Competition Law and Competition Governance in Self-Learning Markets

Introduction

Self-learning markets are markets in which firms use machine-learning or AI systems that continuously process data, observe market conditions, predict consumer or competitor behaviour, and modify commercial decisions without requiring a human to specify every individual decision.

Examples include:

• AI-driven dynamic pricing;
• autonomous advertising auctions;
• algorithmic product rankings;
• recommendation and search systems;
• ride-hailing surge pricing;
• automated credit and insurance pricing;
• AI procurement and bidding;
• hotel and airline revenue management;
• algorithmic inventory allocation;
• autonomous trading;
• platform seller-selection systems; and
• AI agents that negotiate or purchase goods and services.

Competition law generally does not prohibit the use of machine learning itself. The legal issue arises when a self-learning system becomes the mechanism through which firms coordinate prices, exclude competitors, exploit market power, discriminate between trading partners, restrict access, or otherwise reduce competitive constraints.

The central difficulty is that traditional competition law often looks for a human agreement, deliberate decision, or identifiable exclusionary strategy, whereas a self-learning system may continuously modify its behaviour after deployment.

I. Meaning of a Self-Learning Market

A conventional algorithm follows predetermined instructions:

Input → Rule → Output.

A self-learning system is different:

Data → Model → Prediction → Commercial decision → New data → Model adjustment → New decision.

Consequently, the competitive behaviour of the undertaking may evolve over time.

For example, a hotel's AI system may initially compete aggressively on price. After observing competitors' prices for several months, it might learn that maintaining prices above a particular level produces higher returns. If several competitors employ systems that learn from the same market signals, prices may converge without the traditional cartel meeting.

This creates an important distinction:

1. Human-directed algorithmic conduct

The undertaking deliberately programs the system to achieve an anti-competitive objective.

2. Algorithm-assisted coordination

Firms communicate or exchange information and use algorithms to implement the resulting coordination.

3. Algorithmic facilitation

A common platform or algorithm provides information or recommendations that facilitate coordinated behaviour.

4. Autonomous algorithmic convergence

Separate systems independently learn behaviour that produces coordinated market outcomes.

The fourth category presents some of the most difficult unresolved questions for competition law.

II. Objectives of Competition Governance

Competition governance in self-learning markets should pursue several objectives simultaneously:

1. Preservation of independent competitive decision-making
2. Prevention of algorithmic cartels
3. Protection against exclusionary platform algorithms
4. Maintenance of contestability and entry
5. Prevention of discriminatory access to essential data
6. Transparency sufficient for competition-law investigation
7. Protection of innovation
8. Prevention of strategic manipulation of machine-learning systems
9. Interoperability and portability where appropriate
10. Accountability for autonomous commercial systems

The objective is not to require every algorithm to be publicly disclosed. Excessive disclosure can itself weaken competition by revealing commercially sensitive technology.

III. Competition-Law Framework

A. Agreements and Concerted Practices

Traditional cartel provisions can apply where competitors use technology to implement an agreement.

The important principle is:

Technology does not immunise otherwise unlawful coordination.

An agreement to fix prices remains an agreement to fix prices even if the agreed prices are implemented automatically.

This is particularly important where competitors:

• use the same pricing provider;
• upload confidential pricing information;
• establish common pricing rules;
• instruct an algorithm to follow competitors;
• agree to maintain minimum prices; or
• use software to monitor compliance.

The U.S. authorities have expressly stated that companies cannot avoid antitrust law merely by implementing pricing coordination through algorithms.

IV. Algorithmic Collusion

Algorithmic collusion can take several forms.

1. Explicit algorithmic cartel

Competitors expressly agree to use algorithms to coordinate prices.

2. Hub-and-spoke coordination

A common software provider becomes the "hub", while competing businesses become the "spokes".

3. Information-mediated coordination

Competitors supply commercially sensitive information to the same algorithmic system.

