Competitive Harm Prediction Certification Systems

1. Introduction

A Competitive Harm Prediction Certification System is a regulatory, compliance, or institutional mechanism through which an undertaking, regulator, auditor, or independent assessor evaluates whether a proposed business practice, merger, algorithm, digital platform design, AI system, procurement arrangement, or market intervention is likely to produce material harm to competition, and then issues a certification, risk classification, or clearance based on that prediction.

The concept is particularly relevant to AI-driven and digital markets, where competition problems may develop rapidly and may become difficult to reverse once network effects, data advantages, switching costs, interoperability barriers, or ecosystem dependence have become entrenched.

The UK CMA has expressly recognised that increasingly sophisticated algorithmic systems can make competitive harms less transparent and has examined techniques for monitoring and auditing algorithmic systems.

A certification system therefore attempts to move competition policy partly from:

“Detect and remedy harm after it occurs”

towards:

“Predict, document, certify, monitor and prevent foreseeable competitive harm.”

Importantly, such certification would not itself replace competition-law enforcement. A prediction is evidence or a governance mechanism; it does not automatically establish an infringement.

2. Meaning of Competitive Harm Prediction

Competitive harm prediction involves estimating whether conduct is likely to impair one or more dimensions of competition.

These may include:

1. Price competition
2. Quality competition
3. Innovation
4. Consumer choice
5. Market entry
6. Access to essential inputs
7. Interoperability
8. Data access
9. Multi-homing
10. Potential competition
11. Contestability
12. Rival viability
13. Market resilience
14. Competitive neutrality

The modern approach is increasingly based on a theory of harm.

The CMA's current merger guidance describes a theory of harm as a hypothesis concerning how the competitive process could be harmed and emphasises that merger assessment is forward-looking.

Thus:

Conduct → Mechanism of harm → Predicted competitive effect → Evidence → Certification decision → Monitoring

3. What Is a Certification System?

A certification system would establish a formal process through which a firm or regulator demonstrates that a proposed action satisfies predetermined competition-risk requirements.

For example:

Level 1 — Low-risk certification

The system predicts:

• little market-power effect;
• no material foreclosure;
• substantial alternative suppliers;
• low switching costs;
• limited network effects.

The conduct may receive ordinary certification.

Level 2 — Conditional certification

The system identifies risks involving:

• interoperability;
• data access;
• preferential treatment;
• algorithmic pricing;
• exclusive contracts;
• switching costs.

Certification could be conditional upon safeguards.

Level 3 — High-risk certification

The prediction identifies:

• substantial foreclosure;
• ecosystem tipping;
• elimination of an important potential competitor;
• significant innovation harm;
• dependence on an essential platform input.

Enhanced regulatory scrutiny could follow.

Level 4 — No certification

The predicted harm is sufficiently serious or the evidence insufficient to justify certification.

This should not automatically be treated as proof of illegality; it could instead trigger formal investigation.

4. Core Components

A. Market-definition module

The system must identify the relevant competitive space.

In digital markets this may require consideration of:

• multi-sided platforms;
• zero-price services;
• data competition;
• interoperability;
• ecosystem competition;
• potential competitors;
• adjacent markets;
• future products.

Traditional market shares may be inadequate where competition occurs through innovation or ecosystem expansion.

B. Market-power assessment

The system should examine:

• market shares;
• concentration;
• entry barriers;
• switching costs;
• network effects;
• data advantages;
• economies of scale;
• access to compute;
• interoperability;
• vertical integration.

A certification system should avoid treating market share as conclusive.

C. Theory-of-harm engine

The system should identify the mechanism through which harm could occur.

Examples include:

Exclusion

Dominant firm → restrictive conduct → rival foreclosure → reduced competition.

Raising rivals' costs

Platform → discriminatory access → rival costs increase → competitive constraint weakens.

Innovation suppression

Acquisition → removal of innovative competitor → reduced future rivalry.

Coordinated effects

Merger → increased market transparency → easier coordination → reduced competitive pressure.

Ecosystem tipping

Platform advantage → network effects → user migration → reinforcing advantage → market becomes difficult to contest.

5. Predictive Models

A sophisticated system could use several analytical techniques.

