Are AI agents now independently advocating cartel behaviour?

Is competitor collaboration always a bad thing?

Competitor collaboration is not necessarily harmful to competition – given the right circumstances and safeguards, it can produce effects which are beneficial to the competitive process and ultimately consumers. This is reflected in the existence of block exemptions at EU and UK levels in areas including Research & Development, as well as a more general exemption from competition rules under Article 101(3) of the Treaty on the Functioning of the European Union and section 9 of the Competition Act 1998 respectively. 

The issue with AI agents collaborating with each other is to do with the potential scale of the collaboration and lack of human supervision, as well as the lack of guardrails on collaboration spilling over into competitive markets from non-competitive parameters (such as ‘upvoting’, per the original author). This begs the essential question: can/should humans be held responsible for the anti-competitive behaviour of AI agents they design, maintain and/or use to make commercial decisions? 

To answer this question, it would seem helpful to unpack the different types of anti-competitive behaviour which AI Agents might engage in as well as the different categories of humans who may be connected to the behaviour. 

Four main types of AI Agent anti-competitive behaviour 

Looking at previous academic and policy commentary on pricing algorithms (which seem to bear some relevance to AI Agents scenarios too), it is possible to identify four main types of anti-competitive behaviour which AI Agents could engage in.  

Ezrachi and Stucke have written extensively on the various ways in which algorithms can facilitate collusion, both explicit and tacit. They have identified four non-exclusive categories of collusion: messenger, hub and spoke, predictable agent, and digital eye(1). Similarly, the Organisation for Economic Co-operation and Development (OECD) devised a list of possible roles which algorithms may have in facilitating collusion: monitoring, parallel, signalling and self-learning(2). In essence, the OECD’s categories resemble those of Ezrachi and Stucke. 

Broadly speaking, these four types cover: 

• the Messenger/ Monitoring category, whereby humans would rely on AI Agents to directly execute their instructions to support the implementation of a traditional cartel arrangement. Under this scenario, humans remain in control of the collusion, with the algorithmic software acting as an intermediary to communicate and/ or monitor implementation, making the issue of responsibility straightforward to resolve(3);
• the Hub and spoke/ Parallel category, whereby users of AI Agents who compete with each other downstream would use similar algorithms or data pools to make commercial decisions (such as determining prices or other key parameters of competition). In this scenario, humans in competition with each other delegate their decision-making to a single technological tool, leading to the same parameters of strategic thinking being applied and aligned across the market. Again, competition authorities have previously found that responsibility attribution remains relatively straightforward in these cases, as humans are seen to have consciously delegated their decision-making to a common source(4).
• the Predictable agent/ Signalling category, whereby competitors use their own individual AI Agents and unilaterally instruct each of them to monitor and respond to competitors’ pricing or other key parameters of competition. In this scenario, while there is likely to be a harmful effect on competition in the form of sub-competitive conditions for consumers, and human actors can still be considered to play a guiding role for the AI Agent’s behaviour, competition authorities would likely struggle to enforce competition rules, for lack of an agreement or concerted practice between market players (while abuses of collective dominance are notoriously difficult to establish in practice); and 
• finally, the Digital eye/ Self-learning category, whereby, without any specific direction from humans, AI Agents find through a process of self-learning that collusive behaviour (whether explicit or tacit) is the optimal course of action to achieve the overall target that a human has set for them – for instance, to maximise profits. As with the above scenario, although a harmful effect on competition could be evidenced, it is not clear that ex post competition rules could bite, while the question of attributability to humans become more complicated due to the lack of explicit human direction to resort to anti-competitive methods.  This has been referred to as the ‘vending machine test’.  A great example was reported by Sky News last week when Claude Opus 4.6 “passed” the vending machine test by maximising profits through deception, including lying, cheating, and even stealing when instructed to do whatever it took. The episode has been cited as an early warning of how AI systems can independently adopt anti-competitive or unethical strategies when given open-ended optimisation goals

AI Agents pushing ex post competition rules to their limit?

The Moltbook incident mentioned above would seem to best fall under the fourth category – digital eye/ self-learning. In a sense, the fact that the discussion took place on a forum for AI Agents simplifies the issue, in that it involves communication between different actors (albeit not humans). The humans in charge of these AI Agents’ self-learning and communication parameters could potentially be held responsible for failing to put in place safeguards to prevent breaches of competition rules from taking place between their AI Agent and others. This would be a novel development, but one that would seem fairly reconcilable with existing caselaw, particularly in relation to exchanges of commercially sensitive information between competitors, and the related requirement for information recipients to publicly distance themselves if they do not want to be included in the competition authority’s charging sheet(5). In many ways, it is not that far removed from the rules that already apply to algorithm logs and kill switch requirements in high-frequency trading.

However, in the absence of direct communications between AI agents, there is a credible argument that ex post competition rules would not be sufficient to capture behaviour falling within the last two categories – Predictable agent/ Signalling and Digital Eye/ Self-learning. As the CMA put it in its 2021 paper on pricing algorithms: “It is as yet unclear that competition authorities can object to hub and spoke and autonomous tacit collusion situations where, for example, there may not have been direct contact between two undertakings or a meeting of minds between them to restrict competition(6)".

