Model Bias and Discriminatory Enforcement in Competition Authorities

1. Introduction

Model bias and discriminatory enforcement arise when competition authorities use artificial intelligence (AI), machine-learning systems, statistical models, automated screening tools, or algorithmic risk assessments in ways that systematically disadvantage particular firms, industries, business models, or categories of market participants.

Competition authorities increasingly rely on data-driven tools to identify cartels, detect bid-rigging, screen mergers, assess market dominance, investigate digital platforms, and prioritise enforcement resources. These tools can improve efficiency, but they can also reproduce historical enforcement patterns, misinterpret legitimate commercial conduct, or disproportionately target businesses whose operations generate more readily observable data.

The central legal issue is whether algorithm-assisted enforcement remains consistent with equality before the law, procedural fairness, evidential reliability, proportionality, transparency, and independent decision-making.

An important distinction must be maintained: a model that produces unequal outcomes does not automatically establish unlawful discrimination, and an authority's use of AI does not itself invalidate an investigation. The relevant questions concern the source of the disparity, its justification, the reliability of the evidence, and the legal safeguards available to affected parties.

2. Meaning and forms of model bias

A. Historical-data bias

An enforcement model trained on past investigations may reproduce the priorities and assumptions of earlier enforcement practices. If previous investigations disproportionately focused on particular industries, the model may interpret firms in those industries as inherently riskier.

For example, if historic cartel cases mainly involved construction procurement, an automated system may flag construction contractors more frequently than firms in other sectors, even when comparable risk indicators exist elsewhere.

B. Sampling and detection bias

The data available to an authority may not represent the entire market. Large platforms, publicly listed companies, and firms participating in formal procurement systems often generate more structured data than small businesses or informal market participants.

Consequently, the model may identify misconduct more readily among highly observable firms without establishing that those firms engage in more anticompetitive conduct.

C. Proxy discrimination

A model may use variables that indirectly represent characteristics unrelated to the legitimate enforcement objective. Examples include geographical location, company size, language, transaction frequency, or particular business models.

A proxy is not necessarily unlawful. Its legality depends on its relevance, necessity, accuracy, and effect in the particular statutory context.

D. Model specification and feedback bias

A model may incorrectly equate price parallelism with collusion, market share with dominance, or aggressive discounting with exclusionary conduct. If its flags generate more investigations, and those investigations generate more training data, the system can reinforce its original assumptions.

E. Institutional or enforcement bias

Bias can also arise without an AI system. Officials may exercise discretion inconsistently, rely on stereotypes about particular industries, or prioritise politically salient firms. AI may magnify these problems by making subjective assumptions appear scientifically objective.

3. Legal framework

A. Equality and non-arbitrariness

Competition authorities must act within their statutory powers and apply legally relevant criteria consistently. In the UK, relevant principles include the rule of law, procedural fairness, rationality, and applicable equality obligations. In Germany, Article 3 of the Basic Law guarantees equality before the law, while Article 19(4) protects access to judicial review where public authority action infringes rights.

In the European Union, the Charter of Fundamental Rights includes equality before the law under Article 20, non-discrimination under Article 21, and rights to good administration and an effective remedy under Articles 41 and 47, respectively, subject to their respective scopes of application.

B. Competition-law requirements

An algorithmic risk score is not a substitute for the statutory elements of an infringement.

For example:

Under Article 101 TFEU, the authority must establish the legally required elements of an agreement, decision, or concerted practice.

Under Article 102 TFEU, dominance alone is not an infringement; abusive conduct must be established.

Under the UK Competition Act 1998, the applicable Chapter I or Chapter II requirements must be satisfied.

Under Germany's Act Against Restraints of Competition (GWB), the relevant statutory conditions must be established, including those applicable to dominant undertakings and certain digital undertakings under Section 19a.

A model can help identify evidence, but a probability score cannot independently establish liability.

C. Procedural fairness and effective review

Where algorithmic analysis materially influences an investigation or infringement decision, affected firms may need a meaningful opportunity to understand and challenge the relevant evidence. The precise disclosure obligations depend on the applicable legal regime, procedural stage, confidentiality rules, and rights of defence.

