Adversarial Hearing Rights in Algorithmic Antitrust Proceedings — Europe

1. Introduction

Adversarial hearing rights in algorithmic antitrust proceedings concern the procedural rights of an undertaking when a European competition authority relies upon algorithms, automated decision systems, machine-learning models, large datasets, ranking systems, pricing algorithms, ad-tech systems or other computational evidence to investigate or establish an infringement.

The central procedural principle is:

An undertaking must have a genuine opportunity to understand and challenge the evidence and reasoning used against it before an adverse competition decision is adopted.

This becomes more difficult when the authority's case depends upon:

algorithmic pricing;

automated ranking;

recommendation systems;

ad auctions;

machine-learning models;

algorithmic collusion theories;

large-scale transaction datasets;

technical logs;

statistical models;

data-room evidence;

confidential source code;

expert economic models.

EU competition procedure already contains strong rights of defence. The challenge is adapting those rights to evidence that may be technically complex, confidential, enormous in volume or difficult to explain.

The European Commission expressly recognises that access to the file is a procedural guarantee protecting equality of arms and rights of defence, while the right to be heard includes written submissions and, in Article 7 antitrust proceedings, an oral hearing before the independent Hearing Officer. (Competition Policy)

2. Legal Foundation

The principal sources are:

Article 41(2)(a), EU Charter

Right to be heard before an individual adverse measure is taken.

Article 41(2)(b), EU Charter

Right of access to one's file, subject to legitimate confidentiality interests.

Article 47, EU Charter

Right to an effective remedy and fair hearing.

Article 48, EU Charter

Presumption of innocence and rights of defence in proceedings involving penalties.

Article 101 TFEU

Prohibition of restrictive agreements and concerted practices.

Article 102 TFEU

Prohibition of abuse of dominance.

Regulation 1/2003

Provides the Commission's investigative and enforcement framework, including Article 27 rights of defence.

Regulation 773/2004

Contains detailed procedural rules concerning Commission antitrust proceedings.

The Commission confirms that parties receiving a Statement of Objections are entitled to access the investigative file and, in Article 7 proceedings, an oral hearing conducted by an independent Hearing Officer. (Competition Policy)

3. Meaning of an Adversarial Hearing

An adversarial hearing is not merely an opportunity to speak.

It requires a meaningful opportunity to:

know the case against the undertaking;

understand the evidence;

identify the methodology;

challenge factual assumptions;

challenge economic models;

challenge algorithmic evidence;

present counter-evidence;

question the Commission's reasoning;

explain legitimate business justifications;

respond before the final decision.

The Commission's oral hearings occur after written replies to the Statement of Objections and allow accused companies to present their views to Commission officials, the Legal Service and relevant national competition authorities. (Competition Policy)

4. Why Algorithms Create a Special Procedural Problem

Traditional evidence may consist of:

emails;

contracts;

invoices;

witness statements;

meeting records.

Algorithmic cases can instead involve:

millions of transactions;

source-code fragments;

model outputs;

feature variables;

training datasets;

statistical correlations;

automated pricing decisions;

A/B experiments;

ranking signals;

API records;

telemetry;

machine-learning predictions.

The undertaking may therefore face a problem:

How can it meaningfully defend itself if it knows the authority's conclusion but cannot understand how the algorithmic evidence produced that conclusion?

This creates a distinction between:

Formal disclosure

Giving the party documents.

and

Effective disclosure

Giving the party enough information to understand and challenge the evidence.

The second concept is central to algorithmic proceedings.

5. Case Law 1 — Solvay v Commission, C-110/10 P

Case

Solvay SA v European Commission, C-110/10 P

Court: CJEU
Judgment: 25 October 2011

Principle

The CJEU confirmed that access to the Commission's investigation file is an important component of the rights of defence.

The undertaking must have an opportunity to examine documents that may be relevant to its defence. (Infocuria)

Algorithmic application

Suppose the Commission claims:

"The defendant's algorithm excluded competitors."

The defendant should be able to examine relevant evidence concerning:

algorithmic outputs;

relevant comparator data;

counterfactuals;

internal testing;

evidence supporting the foreclosure theory.

The precise disclosure depends upon confidentiality rules, but the party cannot be deprived of material evidence simply because it is technically complex.

Core principle

Access to relevant evidence is an essential component of an effective defence.

6. Case Law 2 — Aalborg Portland v Commission, Joined Cases C-204/00 P and Others

Facts

The cases concerned competition proceedings and access to evidence.

Principle

The CJEU established that the Commission must provide access to evidence relevant to the defence, including exculpatory material, subject to legitimate confidentiality restrictions.