A. Structural economic models

Estimate:

• price effects;
• diversion ratios;
• margins;
• demand elasticity;
• substitution.

B. Event studies

Analyse historical market responses to:

• acquisitions;
• entry;
• exclusion;
• platform-policy changes;
• interoperability restrictions.

C. Simulation

Model possible post-conduct scenarios.

D. Machine learning

Identify patterns associated with:

• foreclosure;
• coordinated pricing;
• customer lock-in;
• declining entry;
• discriminatory ranking.

E. Agent-based modelling

Particularly useful for digital markets because firms, consumers and algorithms can interact dynamically.

F. Counterfactual analysis

The central question becomes:

What would competition look like absent the proposed conduct?

This is particularly important because competition law frequently requires comparison between the actual or proposed situation and an appropriate counterfactual.

6. Certification Evidence

A certification system should not rely exclusively on a black-box prediction.

It should require:

Quantitative evidence

• market shares;
• prices;
• margins;
• switching rates;
• entry rates;
• churn;
• innovation measures.

Qualitative evidence

• internal documents;
• business plans;
• product roadmaps;
• strategic documents;
• technical documentation.

Technical evidence

For AI systems:

• model architecture;
• training objectives;
• datasets;
• optimisation objectives;
• logs;
• model outputs;
• API behaviour;
• deployment constraints.

Experimental evidence

• A/B tests;
• controlled experiments;
• sandbox testing;
• simulated entry;
• interoperability tests.

7. Six Important Case Laws

The concept of competitive-harm prediction certification is not itself an established standalone doctrine in competition law. Its legal foundations can instead be constructed from jurisprudence concerning prospective competitive effects, counterfactual analysis, theories of harm, evidence, potential competition, foreclosure and coordinated effects.

1. Airtours v Commission

Case: Airtours plc v Commission, T-342/99

Principle

The case concerned the Commission's theory that a merger could produce coordinated effects.

The General Court required the Commission to establish the conditions necessary for coordination and to demonstrate the likelihood of such coordination.

Relevance to certification

A prediction system cannot merely say:

“This merger looks dangerous.”

It must identify the mechanism by which competition could be harmed.

A certification framework should therefore contain:

Risk prediction → economic mechanism → supporting evidence → counterfactual → predicted competitive effect.

Airtours is particularly important for coordinated-effects certification.

2. Tetra Laval v Commission

Case: Tetra Laval BV v Commission, C-12/03 P

Principle

The Court emphasised the need for sufficiently convincing evidence when predicting future competitive behaviour following a merger.

The case is highly relevant because merger control inherently involves prediction of future conduct and effects.

Relevance to certification

A certification system should distinguish:

• established facts;
• reasonable economic inference;
• assumptions;
• predictions.

Predictions should be supported by evidence rather than merely by theoretical possibility.

This creates an important principle:

Predictive certification must be evidence-based and logically demonstrable.

3. Microsoft v Commission

Case: Microsoft Corp v Commission, T-201/04

Principle

The case concerned interoperability, tying and exclusionary effects in software markets.

The General Court examined how control over an important technological interface could affect rivals and competition.

Relevance to certification

Modern certification systems could assess:

• API access;
• interoperability;
• technical compatibility;
• data portability;
• interface degradation;
• ecosystem dependence.

For example:

API restriction → rival functionality decreases → switching becomes harder → rival expansion declines.

A certification system could therefore require an interoperability-impact assessment before certifying major platform changes.

4. Intel v Commission

Case: Intel Corp v Commission, C-413/14 P

Principle

The Intel litigation is important for the assessment of exclusionary rebates and the role of economic evidence concerning foreclosure.

The Court's jurisprudence strengthened the importance of examining the actual or potential capacity of conduct to produce exclusionary effects where relevant economic evidence is presented.

Relevance to certification

A predictive certification system should examine:

• the firm's market position;
• coverage of the conduct;
• duration;
• conditions;
• foreclosure mechanisms;
• ability of rivals to compete.

It demonstrates why certification should not simply rely on the formal description of conduct.

Instead:

Conduct + market circumstances + economic mechanism + likely effects

should form the predictive model.

