Competition Law and Artificial Intelligence in Cartel Detection

1. Introduction

Artificial Intelligence (AI) is increasingly relevant to competition-law enforcement, particularly in the detection of cartels. Traditional cartel investigations often begin with whistleblower complaints, leniency applications, dawn raids, suspicious communications, or evidence discovered during another investigation. AI can supplement these methods by identifying statistical patterns, bidding anomalies, parallel pricing, communication networks, suspicious procurement behaviour and other indicators of possible coordination.

AI does not itself establish that a cartel exists. A machine-learning system may identify a pattern that is consistent with collusion, but competition authorities still need to establish the legally relevant elements of an infringement through evidence and appropriate legal procedures.

The relationship between AI and cartel detection therefore has two dimensions:

AI as an enforcement tool used by competition authorities to detect possible cartels; and

AI as a source of new cartel risks, because competing firms may use algorithms to coordinate prices or market behaviour.

The second dimension is especially important because algorithmic coordination can make traditional concepts such as "communication" and "agreement" more difficult to apply.

2. What Is a Cartel?

A cartel generally involves competitors coordinating rather than competing independently.

Typical cartel conduct includes:

price fixing;

market sharing;

customer allocation;

output restriction;

bid rigging;

exchange of competitively sensitive information;

coordination concerning future prices or commercial strategies.

Under EU competition law, Article 101 TFEU prohibits agreements, decisions and concerted practices that have as their object or effect the prevention, restriction or distortion of competition.

Under national competition laws, similar prohibitions exist.

AI does not change the fundamental legal prohibition. Instead, it changes how potentially unlawful coordination can be discovered and how evidence can be analysed.

3. Why AI Is Useful for Cartel Detection

Cartels frequently generate enormous quantities of data.

For example, a public-procurement investigation may involve:

thousands of tenders;

thousands of bids;

bid prices;

winning suppliers;

losing suppliers;

bid timing;

geographic locations;

contract values;

product specifications;

supplier relationships.

Human investigators may have difficulty identifying relationships within such datasets.

AI can process these datasets and identify anomalies such as:

1. Unusual price similarities

Competitors repeatedly submit nearly identical prices.

2. Bid rotation

Different firms repeatedly win tenders according to a suspicious pattern.

3. Geographic allocation

Company A repeatedly wins contracts in one region while Company B wins another.

4. Suspicious losing bids

Certain firms consistently submit bids that are substantially higher than the eventual winning bid.

5. Abrupt changes in behaviour

Several competitors change pricing behaviour simultaneously without an obvious economic explanation.

6. Communication networks

AI-assisted analysis can identify unusual connections among employees, companies and communications.

4. AI Does Not Replace Legal Proof

This distinction is fundamental.

An AI system might conclude:

"The probability of coordinated bidding is high."

That is not equivalent to a legal finding of a cartel.

The authority must still establish the relevant legal standard using admissible evidence.

An algorithm may therefore function as:

screening tool → investigative lead → evidence collection → legal assessment → enforcement decision.

It should not normally be:

algorithmic prediction → automatic finding of infringement.

5. AI Techniques Used in Cartel Detection

A. Supervised machine learning

A system can be trained using historical examples of:

known cartels;

legitimate competitive markets;

suspicious procurement patterns.

It can then classify new datasets according to their similarity to known patterns.

Potential variables include:

price variance;

bid frequency;

winning ratios;

bid timing;

market shares;

tender participation;

geographic distribution.

6. Unsupervised Learning

Unsupervised algorithms search for unusual patterns without requiring investigators to identify beforehand what the cartel looks like.

Examples include:

clustering;

anomaly detection;

network analysis;

dimensionality reduction.

This can be useful where investigators do not know the precise form of coordination.

7. Natural Language Processing

AI can analyse:

emails;

chat messages;

internal documents;

meeting records;

tender correspondence.

Natural-language-processing systems can identify:

discussions concerning future prices;

references to competitors;

suspicious terminology;

communications networks;

changes in communication frequency.

However, context remains essential. A reference to a competitor's price is not automatically evidence of a cartel.

8. Network Analysis

Cartels often involve relationships between several firms and individuals.

AI can create network maps showing:

common directors;

communication relationships;

shared advisers;

repeated contacts;

common addresses;

supplier relationships;

bidding relationships.

Investigators can then examine whether those relationships have legitimate explanations or support a theory of coordination.

9. AI in Public Procurement

Public procurement is particularly suitable for algorithmic cartel screening.

Suppose six companies participate in 2,000 tenders.

AI could detect:

None of these indicators proves an infringement individually.