A related issue is the question of what amounts to “publicly available information”. For example, what is the status of data made available via Open Source, which AI Agents typically rely on for self-learning? Does it qualify as publicly available information if every AI Agent can technically access it? What if humans cannot? This has the potential to create a highly transparent environment (at least for AI Agent capabilities) based on a common source of background information from which each AI Agents may individually develop parallel marketing strategies for its users. In this regard, query whether the existing caselaw on the exchange of sensitive information could be stretched to a scenario where a market is not highly concentrated in the first instance, and where the information exchange took place upstream at the AI Agent self-learning stage, using Open Source data.

Based on the above considerations, it appears that other forms of regulation may need to be introduced in order to prevent anti-competitive effects arising from parallel behaviour carried out by each AI Agent entirely independently. 

A number of key jurisdictions are currently looking to regulate AI services, though not always through the prism of competition law. For example, the EU AI Act(7) is the first comprehensive legal framework on artificial intelligence worldwide. It prohibits AI practices including the exploitation of vulnerabilities, social scoring and practices relating to facial recognition and biometric identification – which gives this piece of legislation a data protection slant. As discussed in more detail in the section below, the ongoing focus from competition authorities in key jurisdictions may lead to further AI regulation which will encompass competition considerations.

It is worth noting that AI Agent designer Anthropic recently publicised a new Constitution for its AI Agent Claude(8). If achievable technically, the setting of overriding parameters/ rules of engagement to secure pro-competitive behaviour as a basic requirement could be an avenue for future regulation, which could even be achieved on an industry, self-regulating basis.

In the meantime: 

• designers of AI Agents may wish to consider putting in place various safeguards to frame self-learning processes and monitor the activities of their AI Agents so as to be in a position to swiftly address any competition concerns; and
• users of AI Agents may wish to consider the terms of their supplier contracts to better understand the basis on which these AI services are being provided to them, and if appropriate implement AI use policies across their business to mitigate any competition risks they might identify.

The rapid evolution of GenAI has prompted a wave of scrutiny from authorities and policymakers around the world. Looking at authorities with competition functions specifically, as AI technologies become increasingly central to economic growth and digital innovation, regulators are grappling with the challenge of fostering innovation while ensuring that markets remain open, competitive, and fair. 

Examples of latest ongoing activities include: 

• The UK Financial Conduct Authority’s launch of the Mills Review(9), which is set to look into how advances in AI could transform retail financial services and in what ways regulators may need to adapt their approach;  
• The Digital Regulation Co-operation Forum’s Agentic AI initiative(10); and 
• The French Competition Authority’s consultation on AI Agents(11).

Recent reports, legislative changes, and policy statements from major jurisdictions reveal a convergence of concerns—alongside some regional nuances—about the risks and opportunities presented by GenAI.

Key competition concerns: market concentration, critical input foreclosure and pricing collusion

A recurring theme across most jurisdictions is the potential for market concentration. The development of GenAI relies on access to substantial resources: vast datasets, specialised computing infrastructure, and highly skilled talent. This creates a natural advantage for large, established technology firms, often referred to as “incumbents,” who can leverage their existing assets and relationships to reinforce their market positions. Several authorities, including those in the UK, France, and the EU, have highlighted the risk that a handful of players could dominate the value chain, from foundational model development through to downstream applications.

Another area of concern is the control of critical inputs such as cloud computing power and proprietary data by vertically integrated firms. These companies are not only suppliers of essential infrastructure but also direct competitors in AI model development. This dual role can create bottlenecks, raise entry barriers for new players, and potentially lead to exclusionary practices. Reports from the US Federal Trade Commission (FTC) and the Portuguese Competition Authority, among others, point to the possibility of preferential treatment for partners, increased switching costs due to exclusivity clauses, and technical lock-in via proprietary hardware and software.

Regulatory approaches and principles

Authorities in the US, UK, and India have raised concerns about algorithmic collusion, where autonomous systems may coordinate pricing or other market-sensitive behaviours, either intentionally or unintentionally. Approaches to this issue vary, particularly in the US, where courts have reached different conclusions (see contrasting judgments in footnote 4 below), and where states such as California and New York have enacted targeted legislation to restrict the use of pricing algorithms in sensitive sectors like rental housing. Despite differences in legal traditions, there is broad alignment among international authorities, including the G7, the European Commission, and regulators in the UK, Japan, on key principles for competition policy in the AI era, such as fair access, transparency, interoperability, and accountability, as well as the need for ongoing market monitoring, adaptive enforcement, and enhanced cooperation between agencies. However, regulatory responses are not uniform: the US FTC has focused on the contractual and technical aspects of AI partnerships, the EU is considering designating certain AI services as “gatekeepers” and addressing enforcement gaps in complex partnerships and minority investments, and countries like Australia and South Korea are updating their merger review regimes to better address the dynamics of digital and AI-driven markets.

