Evidence Standards for Algorithmic Cartel Formation

1. Introduction

Algorithmic cartel formation occurs where pricing or other commercially significant algorithms facilitate, implement, reinforce, or potentially create coordinated conduct between competing firms. The central evidentiary problem is that an algorithm can produce parallel or coordinated outcomes without a conventional human meeting, telephone call, email, or explicit agreement.

Under UK competition law, however, algorithmic coordination does not by itself establish a cartel. The Competition and Markets Authority (CMA) must still establish the elements of an infringement of the Chapter I prohibition of the Competition Act 1998, including an agreement, concerted practice, or decision having the requisite anti-competitive object or effect.

The evidentiary question therefore becomes:

When can algorithmic behaviour, data, communications, code, and market outcomes collectively prove an unlawful agreement or concerted practice?

The answer requires distinguishing parallel algorithmic behaviour from evidence demonstrating communication, coordination, commitment, implementation, or conscious adaptation between competitors.

2. Meaning of Algorithmic Cartel Formation

An algorithmic cartel may arise through several mechanisms.

A. Explicit algorithmic coordination

Competitors directly agree that their algorithms should coordinate prices.

For example:

• A and B agree to use the same pricing algorithm;
• competitors exchange algorithmic instructions;
• firms agree on common pricing parameters;
• firms deliberately configure algorithms to implement an agreed pricing strategy.

This is the clearest case.

B. Algorithm-mediated coordination

Human actors communicate an anti-competitive understanding, while algorithms implement it automatically.

For example:

Competitors agree that neither will price below a particular threshold, and each company's algorithm subsequently maintains that threshold.

The algorithm is essentially an implementation mechanism.

C. Algorithmic monitoring and retaliation

Algorithms may continuously monitor competitors and automatically respond to deviations.

For example:

1. Firm A lowers price.
2. Firm B's algorithm detects the reduction.
3. B immediately matches it.
4. A's algorithm detects B's response.
5. A restores the previous price.
6. The process stabilises the market at an elevated price.

Such conduct may be suspicious, but automated matching alone is not necessarily unlawful.

D. Hub-and-spoke algorithmic coordination

A common intermediary may provide pricing software to competing firms.

The critical question is whether the intermediary merely supplies legitimate software or facilitates an understanding among competing users.

E. Autonomous or emergent coordination

The most difficult scenario is where algorithms independently learn that coordinated behaviour produces higher profits without their human operators expressly agreeing to coordinate.

This raises a fundamental evidentiary problem:

Can competition law establish an infringement where the economic outcome looks cartel-like but there is insufficient evidence of human or legally attributable coordination?

Generally, parallel algorithmic conduct alone should not automatically be treated as proof of a cartel.

3. Applicable UK Legal Framework

3.1 Chapter I prohibition

Section 2 of the Competition Act 1998 prohibits agreements between undertakings, decisions by associations of undertakings, and concerted practices which have as their object or effect the prevention, restriction, or distortion of competition and which may affect trade within the UK.

For algorithmic cartels, the important categories are:

• express agreement;
• tacitly evidenced concerted practice;
• information exchange;
• indirect coordination through an intermediary;
• implementation of an anti-competitive agreement by software.

4. The Evidentiary Threshold

Algorithmic cartel cases require a body of evidence, rather than reliance on a single algorithmic output.

The evidence may include:

Direct evidence

• emails;
• messages;
• contracts;
• meeting records;
• instructions to developers;
• internal pricing documents;
• source-code comments;
• configuration files;
• API instructions;
• communications with algorithm vendors.

Digital evidence

• source code;
• version histories;
• Git repositories;
• deployment logs;
• API calls;
• database records;
• server logs;
• audit trails;
• timestamps;
• model-training records;
• algorithmic parameter changes.

Economic evidence

• price convergence;
• abnormal price stability;
• reduced price dispersion;
• supracompetitive pricing;
• unusual responses to competitor movements;
• structural breaks;
• margins;
• bidding patterns;
• output restrictions.

Circumstantial evidence

• opportunity to coordinate;
• knowledge of competitors' behaviour;
• deliberate adoption of matching mechanisms;
• communications preceding algorithm deployment;
• unusual changes following competitor actions;
• absence of plausible independent explanations.

The strongest cases generally involve convergence between documentary, technical, and economic evidence.

5. Algorithmic Parallelism Is Not Automatically a Cartel

A central principle is that similar algorithmic outcomes do not necessarily establish an agreement.