Their significance increases when several independent indicators occur together and cannot readily be explained by legitimate market conditions.

10. Important Case Law

AI-specific cartel jurisprudence is still developing. Consequently, the most important cases are traditional cartel cases whose legal principles can be applied to algorithmic detection and algorithmic coordination.

Case 1 — Wood Pulp, Joined Cases 89/85 and Others

A. Background

The Wood Pulp litigation concerned parallel pricing behaviour among producers.

The European Commission attempted to establish coordinated behaviour based partly on patterns of pricing.

B. Principle

The Court emphasized that parallel conduct alone does not automatically establish a concerted practice.

There must be sufficient evidence of coordination or other circumstances demonstrating that the conduct cannot adequately be explained by normal competitive conditions.

C. Importance for AI

This is one of the most important principles for algorithmic cartel detection.

An AI system might discover:

"Competitors have followed extremely similar pricing patterns."

That is an investigative lead, not necessarily proof of collusion.

AI must therefore distinguish between:

parallel behaviour caused by market conditions

and

parallel behaviour resulting from coordination.

11. Case 2 — Suiker Unie v Commission, Joined Cases 40–48/73 and Others

This is a foundational EU cartel case.

The Court considered the concept of a concerted practice.

Legal principle

Competition law prohibits competitors from knowingly substituting practical cooperation for the risks of competition.

The concept of concerted practice can therefore extend beyond a formally concluded agreement.

AI significance

This principle is particularly important where algorithms are involved.

Two competitors might never sign a conventional written cartel agreement. Instead, their systems could respond predictably to market information.

AI-based detection therefore needs to examine:

communications;

data sharing;

algorithm design;

pricing instructions;

human decisions;

implementation practices.

12. Case 3 — Anic Partecipazioni, C-49/92 P

The Anic case is important for understanding the concept of concerted practices.

The Court explained that participation in a concerted practice can involve behaviour that reduces the uncertainty that should normally exist between competitors.

AI relevance

AI systems can potentially reduce competitive uncertainty by:

predicting competitors' responses;

monitoring competitors' prices;

automatically adjusting prices;

exchanging market information;

identifying competitors' likely reactions.

Therefore, investigators may need to determine whether the technology has facilitated an exchange of strategically significant information.

13. Case 4 — T-Mobile Netherlands, C-8/08

Background

The case concerned an exchange of information between competitors.

The Court adopted a strict approach to exchanges of strategically sensitive information.

Principle

An exchange of information capable of removing uncertainty concerning competitors' intended future conduct can raise serious Article 101 concerns.

AI significance

AI can make information exchange significantly more powerful.

For example, competitors might use automated systems to exchange:

expected prices;

capacity;

inventory;

production plans;

future commercial strategies.

An AI detection system could therefore search for both direct communication and technical mechanisms facilitating information exchange.

14. Case 5 — Eturas, C-74/14

This case is especially interesting for algorithmic competition law.

Background

Eturas operated an online booking platform used by travel agencies.

A centralized system message imposed restrictions concerning discounts.

The question was whether participating businesses could be regarded as participating in a concerted practice.

Principle

The Court considered circumstances in which undertakings using a common electronic platform could be attributed knowledge of a competitively restrictive message transmitted through that platform.

Importance for AI

Eturas demonstrates that digital infrastructure can become a mechanism through which competitors coordinate.

This has major relevance for:

online marketplaces;

pricing platforms;

reservation systems;

automated bidding;

common software;

AI pricing tools.

The case shows why competition authorities increasingly need to examine the architecture through which competitive decisions are made.

15. Case 6 — AC-Treuhand, C-194/14 P

AC-Treuhand concerned a consultancy that facilitated cartel activity even though it was not itself a traditional competitor in the affected product market.

Principle

EU competition law can reach undertakings that intentionally facilitate cartel conduct.

AI relevance

The modern equivalent could potentially involve:

algorithm providers;

pricing-software suppliers;

data intermediaries;

digital platforms;

consultants;

technology providers.

The legal question would depend upon the precise facts, including knowledge, contribution and participation.

An AI developer does not become a cartel participant merely because competitors use its software. But deliberate facilitation can create serious competition-law risks.

16. Case 7 — ICAP, Case T-180/15

ICAP concerned the role of a financial intermediary in facilitating anti-competitive conduct.

The case is useful because it demonstrates that cartel investigations can extend beyond the immediate competitors.

AI relevance

A similar analytical problem can arise where:

multiple competitors use the same pricing service;

an intermediary supplies competitively sensitive information;

a software platform communicates market information;

a common algorithm determines prices.