4. Tacit algorithmic coordination

Separate algorithms observe one another and progressively learn that aggressive competition is less profitable.

5. Autonomous learning

AI systems independently discover a stable supra-competitive strategy.

The fifth scenario raises the difficult question of whether competition law should attach liability when there was no traditional human agreement at all.

V. Abuse of Dominance by Self-Learning Platforms

Self-learning systems can also create unilateral exclusionary conduct.

A dominant platform may train its algorithm using enormous quantities of market data and then use the resulting system to:

• rank its own products more favourably;
• demote competitors;
• restrict interoperability;
• disadvantage rival sellers;
• increase switching costs;
• favour affiliated businesses;
• discriminate in advertising auctions; or
• deny competitors access to commercially important data.

The Google Shopping litigation is particularly important because Google's ranking systems treated its own comparison-shopping service differently from competing services. The General Court characterised the conduct as an abuse involving favourable positioning of Google's own service and disadvantage to rival comparison-shopping services; the CJEU subsequently confirmed important aspects of that reasoning in 2024.

VI. Six Important Case Laws

1. Eturas UAB and Others v Lithuanian Competition Council — Case C-74/14

Court: Court of Justice of the European Union
Year: 2016

Facts

Eturas operated a common online travel-booking system used by numerous travel agencies.

The system administrator sent a message informing participating agencies that discounts offered through the system would be capped. The technical system was then modified to implement the restriction automatically.

Legal issue

Could participation in the computerised system amount to a concerted practice under Article 101 TFEU?

Decision

The CJEU held that, in appropriate circumstances, knowledge of the system administrator's anti-competitive communication combined with continued participation could support a presumption of participation in a concerted practice, subject to rebuttal.

Importance for self-learning markets

Eturas establishes an important technological principle:

A competition-law agreement or concerted practice does not cease to exist merely because a computer system implements the competitive restriction.

It is particularly relevant to:

• automated pricing;
• common platforms;
• algorithmic discount restrictions;
• digital marketplaces; and
• AI-mediated commercial decisions.

2. United States v David Topkins

Court: U.S. District Court, Northern District of California
Year: 2015

Facts

David Topkins and co-conspirators sold posters through Amazon Marketplace.

The conspirators agreed to fix prices and implemented their agreement through pricing algorithms. The algorithms were programmed to coordinate price changes.

The U.S. Department of Justice described the prosecution as its first criminal prosecution specifically targeting an online marketplace conspiracy.

Legal significance

The case demonstrated that:

An algorithm can be the instrument of a cartel without changing the underlying legal character of the conduct.

The important factor was not that algorithms were used, but that human actors had agreed upon an anti-competitive pricing strategy.

Relevance

Topkins is foundational for:

• algorithmic price fixing;
• automated cartel implementation;
• e-commerce;
• AI pricing systems; and
• platform-based markets.

3. Samir Agrawal v Competition Commission of India

Forum: Competition Appellate Tribunal/NCLAT
India

Facts

The complainant alleged that Ola and Uber used algorithmic pricing mechanisms that prevented drivers from independently negotiating fares and allegedly facilitated price coordination.

The Competition Commission of India did not find sufficient material to establish a prima facie contravention and closed the matter.

The subsequent appellate proceedings examined the allegation that algorithmic pricing could facilitate price fixing.

Importance

The case is particularly significant for Indian competition law because it demonstrates the distinction between:

Algorithmic pricing ≠ automatically illegal price fixing.

An algorithm can independently determine prices without constituting a cartel.

Competition authorities must therefore establish the necessary legal elements of coordination rather than infer illegality merely from the existence of automated pricing.

Principle

The case illustrates the importance of examining:

• control over the algorithm;
• relationship between platform and suppliers;
• contractual arrangements;
• communications;
• pricing autonomy;
• common intention; and
• actual competitive effects.

