New study finds AI agents learn to collude, raising antitrust flags for 2026

Recent research reveals that AI agents learn to collude by independently limiting production to inflate profits, a behavior known as "tacit collusion." The research by Deshpande and Jacobson shows that LLMs, acting as autonomous agents in Cournot oligopoly simulations, consistently settle on anti-competitive outcomes without direct communication (arXiv:2601.17263). This finding intensifies a long-standing antitrust debate, echoing warnings from the OECD that algorithms pose significant "challenges for competition law enforcement" (OECD digital collusion paper).

This emergent behavior presents two major obstacles for regulators. First, traditional antitrust laws require proof of an explicit agreement, a condition that autonomous AI systems circumvent by design. Second, the "black box" nature of their learning processes makes it nearly impossible to assign blame or prove intent. Legal analyses highlight this enforcement gap, arguing that such algorithmic collusion may fall outside existing regulatory frameworks.

Why Cournot Simulations Matter for Policy

The study's use of the Cournot model is significant because it simulates real-world business decisions about production capacity and inventory. In the simulation, independent AI agents, given only basic market data, reduced total output significantly below competitive levels, effectively coordinating to maximize profits. This outcome, achieved through independent fine-tuning, validates concerns from sources like Reuters Practical Law that "algorithmic pricing could foster collusion" even without direct corporate coordination.

A new study found that AI agents can independently learn to restrict market supply to increase profits, a form of tacit collusion. This behavior emerges from the algorithms' optimization processes without any need for communication, challenging traditional antitrust laws that rely on proving an explicit agreement between competitors.

Emerging Enforcement Themes

In response, competition authorities are actively developing new theories of harm to address algorithmic collusion:
- Similar Software: Pursuing cases where rival firms use similar pricing software, leading to coordinated outcomes.
- Hub-and-Spoke Collusion: Investigating platforms that pool competitor data, enabling a central algorithm to make pricing decisions for all, as noted in a Goodwin alert.
- Autonomous Collusion: Targeting the exact scenario highlighted by the Cournot study, where AIs collude without human intervention.

Global regulators are mobilizing, with European agencies reportedly "focused on the competitive risks posed by algorithms" (Loyens & Loeff). In a significant move, the FTC and DOJ have joined their UK and EU counterparts to declare AI a top enforcement priority, per Reuters.

Proposed Safeguards and Proactive Regulation

With traditional enforcement lagging, legal and policy experts are championing a shift toward prevention. Proposed safeguards include mandatory design constraints, regular audits, and official certification for algorithmic pricing tools, as recommended by industry experts. Legal insights suggest regulators must scrutinize "how the system learns, what it observes, how quickly it reacts, and whether competitors share tools" to preemptively block harmful coordination.

Key Takeaways and Open Questions

While the study does not suggest collusion is inevitable in every AI-driven market, it confirms that the risk is real and influenced by factors like market concentration, data sharing, and algorithm design. The key takeaway is clear: the findings provide crucial empirical evidence supporting regulators' demands for proactive oversight. Future antitrust analysis must evolve beyond proving human intent to directly scrutinizing the behavior of the algorithms themselves.

---

What did the new study find about AI agents and collusion?

Deshpande and Jacobson used Cournot oligopoly simulations and discovered that LLM-based pricing agents can learn to sustain prices above the competitive level without any explicit communication or human intent. The agents reached tacit collusion through independent probabilistic optimisation, a finding documented in arXiv:2601.17263. Because the coordination emerges from code rather than from a written or spoken agreement, traditional antitrust doctrines that rely on proving intent or an explicit meeting of the minds may struggle to capture this behaviour.

Why are regulators worried about tacit AI collusion?

Courts and agencies have historically focused on agreement, concerted practice and mens rea. Legal analyses note that "algorithmic tacit collusion via supra-competitive price fixing" may fall outside the scope of existing regulatory frameworks, showing that current laws may leave an enforcement gap. The OECD's foundational paper already warned that "algorithms present [challenges] for both competition law enforcement and market regulation", and that question is now front-and-center as enforcement actions multiply.

Which real-world cases illustrate the risk right now?

• UK CMA has opened investigations into hotel chains suspected of sharing competitively sensitive data via a common analytics provider, a textbook hub-and-spoke model.
• RealPage litigation in the United States: the DOJ and multiple state AGs allege that the firm's software pooled non-public rental data and "induced customers to accept pricing recommendations", leading to higher rents (Wilson Sonsini, 2026).
• Poland's UOKiK has conducted probes into banking and pharmaceutical firms for using algorithms that drew on shared credit-risk databases to coordinate consumer-loan pricing.
  These cases show the risk is not hypothetical; it is already the subject of live enforcement and class-action filings across hotels, rentals, student housing, mobile homes and healthcare.

How are regulators adapting their toolkit?

Rather than waiting for post-hoc enforcement, agencies are shifting toward ex-ante safeguards:

• Mandatory audits and design constraints: Legal experts urge "algorithmic design constraints, mandatory audits and compliance certifications" to future-proof competition law.
• Broader data scrutiny: Industry reports indicate that regulatory agencies are flagging AI as a priority enforcement area.
• Screening shared software layers: Legal alerts warn that using the same algorithmic pricing provider can create a hub-and-spoke risk even when firms never speak directly.
  Taken together, preventive compliance is becoming as important as litigation defense.

What should companies do today to stay compliant?

1. Map AI pricing tools: inventory every third-party model, dataset and recommendation engine used across the business.
2. Implement human review checkpoints: avoid auto-accept features that implement AI suggestions without contemporaneous human judgment.
3. Audit data inputs: confirm that algorithms are not trained on pooled competitor data or sensitive forward-looking pricing signals.
4. Document decisions: maintain clear records showing that final pricing authority rests with people, not code.
5. Stress-test scenarios: run red-team exercises to see whether agents placed in rival shoes would converge on tacitly collusive strategies. Following these steps can reduce exposure under emerging enforcement patterns in multiple jurisdictions.