Relevant safeguards include disclosure of material relied upon, an explanation of the model's role, validation of its accuracy, access to the underlying evidence where legally required, and judicial scrutiny of the authority's reasoning.

D. Data protection and automated decision-making

Where personal data are processed, the GDPR may impose additional requirements. Article 22 concerns certain solely automated decisions producing legal or similarly significant effects on individuals; it does not automatically prohibit every automated investigative tool or decision involving businesses. Articles 5, 13–15 and 35 may also be relevant depending on the processing, transparency obligations, and risk assessment required.

The EU AI Act may impose additional requirements where a system falls within its scope and a relevant classification. Whether a competition authority's particular system qualifies as high-risk must be assessed against the legislation's actual categories and applicable exceptions.

4. Key case laws: at least six judicial authorities

The following cases establish principles relevant to model bias, evidential assessment, equal treatment, procedural safeguards, and review of competition-enforcement decisions. They are not judgments directly deciding whether an AI model used by a competition authority was discriminatory. Their relevance lies in the legal standards they provide for evaluating algorithm-assisted enforcement.

Case 1: Aalborg Portland A/S v Commission (Joined Cases C-204/00 P and others, 2004)

Legal principle: Rights of defence and access to evidence.

The Court of Justice of the European Union emphasised the importance of access to relevant inculpatory and exculpatory material in competition proceedings. An undertaking must have a genuine opportunity to address the facts and evidence used against it.

EUR-Lex

+1

Relevance to model bias: If a competition authority uses an algorithm to identify alleged cartel participation, the resulting suspicion cannot excuse procedural unfairness. Where the model's output or supporting evidence materially informs the case, the authority must comply with applicable disclosure and defence rights.

Application: A company accused of bid-rigging should be able to contest the underlying transaction evidence, explain legitimate pricing patterns, and challenge adverse inferences drawn from the data.

Case 2: Solvay SA v Commission (Case C-109/10 P, 2011)

Legal principle: Effective access to the investigation file.

The Court examined the consequences of restrictions on access to documents in a competition proceeding. It reiterated that infringement of access rights can justify annulment where the undertaking's rights of defence have been affected. Later disclosure during judicial proceedings does not necessarily cure the original defect.

EUR-Lex

Relevance to model bias: An authority should not rely on an opaque automated assessment while withholding material necessary for the undertaking to challenge the case against it.

Application: If a model's classification depends on transaction records, comparator data, or analytical assumptions that materially influence the allegations, the authority should assess what disclosure is required to permit an effective defence, subject to legitimate confidentiality protections.

Case 3: KME Germany AG v Commission (Case C-272/09 P, 2011)

Legal principle: Effective judicial review of competition decisions and fines.

The Court of Justice considered the intensity of judicial review in competition cases, including the General Court's jurisdiction concerning fines.

Relevance to model bias: Judicial review should not become merely formal because an authority relies on sophisticated statistical techniques. Courts must exercise the review powers conferred on them by law and examine relevant challenges to the factual and legal basis of a decision.

Application: Where a fine is based partly on model-generated findings, the authority should be able to explain the evidential connection between the model's output, the established infringement, and the penalty imposed.

Case 4: Chalkor AE Epexergasias Metallon v Commission (Case C-386/10 P, 2011)

Legal principle: Effective judicial scrutiny and the presumption of innocence.

The Court addressed the requirements of effective judicial review in competition proceedings and the relationship between administrative enforcement and judicial control.

Relevance to model bias: An algorithmic score should not be treated as conclusive proof of wrongdoing. The legality of an enforcement decision must remain open to meaningful judicial examination.

Application: A firm flagged by an automated cartel-detection system must not be presumed to have participated in collusion merely because its prices resemble those of competitors. The authority must establish the infringement under the applicable legal standard.

Case 5: Tetra Laval BV v Commission (Case C-12/03 P, 2005)

Legal principle: Careful judicial examination of complex economic assessments and predictive reasoning.

The case concerned merger control and the Commission's assessment of the likely future effects of a proposed concentration. The Court scrutinised the evidential basis for the Commission's economic predictions.