Algorithms may independently produce similar prices because they respond to:

• common demand data;
• common input costs;
• publicly available competitor prices;
• identical market conditions;
• common software;
• common optimisation objectives;
• industry-wide shocks.

Therefore:

Parallel pricing is evidence requiring explanation, not necessarily proof of unlawful coordination.

The CMA would need to distinguish independent rational adaptation from coordination attributable to the undertakings.

6. Evidence of Communication Between Competitors

One of the most powerful forms of evidence is communication showing that competitors understood that their algorithms would coordinate.

Examples include:

"Our pricing engine will follow the competitor's algorithm."

or:

"Configure the system so that we maintain the agreed market price."

Such evidence may transform an otherwise ambiguous algorithmic pattern into evidence of an agreement or concerted practice.

Communications need not necessarily contain the words "cartel" or "fix prices". The surrounding circumstances may establish their meaning.

7. Evidence of Intent and Knowledge

Intent can be particularly important in algorithmic cases.

Relevant questions include:

1. Did management know how the algorithm operated?
2. Did management understand that it responded to competitors?
3. Did the undertaking deliberately choose a coordination-sensitive design?
4. Were safeguards against coordination deliberately removed?
5. Did executives monitor supracompetitive results?
6. Did the undertaking change the algorithm after competitor communications?
7. Did the undertaking benefit from the coordinated outcome?
8. Was the algorithm designed to punish deviations?

However, mere knowledge that an algorithm may produce parallel prices should not automatically equal participation in a cartel.

The evidentiary significance depends upon the total circumstances.

8. Source Code as Competition Evidence

Source code may provide exceptionally powerful evidence.

Investigators could examine whether code contains functions implementing:

• price matching;
• competitor tracking;
• retaliation;
• minimum-price enforcement;
• bid rotation;
• market allocation;
• coordinated output restrictions.

For example, code implementing:

if competitor_price < our_price:       reduce_price()

does not automatically prove an infringement.

But code combined with evidence that competitors agreed to use the same retaliatory mechanism could become highly probative.

9. Algorithm Logs and Digital Forensics

Algorithmic systems generate extensive records.

Investigators may examine:

• input data;
• outputs;
• model versions;
• parameter changes;
• deployment dates;
• system messages;
• API requests;
• competitor-price observations;
• automated responses.

This allows investigators to reconstruct the causal sequence.

For example:

Competitor communication → algorithm modification → synchronized pricing → sustained supracompetitive prices

is much stronger evidence than:

algorithm → similar prices

10. Economic Evidence

Economic evidence can establish whether observed conduct is consistent with coordination.

Important indicators include:

A. Price convergence

Competitors unexpectedly converge on prices despite heterogeneous costs.

B. Reduced price dispersion

Competitive markets often exhibit some price variation. An unexplained reduction may be relevant.

C. Rapid retaliation

Algorithms repeatedly punish deviations from a common price level.

D. Sustained supracompetitive pricing

Prices remain above competitive benchmarks for an extended period.

E. Structural breaks

Pricing behaviour changes sharply following:

• competitor communications;
• algorithm deployment;
• software updates;
• meetings;
• contractual changes.

F. Unusual synchronisation

Prices change simultaneously more frequently than expected under independent decision-making.

Economic evidence is normally most useful when it corroborates documentary or technical evidence.

11. The Importance of Counterfactual Analysis

A major evidentiary question is:

What would the market have looked like if the algorithms had operated independently?

Investigators may construct counterfactual models comparing:

• pre-algorithm pricing;
• post-algorithm pricing;
• algorithm-enabled pricing;
• simulated independent algorithms;
• observed competitor behaviour;
• historical market conditions.

The purpose is to determine whether coordination is a plausible explanation for the observed outcome.

12. Six Important Case Laws

1. T-Mobile Netherlands BV v Raad van bestuur van de Nederlandse Mededingingsautoriteit — C-8/08

This case is fundamental to the evidentiary treatment of concerted practices.

The Court recognised that competition law can address coordination falling short of a formal agreement. The exchange of strategically relevant information can reduce uncertainty about competitors' future conduct and facilitate coordination.

Relevance to algorithms

If competing firms provide information to each other's pricing systems—or deliberately establish mechanisms that reveal future pricing intentions—the evidence may support a concerted-practice theory.

The important principle is that competition law looks beyond formal contractual language to the economic and behavioural substance of coordination.