The rights of defence therefore involve more than access to evidence supporting the Commission's case. (Infocuria)

Algorithmic significance

This becomes especially important where an algorithmic investigation produces mixed results.

For example:

A party should have an opportunity to identify and rely upon relevant exculpatory evidence.

7. Case Law 3 — Corus UK v Commission, C-199/99 P

Principle

The CJEU explained the purpose of access to the file:

it enables the undertaking to acquaint itself with the evidence so that it can effectively comment on the Commission's conclusions.

The Commission's evidence must therefore be accessible in a manner consistent with effective defence rights. (Infocuria)

Algorithmic application

An algorithmic proceeding should not operate on:

"The model says X, therefore X is established."

The defendant should be able to challenge:

input variables;

data quality;

methodology;

model specification;

assumptions;

statistical significance;

causal inference;

alternative explanations.

8. Case Law 4 — Intel v Commission, C-413/14 P

Case

Intel Corporation Inc. v Commission, C-413/14 P

Court: CJEU
Judgment: 6 September 2017

Importance

This is one of the most important modern rights-of-defence cases in EU competition law.

The CJEU held that where the Commission uses the as-efficient-competitor test (AEC test) or relies on evidence concerning the capability of rebates to foreclose equally efficient competitors, it must properly assess that evidence.

The Court required a deeper examination of the foreclosure effects where the undertaking had challenged the Commission's analysis.

Algorithmic relevance

The case establishes an important methodological principle:

An economic model cannot substitute for a properly reasoned and contestable assessment of competitive effects.

In algorithmic proceedings, the equivalent problem could involve:

algorithmic foreclosure models;

price-effect models;

diversion analysis;

ranking simulations;

auction simulations.

If the undertaking produces credible counter-analysis, the authority must deal with the relevant evidence and reasoning.

9. Case Law 5 — Deutsche Bahn v Commission, T-229/94

Case

Deutsche Bahn AG v Commission, T-229/94

Principle

The case concerned competition proceedings, fines, access to the file and rights of defence.

The case is part of the longstanding EU jurisprudence recognising the relationship between:

access to file → ability to defend → legality of Commission decision.

The CJEU/General Court case law has consistently treated access to relevant evidence as a procedural guarantee rather than merely an administrative convenience. (curia)

Algorithmic relevance

If the Commission relies on:

algorithmic logs;

digital communications;

automated pricing records;

the fact that those records are technically difficult to interpret does not remove their relevance to the defence.

10. Case Law 6 — BEH v Commission / Data-Room Proceedings

Recent EU litigation is especially relevant to data rooms and confidential digital evidence.

The Court has reaffirmed that both:

the right to be heard; and

the right of access to the file

are fundamental components of the rights of defence.

The Court has also recognised that these rights must be balanced against protection of business secrets and confidential information. (Curia)

Importance for algorithms

This is highly relevant where an authority has:

confidential source code;

commercially sensitive datasets;

third-party algorithmic information;

confidential auction information.

The solution does not necessarily have to be unrestricted disclosure.

Possible mechanisms include:

data rooms;

confidentiality rings;

restricted expert access;

non-confidential summaries;

redacted datasets;

controlled technical disclosure.

11. Case Law 7 — Brugg Kabel v Commission, T-441/14

Principle

The General Court reiterated that access to the file is intended to enable parties to understand the evidence underlying the Commission's objections.

The file may contain both:

inculpatory evidence;

exculpatory evidence.

The Commission must respect the rights of defence when using evidence obtained during the investigation. (Infocuria)

Algorithmic significance

In a machine-learning case, the relevant "file" could include:

model-validation results;

alternative specifications;

datasets;

rejected models;

statistical tests;

internal analyses.

Whether every underlying technical artefact must be disclosed depends upon relevance and confidentiality, but the defence must receive enough material to challenge the evidence actually relied upon.

12. Case Law 8 — Aalborg Portland and the "Exculpatory Evidence" Principle

A recurring rule from Aalborg Portland is particularly important.

Evidence potentially useful to the defence cannot simply be ignored because the Commission does not rely upon it.

The procedural obligation extends to relevant evidence that could assist the undertaking's defence, subject to the established confidentiality limitations. (Infocuria)

Algorithmic example

Suppose an algorithmic pricing investigation produces:

Model A: suggests coordination.

Model B: finds independent parallel pricing.

Model C: finds pricing explained by common cost shocks.

If the Commission relies upon Model A, the existence and relevance of Models B and C can become important to the defence.

13. Case Law 9 — Toshiba v Commission, T-113/07

The EU courts have also addressed the temporal and procedural boundaries of access to the file.