Relevance to model bias: Competition authorities frequently use predictive models to assess merger effects, innovation incentives, foreclosure, and future market structure. Sophisticated modelling does not remove the requirement for a sufficiently reliable factual and economic foundation.

Application: If an AI model predicts that a merger will reduce innovation, the authority should examine its assumptions, data quality, causal reasoning, alternative scenarios, and the sensitivity of its conclusions to changes in inputs.

Case 6: Intel Corp. v Commission (Case C-413/14 P, 2017)

Legal principle: Proper assessment of evidence and economic arguments in abuse-of-dominance cases.

The Court of Justice clarified the importance of examining relevant arguments and evidence concerning whether allegedly exclusionary rebate practices are capable of restricting competition, where the undertaking advances such arguments.

Relevance to model bias: A risk model that categorises rebates, discounts, or pricing practices as suspicious cannot replace the required legal and economic assessment. Relevant contextual evidence and counterarguments must be addressed.

Application: An authority investigating a dominant undertaking's discounts should not rely solely on an automated classification. It should consider the market context, the firm's conduct, the evidence of foreclosure, and relevant economic analysis.

Case 7: Commission v Dole Food and Dole Germany (Case C-286/13 P, 2015)

Legal principle: Evidential assessment in cartel cases.

The Court considered the Commission's findings concerning information exchanges and price coordination in the banana market.

Relevance to model bias: Automated systems may detect communications, price correlations, or recurring market patterns that warrant investigation. However, the legal significance of those patterns depends on their context and their connection to the alleged coordination.

Application: A model that identifies parallel price movements should distinguish between potentially unlawful coordination and alternative explanations, including common cost changes, public market information, and independent responses to demand.

Case 8: Google and Alphabet v Commission (Google Shopping) (Case C-48/22 P, 2024)

Legal principle: Assessment of alleged self-preferencing and exclusionary conduct by a dominant platform.

The Court of Justice upheld the General Court's judgment in the Google Shopping litigation, concerning the Commission's finding that Google's treatment of its comparison-shopping service constituted an abuse of dominance.

Relevance to model bias: Digital-market investigations may depend on ranking data, traffic allocation, visibility metrics, and comparisons between platform-owned and third-party services. The authority must distinguish a meaningful competitive disadvantage from a disparity that is adequately explained by legitimate differences.

Application: An algorithmic audit of a platform's ranking practices should consider relevant comparators, the mechanism producing differential treatment, and the effects on competition rather than assuming that every ranking difference is discriminatory.

5. How discriminatory enforcement can occur in practice

Consider a competition authority that introduces an AI tool to identify potentially collusive firms.

Stage 1 — Historical enforcement data

Past investigations and infringement decisions supply the training data.

Stage 2 — Model learns historical patterns

The model associates particular sectors, pricing patterns, or business characteristics with higher risk.

Stage 3 — Unequal investigation rates

Some firms are flagged more frequently, potentially because they are more observable or resemble previously investigated firms.

Stage 4 — Feedback loop

The resulting investigations generate further data that may reinforce the original pattern.

This cycle creates a risk of self-reinforcing enforcement bias. The existence of this mechanism does not prove that any particular authority has discriminated; it identifies a risk that must be tested empirically.

6. Legal consequences of biased enforcement

A. Unlawful unequal treatment

If comparable firms are treated differently without an objective and legally relevant justification, the authority may breach applicable equality or administrative-law requirements. Differences in enforcement intensity alone, however, do not necessarily establish a violation.

B. Defective findings of infringement

Where a biased model materially distorts the evidence or leads the authority to overlook exculpatory facts, the resulting decision may be vulnerable to challenge on evidential, legal, or procedural grounds.

C. Invalid or disproportionate penalties

If a model incorrectly estimates the seriousness, duration, or economic impact of conduct, it may distort penalty calculations. The authority must apply the governing fining rules and explain the basis for its assessment.

D. Breach of rights of defence

Failure to provide legally required access to relevant evidence or a meaningful opportunity to challenge the authority's case may constitute a procedural defect. The consequences depend on the applicable law and the effect of the defect on the proceedings.