2. Eturas UAB v Lietuvos Respublikos konkurencijos taryba — C-74/14

Eturas is particularly important for digital and algorithmic competition cases.

An online booking system transmitted a common message to travel agencies concerning restrictions on discounts. The Court considered whether participants could be regarded as engaging in a concerted practice when they were aware of the communication.

Algorithmic significance

The case demonstrates that:

• digital infrastructure can facilitate coordination;
• electronic communications can constitute important evidence;
• participation may depend upon knowledge and response;
• investigators must consider whether undertakings could reasonably be regarded as having participated in the coordinated mechanism.

It is therefore highly relevant to platform-mediated and algorithm-mediated cartel theories.

3. Aalborg Portland A/S v Commission — Joined Cases C-204/00 P etc.

Aalborg Portland is important concerning proof of cartel participation through a body of circumstantial evidence.

The Court recognised that anti-competitive conduct may frequently be concealed, meaning that evidence must sometimes be assessed collectively.

Algorithmic significance

Algorithmic cartels may similarly leave fragmented evidence:

• an email;
• a software instruction;
• an algorithm change;
• a deployment record;
• a pricing pattern.

No individual piece may establish the infringement. Collectively, however, the evidence may form a coherent evidentiary picture.

4. Anic Partecipazioni SpA v Commission — C-49/92 P

Anic is significant for establishing principles concerning participation in concerted anti-competitive conduct.

The case illustrates the importance of determining whether an undertaking participated in a common anti-competitive scheme and whether its conduct demonstrated adherence to that scheme.

Algorithmic significance

Where a company:

• receives competitor information;
• configures its algorithm accordingly;
• deploys the algorithm; and
• continues using it,

the evidentiary question becomes whether those acts demonstrate participation in a common coordinated strategy.

5. Jasper's Limited v Office of Fair Trading

The UK tobacco cases concerning restrictive arrangements demonstrate the importance of examining communications, commercial context, and conduct together when establishing an infringement.

Algorithmic significance

The same methodology can apply to algorithmic evidence.

A regulator should not isolate a pricing graph from:

• internal correspondence;
• commercial negotiations;
• pricing instructions;
• software implementation;
• subsequent conduct.

Algorithmic evidence must be situated within the wider factual record.

6. Tesco Stores Ltd v Competition and Markets Authority

UK competition litigation demonstrates the importance of the evidential and procedural safeguards surrounding CMA enforcement.

Algorithmic significance

Where sophisticated technical evidence is relied upon, undertakings should have a meaningful opportunity to understand and challenge:

• the alleged algorithmic mechanism;
• the data relied upon;
• the methodology;
• the economic inference;
• the causal connection between software behaviour and infringement.

This becomes particularly important where proprietary algorithms make independent verification difficult.

13. Additional Important Authorities

Other authorities that can assist in developing the evidentiary framework include:

Wood Pulp — Joined Cases C-89/85 etc.

Important for distinguishing parallel conduct from unlawful coordination.

Suiker Unie v Commission — Joined Cases 40–48/73

A foundational authority concerning concerted practices and the distinction between independent market conduct and coordinated behaviour.

Hoffmann-La Roche & Co AG v Commission — Case 85/76

Although an Article 102 case rather than a cartel case, it is important for understanding the assessment of commercial conduct through its economic and factual context.

14. Human Attribution in Algorithmic Cartels

A particularly difficult issue is attribution.

Suppose an algorithm independently learns to coordinate prices.

Three possibilities arise:

Scenario 1 — Human instruction

Management expressly instructs the algorithm to coordinate.

Strong basis for attribution.

Scenario 2 — Human awareness and acceptance

Management does not expressly instruct coordination but knows the algorithm is producing coordinated outcomes and deliberately permits the conduct to continue.

Potentially strong circumstantial evidence, depending upon the facts.

Scenario 3 — Genuine autonomous emergence

The algorithm independently discovers a profitable coordination strategy without human knowledge or communication between competitors.

This presents a much more difficult case.

The regulator would need to establish how the conduct can legally be attributed to the undertaking and satisfy the applicable infringement requirements.

15. Evidence of Algorithmic "Knowing Participation"

A useful evidentiary framework is:

Knowledge + Communication + Algorithmic Implementation + Market Effect

For example:

The key is that economic outcomes should not substitute for proof of coordination where the legal infringement requires an agreement or concerted practice.

