Publication date: February 26, 2026
Pricing API language models, collaboratively restricting access to the most advanced models, blocking interoperability, collaborative lobbying for regulations that hinder startup entry – is market ready for this?
An AI cartel is a situation in which AI-developing companies secretly or openly cooperate in a way that restricts competition—for example, by fixing prices, blocking access to technology, or jointly eliminating smaller players from the market.
One of the fundamental principles of a market economy is the protection of competition, and some of the greatest threats to its maintenance stem from agreements that limit competition in a given economic segment, leading, for example, to two entities taking control. In some cases, states recognize that they should have a monopoly in certain sectors of the economy, for example, with respect to key services. In most cases, however, monopolies, oligopolies, and cartels are phenomena and practices that are economically deeply unjustified from a state perspective and, above all, illegal. Most of the world, including the US and the EU, has regulations in place in this area.
Technological advancement is forcing a new perspective on this issue, as agreements concluded by entrepreneurs can take increasingly complex and unusual forms, making it difficult to detect illegal practices, which is ultimately the priority of the entities entering into the agreement. The last few years have forced even deeper reflection. What happens when algorithms co-decide or determine, for example, the price of a product in such a way that the result is a radical reduction in competitiveness in a given segment? Or, how should we assess a situation in which the algorithms themselves “understand” each other? For example, two systems operating for two entities, based largely on generative artificial intelligence, could independently reach the rather logical conclusion that cooperation with the other entity would yield significantly greater benefits, for example, through price dumping or, conversely, through regular, joint price increases. From a legal perspective, such a situation may be unclear, if only because most competition regulations were designed with more “traditional” types of regulations in mind. In the United States, this issue is regulated by an 1890 law, the so-called Sherman Act, which classifies certain agreements between companies as criminal offenses.
In the European Union, the most important provision regulating these matters is Article 101 of the Treaty on the Functioning of the European Union (TFEU), which regulates conduct incompatible with the internal market. Paragraph 1 of the Treaty states that: the following shall be prohibited as incompatible with the internal market: all agreements between undertakings, decisions by associations of undertakings and concerted practices which may affect trade between Member States and which have as their object or effect the prevention, restriction or distortion of competition within the internal market, in particular those which: directly or indirectly fix purchase or selling prices or other trading conditions; limit or control production, markets, technical development or investment; share markets or sources of supply; apply dissimilar conditions to equivalent services to trading partners, thereby placing them at a competitive disadvantage; or make the conclusion of contracts subject to the acceptance by the parties of additional obligations which, by their nature or trade practice, have no connection with the subject matter of those contracts. Importantly, agreements or decisions prohibited under this article are automatically invalid.
Exceptions are agreements that contribute to improving the production or distribution of goods or to promoting technical or economic progress, while allowing users a fair share of the resulting benefits, and without imposing on the undertakings concerned restrictions that are not indispensable to achieving those objectives, and without giving undertakings the possibility of eliminating competition in respect of a substantial part of the products in question.
The situation under this provision is clear with respect to all agreements, including those involving algorithms—classical or artificial intelligence. This is particularly supported by the wording any decision by associations of undertakings and any concerted practice which may affect trade between Member States. This is, of course, provided that we treat the algorithm solely as a tool for implementing the agreement. Technological developments, however, demonstrate that this may not be the case.
Algorithms in the Service of Monopolies? Possible Scenarios
In a document published by the Office of Competition and Consumer Protection, “Measures of Algorithmic Coordination of Business Behavior in the Light of European Union Antitrust Law,” four potential scenarios for the use of algorithms in collusion are identified.
The first is explicit collusion implemented through algorithms. This involves a situation in which the collusion is concluded between humans, and the algorithms are used solely to implement the agreement, possibly also monitoring price levels or automatically responding to deviations. In such cases, attribution of liability may be entirely possible. In such cases, we are dealing with behavior that can be classified as a concerted, conscious, and coordinated practice. The UOKiK document states that: in a scenario where algorithms are used instrumentally to implement or more clearly supervise a previously concluded collusion, the automation of cartel activities is merely an extension of human will. Therefore, as long as coordination corresponding to the premises of price fixing occurs, the precise means used for this coordination are irrelevant. While attributing liability for this type of collusion seems entirely possible, as confirmed by the example cited in the document of the British competition authority’s assessment of such a situation as restrictive action, detecting such agreements seems much more difficult. In this context, the European Commission’s inspection powers granted to it by Article 20 of Council Regulation (EC) No. 1/2003 of 16 December 2002 (as amended) are pointed out.
