The problem appears in how the tools interact. The paper says the resulting outcomes "benefit the colluding parties at the expense of third parties" (p. 22). In a market, those third parties are usually your customers.

The caveats matter

This is not a research finding that pricing bots always collude, or that yours does. The authors are explicit that results depend on the setup:

"The dependence of these outcomes on game structure, training procedure, and initial conditions is itself the point. Cooperation under deployment is context-determined" (p. 9)

They also acknowledge counter-evidence: "Recent work does report cooperation emerging from intergroup competition in language model agents (Tonini & Galke, 2025)" (p. 9). When the paper uses the word "provably", it refers to one specific result about a particular type of learning algorithm in an iterated prisoner's dilemma (Bertrand et al., 2025, p. 9). It doesn't refer to commercial pricing tools.

The paper also makes no legal finding. It doesn't say any of these systems broke a law, and neither do I. Whether a given pricing practice raises competition-law issues is a question for a lawyer, not a blog post.

Why an operator should still care

You don't need certainty to manage a risk. You need to know it exists and have a cheap way to watch for it. The research suggests that "set it and forget it" pricing, running in a market where competitors use similar tools, deserves more oversight than most businesses give it. The paper notes, more broadly, that "Regulation exists because markets, left to their own dynamics, produce externalities, collusion, and instability." (p. 8) Inside your own business, the equivalent of that regulation is a human who's paying attention.

Practical takeaways

1. Put a human check on the price, not just the tool. Set bounds and an escalation path. Unusual or sustained price rises should land on a person's desk before they reach customers.
2. Watch the trend alongside your competitors', using public information. If your prices and theirs are rising together while costs and demand aren't, that pattern deserves a closer look.
3. Ask your vendor hard questions. Find out what the tool optimizes for, what data it uses (including any competitor data), and what controls you have. Write the answers down. Be cautious with tools that pool non-public pricing or sales data from competing businesses, and don't share your own non-public pricing data with competitors, directly or through a shared tool. Ask your vendor whether it does this.
4. Don't count on 'the algorithm did it.' As a working rule, treat the prices your tools set as your business's prices, and ask a lawyer what that means for you.
5. Review regularly, not once. These systems learn and markets change. A setup that looked fine at launch can drift, so review pricing outcomes on a schedule.

This post is general commentary for business operators. It is not legal advice, and nothing here suggests that any particular pricing tool or practice is unlawful. Talk to a qualified attorney about your specific situation. The author is not a lawyer, and reading this doesn't create an attorney-client relationship.

Sources

• Trivedi, R. S., Jaques, N., Cross, L., Vezhnevets, A. S., & Leibo, J. Z. Solipsistic Superintelligence is Unlikely to be Cooperative. Proceedings of the 43rd ICML (PMLR 306), 2026. arXiv:2606.03237. https://arxiv.org/abs/2606.03237. Cited pages 3, 5, 8, 9, 22.
• Studies cited within the paper (not independently reviewed for this post):
  • Calvano, E., Calzolari, G., Denicolò, V., & Pastorello, S. (2020). "Artificial intelligence, algorithmic pricing, and collusion." American Economic Review, 110(10), 3267–3297.
  • Fish, S., Gonczarowski, Y. A., & Shorrer, R. I. (2026). "Algorithmic collusion by large language models." arXiv:2404.00806 (preprint).
  • Lin, R. Y., Ojha, S., Cai, K., & Chen, M. F. (2025). "Strategic collusion of LLM agents: Market division in multi-commodity competitions." arXiv:2410.00031 (preprint).
  • Bertrand, Q., Duque, J. A., Calvano, E., & Gidel, G. (2025). "Self-play Q-learners can provably collude in the iterated prisoner's dilemma." Proceedings of the 42nd ICML.
  • Tonini & Galke (2025), as cited in the paper (p. 9).