5. Google Shopping

Case: Google and Alphabet v Commission, T-612/17

Principle

The Google Shopping litigation involved the preferential positioning and display of Google's own comparison-shopping services relative to competing services.

The case illustrates how a platform can influence competitive conditions through ranking and visibility, rather than simply through conventional pricing.

Relevance to certification

This is particularly important for AI and algorithmic markets.

A certification system could test whether:

• ranking algorithms favour affiliated products;
• recommendation systems disadvantage rivals;
• search visibility is manipulated;
• platform data is used to advantage vertically integrated businesses;
• algorithmic changes increase barriers to entry.

The broader lesson is that competitive harm can arise from control over digital access points, not merely from price.

6. CK Telecoms v Commission

Case: CK Telecoms UK Investments Ltd v Commission, T-399/16

Principle

The case concerned the Commission's assessment of a merger involving mobile telecommunications operators and the loss of competitive pressure.

It is particularly significant for the assessment of unilateral effects and dynamic competitive constraints.

Relevance to certification

A certification system could ask:

What competitive constraint disappears if the transaction or conduct is implemented?

This is particularly important where the target is not yet a major market-share holder but represents:

• potential competition;
• innovation competition;
• disruptive technology;
• future entry;
• an emerging competitive constraint.

A certification system therefore should not measure only current competition.

It should predict the loss of future competition.

8. Additional Relevant Authorities

Several other authorities provide useful foundations.

United States v. Microsoft Corp.

The Microsoft litigation illustrates how exclusionary conduct involving technological platforms can affect future competition and innovation.

Bronner v Mediaprint

Useful for examining exceptional circumstances involving access to indispensable infrastructure.

Post Danmark

Relevant to effects-based analysis of exclusionary conduct.

Tomra v Commission

Important for understanding foreclosure through exclusivity and the significance of market coverage.

Qualcomm

The litigation illustrates the importance of examining economic mechanisms and competitive effects in technologically sophisticated markets.

9. Certification of AI Systems

The concept becomes especially significant with AI.

Imagine a dominant platform introducing an AI recommendation engine.

The certification system could test:

Input layer

Does the system use:

• competitor data?
• customer data?
• proprietary platform data?

Model layer

Does the optimisation function reward:

• platform products?
• affiliated services?
• higher-margin products?
• reduced visibility for competitors?

Output layer

Does the system systematically:

• downgrade rivals;
• increase switching costs;
• restrict discovery;
• reinforce network effects?

Market layer

Does this produce:

• foreclosure;
• tipping;
• increased concentration;
• reduced innovation?

10. Competitive Harm Scorecard

A certification framework could contain several dimensions:

The resulting assessment should preferably produce risk categories and reasons, rather than a simplistic numerical "competition score."

11. Certification and Algorithmic Collusion

One particularly important application concerns pricing algorithms.

The CMA has identified several algorithmic-collusion scenarios, including information exchange through common algorithmic providers, hub-and-spoke arrangements, and the possibility of autonomous systems producing coordinated outcomes.

A certification system could therefore require firms deploying sophisticated pricing algorithms to demonstrate:

1. what data the algorithm receives;
2. whether competitor-specific information is used;
3. whether competitors use the same provider;
4. whether optimisation objectives reward parallel pricing;
5. whether the algorithm responds automatically to competitor prices;
6. whether safeguards prevent coordination.

This would create an Algorithmic Competition Impact Assessment.

12. Certification and Merger Control

Certification could be particularly useful in mergers involving:

• AI companies;
• cloud providers;
• semiconductor firms;
• foundation-model developers;
• data companies;
• app stores;
• digital advertising platforms.

A certification application might contain:

Part I — Current competition

Who competes today?

Part II — Potential competition

Who could compete tomorrow?

Part III — Innovation

Which technologies could emerge?

Part IV — Inputs

Which competitors depend upon:

• compute;
• data;
• cloud infrastructure;
• APIs;
• distribution?

Part V — Counterfactual

What would happen without the transaction?

Part VI — Remedies

Could identified risks be mitigated?

The CMA's current merger guidelines expressly contemplate forward-looking theories of harm, including unilateral effects, potential and dynamic competition, coordinated effects, and foreclosure.