4. Google and Alphabet v European Commission — Google Shopping, T-612/17; C-48/22 P

Court: General Court / Court of Justice of the European Union
Important judgments: 2021 and 2024

Facts

Google used its general search engine to display its comparison-shopping service prominently while competing comparison-shopping services were subjected to ranking adjustments.

The Commission found that Google's specialised comparison-shopping service received favourable treatment.

The General Court upheld the core infringement finding, and the CJEU confirmed important aspects of the judgment in 2024.

Importance for self-learning markets

This case demonstrates that competition law can scrutinise algorithmic architecture itself.

The relevant concern was not merely a conventional contractual restriction. It involved:

• ranking algorithms;
• visibility;
• data traffic;
• search architecture;
• self-preferencing; and
• leveraging dominance from one market into another.

Principle

A dominant digital undertaking cannot necessarily treat its own service more favourably through its algorithm while subjecting rivals to competitive constraints from which its own service is exempt.

This has major implications for AI-powered:

• search engines;
• recommendation engines;
• marketplaces;
• app stores;
• advertising systems; and
• generative-AI ecosystems.

5. Cornish-Adebiyi v Caesars Entertainment

Court: U.S. District Court for the District of New Jersey
Algorithmic pricing litigation

Facts

The litigation concerned allegations involving hotel-room pricing and the use of algorithmic pricing systems.

The FTC and DOJ filed a statement of interest addressing the application of antitrust principles to algorithmic pricing.

The agencies argued that competitors cannot lawfully coordinate prices merely because the coordination is implemented through an algorithm rather than directly by employees.

Importance

The case is important because it addresses the increasingly common algorithm-provider model.

Suppose:

Hotel A + Algorithm X
Hotel B + Algorithm X
Hotel C + Algorithm X

If the common system incorporates competitively sensitive information and produces coordinated pricing recommendations, the fact that each hotel technically presses its own "accept" button does not necessarily resolve the competition issue.

Principle

The use of an intermediary algorithm cannot automatically eliminate the requirement to comply with Section 1 of the Sherman Act.

6. United States v RealPage

Court: U.S. District Court, Middle District of North Carolina
Filed: 2024

Facts

The DOJ and several state attorneys general brought proceedings concerning RealPage's algorithmic rental-pricing system.

The government alleged that competing landlords supplied non-public, competitively sensitive information to RealPage and that the resulting algorithmic system was used to generate rental-pricing recommendations.

The complaint alleged violations of Sections 1 and 2 of the Sherman Act.

Importance

RealPage is particularly significant because it moves beyond the simple question:

"Did two competitors agree on a price?"

toward the broader question:

Can an algorithmic information architecture itself facilitate coordinated conduct between competing firms?

This makes the case highly relevant to self-learning markets.

The litigation also demonstrates why competition authorities increasingly examine:

• training data;
• input data;
• information sharing;
• algorithmic recommendations;
• competitor communications;
• pricing parameters; and
• governance of third-party AI systems.

VII. Amazon Marketplace and Later Algorithmic Enforcement

The U.S. Amazon Marketplace prosecutions also provide an important extension of Topkins.

Several later defendants were prosecuted for price fixing involving Amazon Marketplace products, including DVD and Blu-Ray sales.

The significance is broader than the individual prosecutions:

A digital marketplace can become the technological environment through which traditional cartel conduct is implemented.

Therefore, competition authorities increasingly need to distinguish between:

legitimate automated competitive response

and

automated implementation of an underlying cartel agreement.

VIII. Amazon Buy Box and Algorithmic Self-Preferencing

The Italian Competition Authority's Amazon Marketplace proceedings are also important to understanding algorithmic governance.

The Buy Box system used algorithmic criteria in determining which seller received prominent placement. Competition authorities examined whether the system favoured sellers using Amazon's logistics services.

OECD material discussing the case notes the difficulty of investigating learning algorithms because regulators may have to infer system behaviour from controlled inputs and outputs rather than inspect every internal computational process.

