The AI (artificial intelligence) revolution is making its presence felt in commercial activity, with commercial decisions being increasingly influenced or even dictated by AI agents. Communications between these agents can be analyzed by competition authorities as forms of collusion or behavioral alignment, raising the issue of clearly delineating between illicit behavior and simple market-driven parallelism. [1]

The commercial use of software for recommending or dynamically adjusting prices represents an increasingly widespread practice, but this must not be confused with the autonomous negotiation of an agreement between competitor agents. The existence of technology or clients from the same industry using it does not prove anticompetitive information exchanges between market players, nor communication between their models. [1]

Enterprises must independently determine their conduct, but have the right to intelligently adapt to the observable behavior of competitors. Multiple pricing algorithms may adopt identical strategies if they process the same information, pursue similar objective functions, or operate under comparable economic conditions. The same coincidence may, however, be the result of coordination achieved through information exchanges, and in the absence of evidentiary elements, the observed behavior remains compatible with both explanations. [1]

In the hypothesis of opaque communication between AI agents, the meaning of messages is unknown, and the mere existence of contact does not automatically prove anticompetitive content. When opacity is total, proving collusion becomes considerably more difficult. Although the competition authority may demonstrate the existence of communication, if the content remains opaque, it will not be able to directly prove whether the exchanged messages had an anticompetitive purpose or not. [1]

With the burden of proof resting on the competition authority, it may trace the chronology and metadata of exchanges, logs and price decision history, parameters, versions and documentation of systems, as well as reactions to deviations. Additionally, to demonstrate the existence of an agreement and/or concerted practice, the authority may resort to controlled experiments, including interrupting communication between AI agents, and to counterfactual economic analyses. [1]

The evidentiary instruments mentioned cannot, in isolation, demonstrate the existence of coordination, requiring broad corroboration. The authority could attempt to prove that the totality of evidence regarding the effect of communication leaves no room for the hypothesis of a plausible alternative explanation. Thus, the authority can monitor the actions of AI agents both while communication is active and when it is interrupted. [1]

A causal relationship between AI agents' communication and prices may be indicated when, while opaque communication is active, agents maintain high prices and react to deviations, and at the moment communication is interrupted, prices fall and parallelism decreases. Additionally, the authority can analyze price evolution during the period of communication and in its absence, to prove that high prices are due to collusion and not independent reactions. [1]

The competition authority must carry out a thorough analysis of the context, market structure, and nature of information, ruling out possible alternative explanations. It must be able to prove that the supposed information exchanges between autonomous AI agents substituted competitive uncertainty with practical cooperation. [1]

The situation differs in the hypothesis of a single opaque contact, through which the entire anticompetitive mechanism is established, followed by the cessation of transmissions. Such a single contact may, in principle, be sufficient for the existence of a concerted practice, but the difficulty remains proving its content. If there is human intervention regarding the AI agent's proposals, the theory of exclusive coordination through agents is called into question, imposing on the authority the need to prove collusion also at the level of human validations. [1]

The authority can track the structural discrepancy of behavior immediately after contact, its selectivity (an AI agent following only the other’s price increases), and can compare previous and subsequent versions of the agents in an isolated environment. The use of AI agents does not lack competition risks, but these can be reduced or eliminated through measures such as using internal AI solutions, which do not communicate with public solutions and are subject to human validation, ensuring an architecture compliant with competition regulations. [1]

As the story develops

Update

Algorithmic collusion: the red line between lawful parallelism and illicit coordination

Investigators face the dilemma of determining whether similar prices set by competing firms using artificial intelligence are the result of collusive agreements or lawful parallelism, prompting a complex analysis of algorithmic behavior. [2]

The use of price adjustment software does not itself prove the existence of an anticompetitive agreement, and price similarity is not direct proof of coordination. [2]

The risk of algorithmic collusion arises when competitors' systems transmit information to each other, not when they process similar data to obtain identical results. [2]

Communication between algorithms does not automatically explain the final result; the investigation must identify the common source of information. [2]

In the case of total opacity of messages exchanged between AI agents, the authority can demonstrate the existence of contact, but proving anticompetitive content becomes extremely difficult. [2]

The burden of proof rests with the competition authority, which can analyze the chronology of exchanges, metadata, decision logs, and algorithm parameters. [2]

Direct evidentiary instruments are insufficient in isolation; broad corroboration is necessary that rules out plausible alternative hypotheses, including tests involving communication interruption. [2]

If prices remain high while communication is active and fall upon its interruption, this constitutes strong evidence of collusion, but not conclusive proof. [2]

Even a single opaque contact may be sufficient for a concerted practice if followed by the cessation of transmissions, but the difficulty of proving its content remains. [2]

To avoid sanctions, companies must maintain human validation, limit communication between agents, and use confidential internal architectures. The conclusion for consumers is that similar prices raise suspicion, but the difference between legitimate competition and illegal coordination must be determined exclusively through concrete evidence. [2]

Categories:
Economy
Entities:
Tags:
Romania

Sources

  1. Ziarul Financiar Cătălin Suliman, partener, şi Germin Petcu, senior associate: Coluziunea algoritmică în era inteligenţei artificiale. Care este linia roşie în cazul încălcării legilor de concurenţă în deciziile comerciale
  2. Adevărul Cum poate inteligența artificială să mențină prețurile ridicate fără ca firmele să se înțeleagă între ele