16. Algorithmic Evidence and the Burden of Proof

Competition authorities must establish infringement according to the applicable standard of proof.

In cartel proceedings, the evidential assessment may involve:

1. direct evidence;
2. circumstantial evidence;
3. economic evidence;
4. digital forensic evidence;
5. expert analysis.

The authority should be able to explain why the evidence supports coordination rather than independent algorithmic optimisation.

This is particularly important because machine-learning systems can produce complex outcomes that even their designers may find difficult to explain.

17. Explainability as an Evidentiary Requirement

Algorithmic cartel cases raise a new question:

Can a regulator establish an infringement from an algorithm that cannot explain why it produced a particular price?

The answer should generally be cautious.

An opaque algorithm does not itself prove illegality.

Investigators may therefore need:

• model documentation;
• training data;
• parameter histories;
• audit logs;
• decision traces;
• testing records;
• developer documentation;
• deployment records.

Where explainability is impossible, economic evidence may become more important, but it should still be connected to the legal elements of the infringement.

18. Evidence Preservation

Algorithmic investigations create unusual preservation challenges.

Important evidence may disappear through:

• model retraining;
• automatic deletion;
• software updates;
• version replacement;
• cloud migration;
• changing parameters;
• ephemeral API logs.

Accordingly, competition investigations may need preservation of:

• source-code versions;
• model checkpoints;
• configuration files;
• logs;
• databases;
• API records;
• deployment histories;
• communications between developers and commercial teams.

A cartel case involving continuously learning systems may therefore require temporal reconstruction.

19. Evidentiary Problems with Machine Learning

Machine-learning algorithms create several special problems.

A. Non-determinism

The same system may produce different results depending on data and model state.

B. Continuous learning

The algorithm may change after deployment.

C. Black-box decision-making

The precise reason for an output may be difficult to identify.

D. Data dependence

Competitor prices may be incorporated into training or real-time decision systems.

E. Emergent behaviour

The system may develop strategies not expressly programmed by humans.

These characteristics make conventional cartel evidence increasingly difficult to apply mechanically.

20. Recommended Evidentiary Test

A robust UK framework could examine five stages:

Stage 1 — Identify the algorithmic conduct

What exactly did the algorithm do?

Stage 2 — Establish the human or undertaking connection

Who designed, selected, configured, deployed, or controlled it?

Stage 3 — Establish communication or coordination

Was there:

• direct communication;
• information exchange;
• common instructions;
• intermediary facilitation;
• knowledge of a coordination mechanism?

Stage 4 — Establish implementation

Did the algorithm actually implement the alleged arrangement?

Stage 5 — Test alternative explanations

Could the observed outcome reasonably result from:

• independent optimisation;
• common market conditions;
• legitimate price matching;
• common software;
• demand shocks?

This final stage is essential.

21. Difference Between Legitimate Algorithmic Pricing and Cartelisation

The same technical behaviour can therefore have very different competition-law implications depending upon the underlying evidence.

22. Key Legal Principle

The most important principle is:

An algorithm is evidence of conduct; it is not, by itself, evidence of an unlawful agreement.

A regulator must connect algorithmic behaviour to the legal elements of the competition infringement.

Consequently, the strongest algorithmic cartel case will normally combine:

Documentary evidence
↓
Digital/technical evidence
↓
Algorithmic behaviour
↓
Economic evidence
↓
Alternative-explanation analysis
↓
Proof of coordination

23. Conclusion

Evidence standards for algorithmic cartel formation require competition authorities to adapt traditional cartel-investigation techniques to technologically complex systems without lowering the substantive evidentiary requirements of competition law.

The most probative evidence is likely to arise from the interaction between human conduct and algorithmic implementation: communications between competitors, algorithm specifications, source code, configuration changes, API records, deployment logs, internal documents and economic evidence.

The difficult boundary is autonomous algorithmic coordination. Parallel pricing, even when persistent and economically suspicious, does not necessarily establish an agreement or concerted practice. The authority must distinguish independent algorithmic optimisation from coordination attributable to competing undertakings.

The emerging evidentiary model can therefore be expressed as:

Algorithmic outcome + communication/knowledge + implementation + contextual evidence + economic corroboration + rejection of plausible independent explanations = strong cartel evidence.

For UK competition law, the central challenge is not simply determining what the algorithm did, but establishing why it did it, who caused or accepted that behaviour, whether competitors coordinated, and whether the resulting conduct satisfies the legal test for an anti-competitive agreement or concerted practice.