This illustrates a fundamental problem:

The black-box problem

A regulator may observe:

Input → Algorithm → Output

without being able to fully observe:

Training data → model architecture → weights → reinforcement process → internal optimisation → output.

This creates an evidentiary challenge for competition authorities.

IX. Self-Learning Algorithms and Market Definition

Traditional market definition can become difficult in AI markets.

For example, an AI platform may simultaneously provide:

• search;
• advertising;
• recommendations;
• payments;
• cloud computing;
• data analytics;
• AI models.

The relevant market may therefore be:

A. Product-specific

One identifiable service.

B. Ecosystem-based

Multiple interconnected services.

C. Multi-sided

Different user groups interacting through the same platform.

D. Dynamic

The relevant competitive constraint changes as the technology develops.

Self-learning systems make this more complicated because a platform may rapidly improve its product through accumulated data.

X. Data as a Competitive Asset

In self-learning markets, data can function as a competitive input.

A dominant undertaking may possess:

• transaction data;
• behavioural data;
• search data;
• pricing data;
• location information;
• customer preferences;
• product-performance data;
• supplier data; and
• interaction histories.

The competitive advantage may therefore be represented as:

More users → more data → better model → better service → more users.

This creates a potential data-feedback loop.

Competition law may consequently need to consider whether control over data:

• creates barriers to entry;
• prevents interoperability;
• facilitates exclusion;
• strengthens dominance;
• enables discriminatory pricing;
• makes switching difficult; or
• permits superior algorithmic prediction.

XI. Network Effects and Machine-Learning Effects

Traditional digital markets already exhibit network effects.

Self-learning markets can produce an additional effect:

Machine-learning network effect

More users
↓
More data
↓
Better model
↓
Better predictions
↓
Better service
↓
More users

This can produce a reinforcing cycle.

A new entrant may therefore face a disadvantage even if it possesses technically sophisticated software because it lacks comparable training data.

Competition authorities may consequently have to examine:

• data accumulation;
• data portability;
• interoperability;
• access to public datasets;
• switching costs;
• data exclusivity; and
• whether incumbent data advantages are replicable.

XII. Personalized Pricing

Self-learning systems can estimate an individual's:

• willingness to pay;
• purchasing probability;
• price sensitivity;
• urgency;
• location;
• previous purchases;
• browsing behaviour.

The system can then potentially offer different prices to different consumers.

Personalization is not inherently anti-competitive.

However, concerns arise when personalized pricing is combined with:

• market power;
• exclusionary conduct;
• discriminatory treatment;
• exploitation of locked-in consumers;
• tying;
• foreclosure of competitors; or
• coordinated pricing.

Competition law therefore has to distinguish efficient price discrimination from conduct that weakens competitive constraints.

XIII. Algorithmic Discrimination Between Competitors

A self-learning platform may automatically decide which sellers receive:

• higher search rankings;
• advertising exposure;
• lower commissions;
• preferential delivery;
• access to consumers;
• recommendation slots; or
• premium placement.

If the platform is dominant, the algorithm may become a mechanism for exclusionary discrimination.

The Google Shopping litigation demonstrates the significance of differential algorithmic treatment of the dominant firm's own service versus rivals.

XIV. Algorithmic Refusal to Deal

A self-learning system may automatically deny or reduce access to competitors.

For example:

A dominant marketplace's AI determines that a rival seller's products should receive almost no visibility.

Competition-law questions could include:

1. Is the platform dominant?
2. Is the input or platform access indispensable?
3. Is the conduct objectively justified?
4. Does it foreclose efficient competitors?
5. Does it harm competition rather than merely an individual competitor?
6. Is there a less restrictive technological alternative?

The fact that the decision was produced automatically should not itself determine legality.

XV. Competition Governance and Explainability

Self-learning markets require a new form of competition-law explainability.

Traditional investigation asks:

"Who made the decision?"

AI markets may require:

"Which data, model, parameter, instruction, feedback mechanism and governance process produced the decision?"