Evidence submitted by other parties during the administrative procedure does not automatically become available at every stage simply because it exists somewhere within the broader investigative process.

The relevant question is whether the evidence forms part of the material necessary for effective defence at the relevant procedural stage. (Infocuria)

Algorithmic importance

Large computational investigations may continuously generate:

new datasets;

revised models;

new simulations;

new expert reports.

The authority cannot necessarily treat every internal development as automatically disclosable, but if it becomes part of the evidentiary basis for the final infringement finding, procedural fairness becomes critical.

14. Case Law 10 — Mitsubishi Electric v Commission, T-133/07

This case reinforces the distinction between:

the Commission's internal material; and

evidence forming part of the adversarial administrative file.

The undertaking does not have an unrestricted right to every internal Commission document.

However, it must have access to relevant evidence necessary for its defence. (Infocuria)

Algorithmic significance

A defendant cannot ordinarily demand the Commission's:

internal emails;

deliberative notes;

confidential internal assessments;

merely because an algorithmic case is complicated.

But the defendant can challenge the evidentiary basis actually relied upon for the infringement decision.

15. Algorithmic Evidence: What Should Be Disclosed?

There is no single rule requiring every competition authority to hand over source code.

Instead, disclosure should focus on material necessary for effective defence.

Potentially relevant material includes:

A. Model description

purpose;

methodology;

variables;

assumptions;

outputs.

B. Data description

period;

population;

sampling;

relevant markets;

data cleaning.

C. Statistical methodology

regression specification;

confidence intervals;

robustness testing;

sensitivity analysis.

D. Counterfactual

What would competition have looked like without the allegedly abusive conduct?

E. Validation

model accuracy;

error rates;

alternative specifications.

F. Relevant outputs

The actual results relied upon by the authority.

16. Source Code Is Not Automatically the Same as Effective Disclosure

An authority might theoretically provide thousands of lines of source code.

That does not necessarily mean the party has received meaningful procedural disclosure.

For example:

Source code + no explanation + no relevant data + no model assumptions

may be practically useless to a defence team.

Conversely:

Detailed model description + relevant outputs + methodology + controlled expert access

may sometimes provide a more effective mechanism.

Therefore:

Procedural fairness concerns the practical ability to challenge the evidence, not merely the formal transfer of files.

17. Black-Box Algorithms

A black-box algorithm is particularly problematic where:

the decision logic is opaque;

the model is proprietary;

the authority itself cannot fully explain the model;

machine learning generates complex predictions.

Competition law does not necessarily prohibit reliance on complex evidence.

But a serious procedural problem arises if:

the authority relies on an inference that the undertaking cannot meaningfully test or challenge.

The more important the algorithmic output is to the infringement finding, the more important effective explanation becomes.

18. Algorithmic Collusion Cases

Algorithmic antitrust investigations can involve alleged:

price coordination;

information exchange;

hub-and-spoke coordination;

signalling;

automated adaptation;

parallel algorithmic behaviour.

A company may argue:

"Our algorithm independently responded to market conditions."

The authority may argue:

"The algorithm was designed or configured in a way that facilitated coordination."

The adversarial process must allow the undertaking to challenge:

source of the coordination;

human involvement;

algorithm design;

communication channels;

common data;

market conditions;

alternative explanations.

19. Human Agency vs Algorithmic Output

A crucial question is:

Who made the competitive decision?

An algorithm may:

execute instructions;

learn from historical data;

optimise prices;

respond to competitors.

But an algorithm's output does not automatically establish an unlawful agreement.

The authority must establish the elements of the applicable infringement.

Therefore:

Algorithmic similarity ≠ automatically an antitrust agreement.

Similarly:

Automated price increases ≠ automatically collusion.

The evidence must be assessed under Article 101 or Article 102 and the relevant case-law standards.

20. Algorithmic Dominance Cases

In Article 102 proceedings, algorithms may be used to analyse:

ranking;

self-preferencing;

foreclosure;

tying;

discriminatory access;

exclusionary pricing.

For example:

Search algorithm → preferential ranking → traffic diversion → reduced rival traffic → foreclosure.

The defendant must be able to challenge each causal link.

This is particularly important following the Google Shopping jurisprudence, where the legal assessment involved preferential positioning and display of Google's comparison-shopping service.

21. Economic Models and Adversarial Rights

Suppose the Commission's case depends on an economic model.

The undertaking should be able to challenge:

Inputs

Were the correct data used?

Assumptions

Are the assumptions realistic?

Model selection

Why was this model chosen?

Counterfactual

What is the relevant alternative scenario?

Causation

Does correlation establish the claimed competitive effect?