13. Certification and Essential Facilities

For AI infrastructure, certification could assess whether a firm controls an input that competitors cannot reasonably reproduce.

Examples:

• specialised compute;
• unique datasets;
• cloud infrastructure;
• model-access APIs;
• specialised chips;
• technical standards.

The assessment should consider:

Indispensability → availability of alternatives → replication possibility → access conditions → foreclosure effects.

However, certification should not automatically transform every important digital resource into an "essential facility."

14. Ex-Ante and Ex-Post Certification

Ex-ante certification

Occurs before conduct.

Examples:

• merger;
• algorithm deployment;
• API restructuring;
• platform redesign;
• exclusive agreement.

Advantages:

• prevents irreversible harm;
• identifies risks early;
• encourages compliance by design.

Disadvantage:

• predictions may be uncertain.

Ex-post certification

Occurs after implementation.

The regulator examines whether predicted harms actually materialised.

This creates a feedback mechanism:

Prediction → Implementation → Observation → Verification → Model correction

This is particularly valuable for machine-learning systems because their behaviour may change after deployment.

15. The Counterfactual Problem

The biggest difficulty is determining what would have happened without the conduct.

For example:

A rival's market share falls 20% after a platform changes its ranking algorithm.

That does not automatically prove that the algorithm caused the decline.

Alternative explanations might include:

• superior rival products;
• changing consumer preferences;
• macroeconomic conditions;
• technological change;
• new entrants;
• pricing changes.

Certification must therefore distinguish correlation from causation.

16. False Positives and False Negatives

A prediction system creates two major risks.

False positive

The system predicts serious competitive harm, but the harm does not occur.

Consequences:

• unnecessary regulatory intervention;
• reduced innovation;
• delayed investment;
• excessive compliance costs.

False negative

The system predicts low risk, but serious harm occurs.

Consequences:

• foreclosure;
• market tipping;
• increased concentration;
• irreversible innovation loss.

Therefore, certification should not treat predictive accuracy as absolute.

17. Explainability Requirement

A certification decision should explain:

1. What was predicted?
2. Why was it predicted?
3. Which data was used?
4. What assumptions were made?
5. What counterfactual was used?
6. What uncertainty exists?
7. What evidence contradicts the prediction?
8. How can the prediction be challenged?

This is particularly important for AI-based certification.

A regulator should not simply receive:

"Model predicts high competitive harm."

It should receive:

"Model predicts high foreclosure risk because X controls Y input, rival Z depends upon Y, switching costs are high, and the counterfactual indicates that Z would otherwise expand."

18. Independent Audit

Certification should ideally involve independent review.

Possible architecture:

Firm's internal assessment
↓
Independent competition auditor
↓
Technical/algorithmic audit
↓
Competition authority review
↓
Certification / conditional certification / investigation

This reduces the danger of firms designing their own predictive models in a way that systematically understates harm.

19. Certification Does Not Create Immunity

This is a critical legal principle.

A firm should not be able to argue:

"We received a competition certificate, therefore competition law cannot apply."

Certification should instead function as:

• evidence of compliance;
• risk-management documentation;
• regulatory monitoring;
• an early-warning mechanism.

If circumstances change, the certification should be capable of:

• review;
• suspension;
• modification;
• withdrawal.

20. Continuous Certification

Static certification is poorly suited to AI markets.

A better model is:

Initial certification → continuous monitoring → periodic reassessment → incident reporting → re-certification.

Triggers for reassessment could include:

• major market-share changes;
• acquisition of a competitor;
• new algorithm deployment;
• significant API changes;
• deterioration in interoperability;
• increased switching costs;
• evidence of coordinated behaviour.

The CMA has itself emphasised that digital markets are dynamic and that algorithmic systems can change rapidly, creating new competitive risks.

21. Competition Certification and Regulatory Sandboxes

Certification can be combined with regulatory sandboxes.

A firm could deploy a new AI system in a controlled environment.

The regulator could observe:

• rival treatment;
• consumer switching;
• prices;
• recommendations;
• interoperability;
• innovation effects.

Only after satisfactory testing would full certification be considered.