E. Institutional legitimacy and public confidence

Repeatedly concentrating enforcement on particular sectors or business models without adequate justification can undermine confidence in the authority's impartiality. Even where no legal infringement is established, unexplained disparities may indicate weaknesses in the authority's governance arrangements.

7. Comparative legal position: UK, EU and Germany

These frameworks do not establish a universal requirement that every competition-enforcement model be fully open-sourced. Transparency obligations must be balanced against confidentiality, business secrets, investigative integrity, and the rights of other parties.

8. A practical framework for preventing model bias

Competition authorities should adopt a lifecycle approach rather than relying on a single fairness audit.

Audit the training data. Identify gaps in sector coverage, historical enforcement patterns, missing observations, and unreliable labels.

Test disparate outcomes. Compare investigation rates, false-positive rates, and error rates across relevant sectors, firm sizes, and business models. Where legally and ethically appropriate, examine protected characteristics and indirect proxies.

Validate the model independently. Test performance on data not used for training and measure whether the model remains reliable across markets and time periods.

Require human review. A qualified official should evaluate the context and evidential basis of material flags instead of mechanically accepting model scores.

Preserve an audit trail. Record model versions, input data, material assumptions, thresholds, overrides, and the reasons for investigative decisions.

Protect defence rights. Ensure that the model's use does not prevent affected parties from challenging material evidence or presenting relevant counterarguments.

Provide independent oversight. Establish internal audit, legal review, and appropriate external scrutiny, with procedures for correcting errors and suspending unreliable systems.

9. Hypothetical case study

Suppose a competition authority uses a machine-learning system to identify potential price-fixing in retail markets. The model repeatedly flags small independent retailers because their prices move closely with those of major supermarket chains.

An investigation reveals that both groups purchase goods from the same wholesalers and face identical changes in wholesale costs.

The model's correlation-based flag is therefore insufficient to establish collusion.

A legally sound response would involve:

Examining whether the model has confused common market shocks with coordinated conduct.

Reviewing evidence of communications, information exchange, or other facts relevant to the alleged infringement.

Comparing false-positive rates across different retail formats.

Documenting why the authority proceeds with or closes an investigation.

Reassessing the model if the same error systematically affects a particular category of retailer.

If the authority disregards this evidence and relies on a defective model, the resulting decision may be open to challenge. If it investigates the flag impartially and ultimately finds no infringement, the initial flag alone does not establish discriminatory enforcement.

10. Critical evaluation

The principal difficulty is that enforcement bias can be both statistical and institutional. A model may be technically accurate according to its training objective while still being unsuitable for the legal question it is intended to support. For example, predicting which firms are likely to be investigated successfully is not the same as determining whether a firm has infringed competition law.

Three safeguards are particularly important:

Separate risk prediction from proof of infringement. Model outputs should guide investigative priorities, not replace the legal standard of proof.

Evaluate outcomes, not merely accuracy. Overall predictive accuracy can conceal disproportionate errors affecting particular categories of firms.

Maintain accountable human decision-making. Officials must be able to explain the legal and evidential basis for their decisions rather than treating an algorithm as an unquestionable authority.

There is also a risk of overcorrection. Forcing equal investigation numbers across every sector could prevent authorities from responding appropriately to genuine differences in cartel risk, market structure, or evidence. The objective is not identical treatment in every circumstance, but consistent application of lawful criteria, supported by reliable evidence and objectively justified distinctions.

11. Conclusion

Model bias and discriminatory enforcement pose a significant governance challenge for data-driven competition authorities. The principal legal concern is not simply whether an algorithm produces unequal results, but whether its use undermines lawful discretion, equality, evidential reliability, procedural fairness, or effective judicial review.

The eight cases discussed above—Aalborg Portland, Solvay, KME, Chalkor, Tetra Laval, Intel, Commission v Dole Food, and Google Shopping—provide relevant judicial principles concerning defence rights, evidence, economic analysis, and judicial scrutiny. They should be understood as analogical legal authorities, not direct precedents establishing a freestanding prohibition on AI bias in competition enforcement.