The crucial question remains the undertaking's role and knowledge rather than simply the existence of common technology.

17. Case 8 — Infineon Technologies and Others, Joined Cases C-99/17 P and Others

This litigation concerned the automotive chip cartel and illustrates the evidentiary challenges associated with complex cartel investigations.

Relevance to AI

AI can help authorities analyse:

large numbers of communications;

pricing records;

customer allocation;

meeting patterns;

temporal relationships;

transaction data.

But the legal evaluation still requires evidence establishing participation and the relevant infringement.

AI therefore assists evidence discovery, rather than replacing the judicial assessment of evidence.

18. Algorithmic Collusion

A particularly important distinction is between:

A. AI-assisted cartel detection

Authorities use AI to find existing cartel behaviour.

B. AI-enabled collusion

Companies use AI systems that may facilitate coordination between competitors.

These are legally different questions.

19. Explicit Algorithmic Collusion

The clearest situation would be where competing firms deliberately program their systems to coordinate.

For example:

Company A and Company B agree to use a common pricing algorithm;

the algorithm is programmed to maintain an agreed price;

both companies implement the arrangement.

The technology does not eliminate the underlying agreement.

The cartel is still the agreement or concerted practice; the algorithm is simply the mechanism through which it is implemented.

20. Tacit Algorithmic Coordination

A more difficult situation arises when companies independently use algorithms that observe competitors and respond to their pricing.

For example:

Firm A's algorithm observes Firm B's price.

Firm A automatically raises its price.

Firm B's algorithm observes A.

Firm B raises its price.

The process repeats.

Prices may converge without direct communication.

The critical legal question becomes whether the behaviour constitutes a prohibited concerted practice or merely independent adaptation to market conditions.

Independent parallel conduct is not automatically a cartel.

This is why traditional principles concerning concerted practices remain important.

21. Predictable Algorithmic Coordination

Another possibility involves competitors using algorithms designed to react predictably to one another.

Suppose competing firms know that:

whenever another firm raises prices, the algorithm automatically follows.

Even without a conventional communication channel, the algorithms can create highly stable coordination.

Competition authorities may therefore need to examine:

who designed the algorithm;

what instructions were given;

what data it receives;

whether competitors use identical systems;

whether the system was intentionally configured to facilitate coordination;

whether there was communication between competitors;

whether human decision-makers understood the system's operation.

22. AI and Bid Rigging

Bid rigging is particularly suitable for AI detection.

An authority could develop an algorithm that calculates a bid-rotation indicator based on:

frequency of winning;

tender sequence;

bid differences;

market shares;

geographical distribution;

tender participation;

withdrawal patterns.

For example:

Such a pattern might trigger investigation.

But legitimate explanations must be considered, such as:

geographic specialization;

capacity constraints;

different production costs;

technical qualifications;

customer preferences.

23. AI and Price-Fixing Detection

AI can compare prices across:

competitors;

locations;

time periods;

products;

customer groups.

It can identify unusually synchronized movements.

For example:

Price movement A: +5%
Price movement B: +5.1%
Price movement C: +4.9%

Repeated synchronization may be suspicious.

But synchronized prices may also arise because competitors face:

identical input costs;

tax changes;

commodity prices;

common suppliers;

exchange-rate movements;

regulation.

Consequently, AI-generated anomalies require economic investigation.

24. AI and Communication Evidence

AI can help investigators process enormous document collections.

For example, an authority may possess:

10 million emails;

2 million instant messages;

500,000 documents;

several years of pricing data.

Natural-language processing can identify documents containing combinations of:

competitor names;

future price references;

market-allocation language;

tender information;

customer-allocation references.

Investigators can then examine the relevant documents manually.

25. Explainability Problem

One of the most significant legal problems is explainability.

A machine-learning system may identify a suspicious pattern but fail to explain why it considers the pattern suspicious.

A competition authority therefore needs to know:

what data the model used;

what variables were important;

how the model was trained;

whether the data were reliable;

whether false positives were tested;

whether the model has systematic biases;

whether the result can be independently reproduced.

This is particularly important where AI-generated evidence influences enforcement action.

26. False Positives

AI cartel detection can generate false positives.

For example, three competing firms may charge nearly identical prices because:

they purchase the same raw material;

they operate under the same regulation;

their products are homogeneous;

customers can easily compare prices;

transportation costs are similar.

The algorithm could flag the pattern even though no cartel exists.

Therefore:

statistical suspicion is not equivalent to legal proof.

27. False Negatives

The opposite problem also exists.