Robustness

Does the conclusion survive alternative specifications?

This is where Intel is particularly important.

22. Confidentiality vs Rights of Defence

Algorithmic antitrust proceedings frequently involve genuine confidentiality.

Examples:

source code;

trade secrets;

proprietary algorithms;

confidential customer data;

third-party business information.

The EU framework therefore does not provide an unlimited right of access.

The Commission recognises that confidential information can be protected, while the Hearing Officer may resolve disputes concerning whether information must be disclosed for effective defence. (Competition Policy)

23. Data Rooms

A data room can reconcile competing interests.

For example:

Competition authority

→ provides sensitive algorithmic evidence

↓

Independent external lawyers/economists

↓

Restricted data room

↓

Defence submissions

The defendant may therefore obtain meaningful technical understanding without receiving unrestricted access to another company's source code.

Recent EU case law specifically discusses data-room procedures as a mechanism for balancing rights of defence against protection of business secrets. (Curia)

24. Equality of Arms

The principle of equality of arms is particularly important.

An authority may possess:

thousands of experts;

economists;

data scientists;

programmers;

investigators.

The defendant may face the authority with a relatively small legal team.

Procedural fairness therefore requires a meaningful opportunity to understand the technical case.

This does not mean equal resources.

It means that the defendant must have a genuine opportunity to respond.

25. Right to Challenge Expert Evidence

Where expert economic or computational analysis is central, the defence may challenge:

qualifications;

methodology;

assumptions;

data;

statistical techniques;

interpretation;

causal inference.

A hearing is especially valuable because written submissions can be supplemented by direct technical explanation.

The Commission's Hearing Officer can allow discussion of particular topics and deal with questions during the hearing. (Competition Policy)

26. Oral Hearing

In ordinary Article 7 antitrust proceedings, an undertaking receiving a Statement of Objections can request an oral hearing.

The hearing is:

non-public;

organised by the Hearing Officer;

conducted after written replies;

intended to allow parties to present their views.

The Hearing Officer also determines procedural questions concerning participation, documents, questions and post-hearing submissions. (Competition Policy)

27. Important Limitation: Commitment Proceedings

Adversarial hearing rights differ depending upon the procedural route.

Under Article 9 Regulation 1/2003 commitment proceedings, there is no equivalent right to request an oral hearing under Article 12 of Regulation 1/2003, and there is no formal access-to-file procedure equivalent to Article 7 proceedings. (Competition Policy)

This distinction is crucial.

Article 7

Infringement procedure

→ Statement of Objections

→ access to file

→ written response

→ oral hearing available upon request.

Article 9

Commitment procedure

→ competition concerns

→ proposed commitments

→ no formal Article 7-style oral hearing.

Therefore, the procedural architecture itself affects the scope of adversarial participation.

28. Algorithmic Evidence and the Statement of Objections

The Statement of Objections should communicate the Commission's case sufficiently clearly for the undertaking to respond.

For an algorithmic case, this may require explaining:

alleged conduct;

relevant algorithm;

relevant period;

market;

mechanism of harm;

economic theory;

evidence;

causal connection.

A final decision should not fundamentally transform the case into a materially different theory on which the undertaking had no opportunity to comment.

The Commission's Hearing Officer expressly has responsibility to ensure that the final decision is not based upon objections on which the parties have not been heard. (Competition Policy)

29. New Algorithmic Theory at Final Decision Stage

Consider:

Statement of Objections

Algorithm A foreclosed rivals through discriminatory ranking.

Final decision

Algorithm A additionally constituted an exploitative abuse because it manipulated consumer attention.

If the second theory materially changes the legal or factual case, the undertaking may argue that it was denied an adequate opportunity to respond.

This is a classic rights-of-defence problem.

30. Burden of Proof

The Commission bears the burden of establishing an infringement.

The undertaking does not have to prove:

"Our algorithm is innocent."

Instead, the authority must establish the elements of the alleged infringement according to the applicable legal standard.

The defence can then challenge:

factual evidence;

economic analysis;

algorithmic inference;

causation;

legal qualification.

31. Standard of Review by EU Courts

The General Court and CJEU can review Commission decisions.

Depending on the issue, judicial review can involve:

legal interpretation;

factual assessment;

evidence;

economic analysis;

procedural legality;

proportionality;

fines.

Where procedural defects affect the rights of defence, the decision may be annulled if the defect is sufficiently significant under the applicable jurisprudence.

The case law therefore treats procedural rights as substantive safeguards rather than mere administrative formalities.

32. Digital Evidence and the Size of the File

Algorithmic cases can generate enormous files.