Emerging issues: environmental impact and talent

Some jurisdictions are beginning to look beyond traditional competition issues. The French Competition Authority, for example, has flagged the energy and environmental impact of AI as a potential source of competitive distortion, particularly where access to electricity or privileged supply agreements could create new barriers to entry – for more information on this, see our January 2026 briefing here. The acquisition and retention of expert talent is also coming under the spotlight, with concerns that dominant firms could restrict labour mobility and innovation through non-compete clauses or aggressive hiring.

One of the biggest challenges facing regulators is the speed at which AI markets are evolving – and the Moltbook incident is a case in point. The direction of travel is clear: competition authorities are determined to ensure that the transformative potential of GenAI is realised in a way that benefits consumers, supports innovation, and avoids the pitfalls of excessive market concentration. But traditional competition enforcement tools may be too slow or ill-suited to address the rapid emergence of new business models and partnership structures. There is also the risk that some anti-competitive conduct may fall outside the scope of current laws, especially where it involves non-obvious or novel forms of collaboration.

1. Ezrachi and Stucke, ‘Artificial Intelligence & Collusion: When Computers Inhibit Competition’ (2017) University of Illinois Law Review 5, 1782. 
2. Organisation for Economic Co-operation and Development, ‘Algorithms and Collusion: Competition Policy in the Digital Age’ (2017); accessed on 9 February 2026 via this link: https://web-archive.oecd.org/pdfViewer?path=/2019-02-17/449397-Algorithms-and-colllusion-competition-policy-in-the-digital-age.pdf.  
3. For example, the UK CMA addressed this scenario in its August 2016 Infringement Decision in the Online Sales of Posters and Frames Investigation (Case 50223, Trod Ltd/GB Eye Ltd): accessed on 9 February 2026 via this link: https://assets.publishing.service.gov.uk/media/57ee7c2740f0b606dc000018/case-50223-final-non-confidential-infringement-decision.pdf. Two companies – Trod Ltd and GBE – engaged in an anticompetitive agreement concerning the sale of licenced sport and entertainment posters sold on Amazon UK. The CMA found that the companies agreed not to undercut each other on prices for posters and frames, and that this arrangement was implemented using automated repricing software, and the maintenance of contact between the parties throughout the implementation period. The CMA was able to easily fit the practice into the existing legal framework and found it in breach of UK anti-competitive rules.
4. For example, the CMA secured commitments from motor insurance providers in December 2011 in relation to the Whatif? pricing software product – decision accessible via this link (accessed on 9 February 2026): https://assets.publishing.service.gov.uk/media/555de4dbe5274a74ca000159/OFT1395.pdf; the Court of Justice of the European Union confirmed to the Lithuanian Supreme Administrative Court in 2016 in Case C-74/14 Eturas that liability could be attributed to travel agents using a common booking software which had capped the level of discounts that could be made via the booking platform, provided they were aware of the practice and had not publicly distanced themselves from it (akin to “tacit collusion” in a commercially sensitive information sharing context), accessible via this link (accessed on 9 February 2026): https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex:62014CJ0074; while there have been a string of cases in the US: U.S. v David Topkins (US Department of Justice, ‘Former E-Commerce Executive Charged with Price Fixing in the Antitrust Division’s First Online Marketplace Prosecution (2015); accessible via this link (accessed on 9 February 2026): https://www.justice.gov/opa/pr/former-e-commerce-executive-charged-price-fixing-antitrust-divisions-first-online-marketplace); and more recently Gibson v Cendyn Group, In Re. RealPage Inc., Cornish-Adebiyi v. Caesars Entertainment Inc., Duffy v. Yardi Systems Inc., and Mach v. Yardi Systems, Inc.
5. See, e.g. Ofcom decision published in February 2023 in Motorola/ Sepura, accessed on 9 February 2026 via this link: https://www.ofcom.org.uk/siteassets/resources/documents/about-ofcom/bulletins/competition-bulletins/all-cases/cw_01241/motorola-sepura-non-confidential-ca98-decision.pdf?v=320574. 
6. CMA, Algorithms: How they can reduce competition and harm consumers, 2021, paragraph 2.87. Accessed via this link on 9 February 2026: https://www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-and-harm-consumers. 
7. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. 
8. https://www.anthropic.com/constitution (accessed on 9 February 2026).
9. See Financial Conduct Authority webpage here (accessed on 9 February 2026): https://www.fca.org.uk/publications/calls-input/review-long-term-impact-ai-retail-financial-services-mills-review. For a more general discussion on the potential liability of senior financial services managers for AI-based decisions, you can access AG’s recent briefing here: https://www.addleshawgoddard.com/en/insights/insights-briefings/2026/global-investigations/senior-managers-liable-under-uk-regulatory-regime-decisions-made-ai// 
10. See DRCF webpages (both accessed on 9 February 2026): https://www.drcf.org.uk/news-and-events/news/drcf-launches-thematic-innovation-hub-for-innovators-with-first-focus-on-agentic-ai; and https://www.drcf.org.uk/news-and-events/news/call-for-views-agentic-ai-and-regulatory-challenges.
11. See French Competition Authority announcement in English (accessed on 10 February 2026): https://www.autoritedelaconcurrence.fr/en/article/autorite-launches-public-consultation-conversational-agents