Sophisticated cartels may deliberately avoid obvious patterns.

Participants may:

vary prices slightly;

rotate winners irregularly;

communicate indirectly;

use intermediaries;

conceal communications;

manipulate data.

AI systems trained on conventional cartels may fail to detect novel forms of coordination.

28. Data Protection and Procedural Issues

AI cartel detection also creates legal issues concerning:

confidentiality;

personal data;

employee communications;

privileged material;

business secrets;

proportionality;

investigative powers.

Competition authorities must ensure that data processing is consistent with applicable legal requirements.

29. Evidentiary Chain

A sound AI-assisted investigation should preserve a chain such as:

Raw data → data validation → algorithmic analysis → anomaly → human investigation → corroborating evidence → legal assessment → decision.

The weakest model would be:

Raw data → AI score → cartel finding.

The latter creates serious problems concerning procedural fairness and reliability.

30. Role of Human Investigators

Human investigators remain essential because they can understand:

market structure;

industry practices;

contractual arrangements;

legitimate business explanations;

technological architecture;

communications;

economic conditions.

AI is particularly powerful at finding patterns that humans may otherwise overlook.

The best institutional model is therefore generally human-led, AI-assisted enforcement.

31. Leniency and AI Detection

AI does not eliminate the importance of leniency programmes.

A cartel may be detected through:

AI screening + leniency evidence + documentary evidence + economic analysis.

For example, AI might identify suspicious bid rotation, after which investigators discover an internal document confirming an agreement.

The documentary evidence, rather than the AI score alone, could become central to the legal case.

32. AI and the Standard of Proof

The legal standard applicable to cartel enforcement remains crucial.

Authorities should establish:

the relevant conduct;

the identity of participating undertakings;

the relevant agreement or concerted practice;

the infringement period;

the relevant market circumstances;

the evidence connecting the undertaking to the conduct.

AI can assist with steps 1, 4 and 5, but it cannot automatically satisfy every legal element.

33. Competition Law Challenges Created by AI

AI therefore creates at least six major challenges:

1. Detection challenge

Authorities must identify sophisticated algorithmic coordination.

2. Attribution challenge

Authorities must determine who is legally responsible for an algorithm's conduct.

3. Intent challenge

Investigators may need to determine whether coordination was deliberately facilitated.

4. Evidence challenge

AI outputs must be distinguished from legally sufficient evidence.

5. Explainability challenge

Authorities must understand and explain how an algorithm generated its findings.

6. Dynamic-market challenge

Algorithms can change their behaviour faster than conventional investigations.

34. Practical Investigation Model

A competition authority could adopt the following model:

Phase 1 — Data collection

Collect:

tender data;

transaction data;

prices;

communications;

market shares;

customer information.

Phase 2 — AI screening

Run:

anomaly detection;

clustering;

network analysis;

NLP;

price correlation analysis.

Phase 3 — Risk identification

Identify:

suspicious firms;

suspicious transactions;

suspicious communications;

suspicious tender patterns.

Phase 4 — Human investigation

Investigators examine the underlying evidence.

Phase 5 — Economic analysis

Determine whether the observed conduct has legitimate economic explanations.

Phase 6 — Legal analysis

Apply Article 101 or the applicable national competition-law provision.

Phase 7 — Enforcement

Only after sufficient evidence is obtained should the authority proceed toward an infringement decision.

35. Key Case-Law Principles

36. Conclusion

Artificial Intelligence can significantly strengthen cartel detection by allowing competition authorities to analyse enormous datasets that would be difficult to investigate manually. Its greatest applications are likely to involve bid-rigging detection, price-pattern analysis, communication analysis, network analysis, anomaly detection and identification of suspicious information exchanges.

However, AI should be treated principally as an investigative and evidentiary-support mechanism, not as an autonomous adjudicator.

The central legal distinction remains:

An AI-detected anomaly is evidence requiring investigation; it is not itself proof of a cartel.

The cases of Wood Pulp, Suiker Unie, Anic, T-Mobile Netherlands, Eturas, AC-Treuhand, ICAP and Infineon provide the doctrinal foundation for dealing with this problem. They show that cartel law focuses not merely on identical outcomes but on whether competitors have substituted coordination for independent competition.

The future of cartel enforcement is therefore likely to involve a combination of:

AI-based screening + economic analysis + digital forensics + documentary evidence + human investigation + established competition-law principles.

At the same time, competition authorities must ensure that AI systems are sufficiently transparent, reliable and carefully validated so that statistical correlation is not mistakenly treated as proof of unlawful coordination.