The Commission itself has acknowledged that digitalisation and the proliferation of data have made files increasingly large and that producing non-confidential versions can be extremely time-consuming. (Competition Policy)

This creates a practical problem:

More disclosure does not necessarily mean better defence if the evidence is impossible to process.

Possible solutions include:

searchable databases;

structured datasets;

data rooms;

metadata;

technical summaries;

expert protocols;

confidentiality rings.

33. Algorithmic Explainability

A useful procedural distinction is:

Technical explainability

How does the algorithm operate?

Evidentiary explainability

Why does the algorithmic evidence support the alleged infringement?

Legal explainability

Why does that evidence satisfy Article 101 or 102?

The third is ultimately the most important.

An authority does not establish an infringement merely by showing:

"The algorithm produced this result."

It must connect the result to the legal elements of the competition offence.

34. Automated Evidence Does Not Eliminate Human Adjudication

Even where an authority uses:

AI;

machine learning;

automated screening;

anomaly detection;

the ultimate infringement decision remains a legal decision.

Automated systems may identify:

suspicious price movements;

communication patterns;

ranking changes;

exclusionary conduct.

But investigators must still assess:

context;

causation;

legal relevance;

alternative explanations;

evidence quality.

35. Defence Rights in Algorithmic Dawn-Raid Evidence

During inspections, authorities may obtain:

source code;

emails;

algorithm documentation;

developer communications;

pricing records;

technical specifications.

Later, if that material becomes part of the case, the undertaking should have the procedural opportunity provided by EU competition law to understand the evidence relevant to the objections.

The right of access is therefore particularly important after the investigation develops into an adversarial proceeding.

36. Practical Disclosure Matrix

37. Main Procedural Risks

Risk 1 — Black-box evidence

The authority relies on an unexplained model.

Risk 2 — Insufficient data

The undertaking receives conclusions without enough underlying information.

Risk 3 — Excessive confidentiality

Confidentiality claims prevent meaningful defence.

Risk 4 — Data overload

Millions of documents technically become available but cannot realistically be analysed.

Risk 5 — Changing theory

The final decision relies on a materially different theory.

Risk 6 — Hidden exculpatory evidence

Relevant evidence favourable to the undertaking is not properly accessible.

Risk 7 — Model opacity

The authority cannot adequately explain how its computational model supports the legal conclusion.

38. Legal Test for Algorithmic Hearing Rights

A useful framework is:

Step 1 — Identify the algorithm

What automated system is involved?

Step 2 — Identify its evidentiary role

Is it:

investigative;

corroborative;

central;

determinative?

Step 3 — Identify the legal allegation

Article 101 or Article 102?

Step 4 — Identify the evidence relied upon

What data and model produced the conclusion?

Step 5 — Assess disclosure

Has the undertaking received sufficient information to understand and challenge the evidence?

Step 6 — Assess confidentiality

Is restricted access justified?

Step 7 — Assess exculpatory evidence

Has potentially favourable evidence been made available?

Step 8 — Assess opportunity to respond

Was there adequate time and opportunity for written and oral submissions?

Step 9 — Assess final decision

Did the final decision rely upon objections materially different from those communicated earlier?

Step 10 — Assess prejudice

Did the procedural defect impair the undertaking's ability to defend itself?

39. Six Core Case Laws for Examination

40. Conclusion

Adversarial hearing rights in algorithmic antitrust proceedings are essentially an application of traditional EU rights of defence to increasingly complex digital evidence.

The governing principle can be expressed as:

Complexity of technology cannot reduce the substance of procedural rights.

An undertaking investigated for algorithmic antitrust conduct should have a meaningful opportunity to understand and contest:

Algorithm → Data → Methodology → Inference → Competitive Effect → Legal Qualification

The most important safeguards are:

Statement of Objections

Access to the investigation file

Access to relevant inculpatory and exculpatory evidence

Adequate disclosure of economic/algorithmic methodology

Protection balanced against legitimate confidentiality

Written response

Oral hearing in Article 7 proceedings

Opportunity to challenge expert and economic analysis

No material new objection without an opportunity to respond

Effective judicial review

The EU Commission expressly treats access to the file and the right to be heard as core procedural guarantees, and its Hearing Officer can resolve disputes concerning access, confidentiality, hearing participation and procedural fairness. (Competition Policy)

Core legal formula

Algorithmic Antitrust Investigation + Relevant Evidence + Effective Disclosure + Ability to Challenge Methodology + Confidentiality Balance + Written/Oral Hearing + Judicial Review = Adversarial Procedural Fairness

The key conceptual point is that an algorithm may generate evidence, but it cannot replace the adversarial legal process through which that evidence is tested.