This is especially appropriate where traditional ex-ante economic models have difficulty predicting novel technology markets.

22. Legal Standards for Certification

A legally robust framework should contain:

1. Clear statutory authority

The regulator must have legal authority to require certification.

2. Defined assessment criteria

Firms must know what factors are evaluated.

3. Evidentiary standards

Predictions must be based on sufficiently reliable evidence.

4. Procedural fairness

Affected firms should have an opportunity to respond.

5. Transparency

Reasons for certification decisions should be recorded.

6. Confidentiality

Trade secrets and sensitive technical information require protection.

7. Appeal

There should be judicial or administrative review.

8. Periodic reassessment

Certification should not become permanent immunity.

23. Major Advantages

Competitive harm prediction certification could:

• encourage competition-by-design;
• detect risks before market tipping;
• improve regulatory transparency;
• create audit trails;
• reduce information asymmetry;
• improve AI governance;
• facilitate continuous monitoring;
• encourage interoperability;
• identify potential foreclosure;
• improve merger-risk analysis.

24. Major Risks

However, the system could itself generate competition problems.

Regulatory overreach

Prediction may become a substitute for proof.

Model bias

The certification model may systematically overestimate or underestimate certain forms of harm.

Strategic gaming

Companies could optimise behaviour to pass the certification test without changing the underlying competitive effect.

Regulatory capture

Dominant firms may influence certification standards.

False certainty

A numerical prediction can create an illusion of scientific precision.

Innovation chilling

Overly conservative predictions may discourage experimentation.

Confidentiality problems

Certification requires access to commercially sensitive information.

25. Proposed Legal Framework

A useful model would be:

Stage 1 — Screening

Determine whether the conduct presents meaningful competitive risk.

Stage 2 — Market analysis

Identify relevant markets, competitors, potential entrants and ecosystem dependencies.

Stage 3 — Theory of harm

Identify the specific mechanism of potential harm.

Stage 4 — Counterfactual

Establish the relevant alternative scenario.

Stage 5 — Predictive assessment

Use economic, technical and empirical evidence.

Stage 6 — Independent validation

Audit the assumptions and methodology.

Stage 7 — Certification

Issue:

• certified;
• conditionally certified;
• enhanced monitoring;
• certification refused.

Stage 8 — Continuous monitoring

Measure actual market outcomes.

Stage 9 — Reassessment

Compare predicted and observed outcomes.

Stage 10 — Enforcement

Where actual conduct infringes competition law, ordinary enforcement mechanisms remain available.

26. Relationship With Traditional Competition Law

The system should complement, rather than replace:

• Article 101 TFEU / Chapter I prohibition
• Article 102 TFEU / Chapter II prohibition
• UK merger control
• Competition Act enforcement
• digital-market regulation
• sectoral regulation
• consumer protection law

Its principal contribution is anticipatory evidence and governance.

27. Key Doctrinal Proposition

The emerging legal principle can be expressed as:

The more consequential, irreversible and technologically opaque the potential competitive harm, the greater the justification for structured ex-ante risk assessment, provided that predictive certification remains evidence-based, reviewable and distinct from a finding of infringement.

This reconciles technological innovation with competition-law precaution without treating every predicted harm as an established violation.

28. Conclusion

Competitive Harm Prediction Certification Systems represent a possible evolution of competition governance from purely reactive enforcement toward predictive, auditable and continuously monitored competition protection.

The jurisprudence of Airtours, Tetra Laval, Microsoft, Intel, Google Shopping and CK Telecoms demonstrates different components necessary for such a system: identifiable theories of harm, credible evidence, counterfactual analysis, effects on potential or actual competition, technological foreclosure analysis and assessment of future competitive constraints.

The strongest legal architecture would therefore not create a simple "AI competition score." Instead, it would create a reasoned certification process consisting of:

Market assessment → Theory of harm → Counterfactual → Evidence → Prediction → Independent validation → Certification → Monitoring → Reassessment.

Such a system would be particularly relevant to AI platforms, foundation models, algorithmic pricing, digital ecosystems, APIs, cloud infrastructure, data markets and technology mergers, where competitive harm may develop faster than conventional ex-post enforcement can address it.