A competition authority may therefore require access to:

• model documentation;
• audit logs;
• version histories;
• training-data categories;
• pricing inputs;
• ranking criteria;
• model updates;
• human interventions;
• system communications;
• API logs; and
• testing results.

The objective should be regulatory explainability, not necessarily complete public disclosure of source code.

XVI. Auditability

A self-learning platform should ideally maintain an auditable record of material commercial decisions.

For competition purposes, useful records can include:

XVII. Competition Governance of AI Agents

The next stage of self-learning markets may involve autonomous AI agents.

An AI agent could:

1. search suppliers;
2. negotiate prices;
3. select products;
4. place orders;
5. change prices;
6. monitor competitors;
7. renegotiate contracts; and
8. independently repeat the process.

This creates an important legal problem.

Who controls the competitive decision?

Possibilities include:

• the company;
• the software developer;
• the platform;
• the AI agent;
• the data provider; or
• multiple entities simultaneously.

Competition law will likely continue to attribute commercial conduct to human or corporate actors rather than treating AI as a separate legal person. The difficult issue is determining which undertaking exercised sufficient control or responsibility over the system.

XVIII. The Problem of Autonomous Collusion

Suppose:

• Firm A uses AI-A;
• Firm B uses AI-B;
• neither communicates with the other;
• both systems observe market prices;
• both systems learn that aggressive price cuts reduce profits;
• both independently raise prices;
• the systems maintain those prices.

The resulting market could resemble a cartel.

But traditional cartel law generally looks for:

Agreement + coordination + intention/knowledge

The central unresolved question becomes:

Can competition law respond effectively to anti-competitive coordination produced without conventional human communication?

This is one of the most important future questions in AI competition law.

XIX. Regulatory Responses

Competition governance can employ several mechanisms.

1. Ex-ante regulation

For systemically important digital platforms, regulation can impose obligations before competitive harm occurs.

Examples include:

• interoperability;
• data portability;
• restrictions on self-preferencing;
• transparency;
• non-discrimination;
• restrictions on combining data sets.

2. Ex-post enforcement

Traditional competition law remains important.

Authorities can investigate:

• cartels;
• abuse of dominance;
• exclusionary conduct;
• tying;
• refusal to deal;
• discriminatory access;
• information exchange;
• anti-competitive mergers.

3. Algorithmic audits

Authorities can test algorithms using controlled inputs and observe outputs.

This is particularly important when the internal system is difficult to inspect. The Amazon Buy Box experience illustrates the evidentiary difficulties of investigating systems that can learn and adapt.

4. Data governance

Competition authorities may examine:

• data exclusivity;
• data portability;
• access restrictions;
• interoperability;
• data pooling;
• commercially sensitive information exchange.

5. Merger control

AI markets can generate powerful data and learning advantages after mergers.

Authorities should examine whether a transaction creates:

Data + users + computing + model + distribution

combinations that substantially increase entry barriers.

XX. Indian Competition-Law Perspective

India's Competition Act, 2002 can address many self-learning-market problems through existing concepts.

Section 3

Relevant to:

• algorithmic cartels;
• coordinated pricing;
• information exchange;
• agreements restricting competition.

Section 4

Relevant to:

• algorithmic self-preferencing;
• discriminatory access;
• exclusionary ranking;
• refusal to provide access;
• leveraging;
• unfair conditions.

Sections 5 and 6

Relevant to:

• AI-sector mergers;
• data-driven acquisitions;
• platform acquisitions;
• acquisitions of emerging AI competitors.

Competition Commission of India

The CCI can therefore potentially investigate the economic substance of algorithmic conduct, rather than simply asking whether the conduct was manually performed.

The Samir Agrawal/Ola-Uber proceedings are particularly relevant because they show that automated pricing by itself is insufficient to establish an anti-competitive agreement; the legally required elements of coordination still have to be demonstrated.

XXI. Challenges for Competition Authorities

1. Black-box decision making

The authority may not know why the AI made a particular decision.