The second scenario is interesting because we are not dealing with a classic agreement in this case. It concerns a configuration involving the use of an algorithm in a so-called a hub-and-spoke relationship, where the hub is an entity higher than the competing entities and manages the algorithm. Therefore, there is no horizontal exchange of information between entities, but the effect and purpose are similar and restrictive of competition. In such a case, proving the existence of a cartel is obviously much more difficult, but still possible. In the CJEU judgment in Case C-74/14, the Court stated that the existence of a concerted practice or agreement, and thus a means of coordinating the conduct of undertakings, often stems from coincidences and circumstantial evidence which, when considered together, may constitute evidence of an infringement of Article 101(1) TFEU. The mere sending of information via a computerized system may justify a presumption that the travel agencies were aware of it and, therefore, that they participated in the concerted practice. However, this presumption may be rebutted by proving that they did not receive the message or that they did not read it. The moment of acceptance of participation in the concerted practice was therefore determined by the moment of becoming aware of the content of the message. However, in a case where the travel agency openly distanced itself from the practice or notified the relevant authorities of it, there is no liability for the infringement.
What if the algorithms themselves lead to the creation of cartels?
An even more subtle and difficult-to-analyze scenario is one in which there is little or no communication between individuals. One can imagine a scenario of tacit collusion achieved through algorithms, in which companies with access to enormous amounts of data, using independently applied pricing algorithms, achieve a monopolistic price level in the market without communication. The UKOiK report illustrates this case by citing the American island of Martha’s Vineyard, where only a few businesses operate gas stations, and the supply and demand situation is highly predictable. In such circumstances, any price reduction by one business proves unwise, because each such reduction will also generate reductions by competitors, which will consequently mean a reduction in margins for each business. Therefore, the only profitable strategy from the gas stations’ perspective is to consistently maintain prices above the competitive level. Exactly the same situation can be imagined in this scenario. The increasing use of algorithms in pricing (common in some industries, such as the aviation market) may lead to a situation in which the rational decision of every business in a given segment will also be to use pricing algorithms to keep up with dynamic market price changes resulting from algorithmic manipulation. However, if almost every or every significant market player uses algorithmic pricing in this way, the situation will resemble that of the American island. Ultimately, any price reduction for one business by an algorithm will not prove effective and will not encourage purchases, as it will automatically result in a price reduction for competitors as well. In a situation where businesses cannot compete for customers by lowering prices, it is rational to raise prices to a level higher than would result from the balance between supply and demand. As the report indicates, such a situation will primarily occur in highly homogeneous markets. From a legal perspective, such a situation is difficult to assess because it clearly restricts competition, but without the communication that would allow for the recognition of a restriction of competition as a result of an agreement between businesses. At this point, it seems that this situation is not clearly regulated, but is clearly a consequence of technological development and the automatic collection of large aggregates of input data about the environment and processing it using algorithms, which, of course, is not forbidden in itself.
The final and most far-reaching scenario is currently still a theoretical construct, which, however, due to the progress achieved in the field of artificial intelligence in recent years, seems not so unlikely. This involves a situation in which, instead of classical programmed algorithms used to establish price levels, but AI-based systems with the ability not only to reproduce learned rules but also to infer, generalize, and generate entirely new content, would lead to such collusion. One of the key characteristics of artificial intelligence, in addition to those mentioned above, is a certain level of autonomy, as well as the ability to self-learn and train. In the third scenario mentioned earlier, it would be possible to analyze whether the algorithm created by the programmer was designed to lead to the creation of a cartel and restrict competition, or whether, for example, it is the result of widespread use of algorithms by all market players, and such liability cannot be attributed to it. However, this may be different in the case of artificial intelligence-based systems, if only because, once designed, they can “evolve” in a slightly different direction, meaning that the static element present in classic algorithms is missing. A situation in which advanced algorithms independently recognize interdependencies and behave independently in parallel does not differ from the third scenario regarding less complex algorithms in homogeneous product markets. Such behavior should likely be considered consistent with Article 101(1) TFEU.
In some cases, the algorithm provider may also be liable. The UOKiK report indicates that several conditions must be met, however. First, if the supplier acted under the control of an enterprise engaged in a practice that violates competition law. Second, if the algorithm creator is aware of the anticompetitive objective pursued by enterprises in another market and intends to contribute to it through their conduct. Third, and finally, if the supplier could, with due diligence, have foreseen the occurrence of a practice that restricts competition and was willing to accept the resulting risk. Important changes may result from regulations not directly related to competition protection, but others imposing numerous obligations on suppliers and software developers, such as the AI Act, which establishes transparency obligations and supervisory instruments, or the DSA, which introduces disclosure requirements and algorithm audits. Such regulations can help ensure system transparency and limit the potential use of algorithms, including artificial intelligence-based systems, for monopolistic practices.
Pricing for language model APIs could be an example of an AI cartel. A pricing cartel occurs when competing companies agree on prices among themselves, rather than setting them independently. In the context of AI, this would mean, for example, jointly setting a minimum price for 1 million tokens, agreeing on API price increases, establishing common free access thresholds, or eliminating cheaper plans. This would be a classic antitrust violation. In such a case, regulators (e.g., the Federal Trade Commission or the European Commission) would look for signals such as identical and simultaneous price increases without cost justification, disclosed communications between companies, joint agreements on standards that restrict competition, or closing off access to GPU infrastructure to new players. In this context, for example, cloud investments in AI startups or long-term infrastructure contracts, while not illegal in themselves, can reduce real competition.