2. Rapid model evolution

An algorithm may behave differently six months after deployment.

3. Attribution

It can be difficult to identify the undertaking responsible for an autonomous outcome.

4. Evidence preservation

Relevant evidence may exist only in:

• logs;
• temporary model states;
• APIs;
• cloud infrastructure;
• training records.

5. False positives

Price convergence does not necessarily prove collusion.

6. False negatives

Traditional enforcement may miss sophisticated AI-enabled coordination.

7. Cross-border operation

An AI system can be trained in one country, hosted in another and affect consumers globally.

XXII. Pro-Competitive Functions of Self-Learning Systems

Competition governance should not assume that machine learning is inherently harmful.

Self-learning systems can:

• reduce transaction costs;
• identify cheaper suppliers;
• improve inventory management;
• reduce waste;
• lower search costs;
• improve product matching;
• increase price transparency;
• improve logistics;
• personalise products;
• encourage innovation.

Algorithmic pricing can also increase competitive responsiveness. The U.S. authorities have recognised that automated pricing can be entirely legitimate and potentially pro-competitive; the legal concern arises when it is used to implement or facilitate unlawful coordination.

Thus:

Algorithmic sophistication is not itself an antitrust offence.

XXIII. Emerging Doctrine: From Algorithmic Compliance to Algorithmic Governance

A future competition-compliance programme should not merely tell employees:

"Do not form a cartel."

It should also establish controls over the AI itself.

Algorithmic Competition Compliance Framework

1. Design

↓

Identify potential competition risks before deployment.

2. Data

↓

Prevent unauthorised use of competitors' confidential information.

3. Training

↓

Test whether training data can produce discriminatory or coordinated outcomes.

4. Deployment

↓

Establish human accountability.

5. Monitoring

↓

Continuously test outputs for suspicious patterns.

6. Audit

↓

Preserve logs and model versions.

7. Intervention

↓

Provide mechanisms to suspend problematic behaviour.

8. Review

↓

Reassess competition risks whenever the model materially changes.

XXIV. Comparative Case-Law Principles

XXV. Key Legal Principles

The emerging law of self-learning markets can be condensed into ten principles:

Principle 1

AI does not create an exemption from competition law.

Principle 2

Automated implementation does not eliminate an underlying agreement.

Principle 3

Algorithmic pricing is not automatically unlawful.

Principle 4

Independent algorithmic responses should not automatically be treated as collusion.

Principle 5

Shared competitively sensitive data can materially increase algorithmic-collusion risks.

Principle 6

Dominant platforms may face scrutiny when algorithms systematically favour their own services.

Principle 7

Competition authorities increasingly need algorithmic evidence, not merely traditional documents.

Principle 8

Model explainability and auditability can become important components of competition compliance.

Principle 9

Data advantages can become barriers to entry where they reinforce machine-learning advantages.

Principle 10

Competition governance must regulate anti-competitive outcomes without unnecessarily suppressing legitimate AI innovation.

Conclusion

Self-learning markets represent a shift from human-directed competition toward machine-mediated competition. The central competition-law problem is not the use of artificial intelligence itself, but the possibility that autonomous systems can transform pricing, ranking, information exchange, market access and competitive strategy.

The existing cases already establish several foundations. Eturas demonstrates that computerised systems can facilitate concerted practices; Topkins demonstrates that algorithms can implement conventional price-fixing agreements; Samir Agrawal illustrates that algorithmic pricing alone is not sufficient to prove cartelisation; Google Shopping demonstrates the importance of algorithmic self-preferencing in dominance analysis; while Cornish-Adebiyi and RealPage show the increasing focus on algorithm-mediated coordination.

The future of competition governance will therefore increasingly involve algorithmic auditing, data governance, model accountability, evidence preservation, interoperability, non-discrimination and continuous monitoring, alongside conventional Sections 3 and 4-type competition-law analysis.

The fundamental legal proposition remains: