Competition Law and Governance of Self-Evolving Competitive Systems

1. Introduction

A self-evolving competitive system is a market environment in which firms increasingly rely on algorithms, artificial intelligence, machine learning, automated pricing, recommendation engines, ranking systems, dynamic allocation tools and other technologies that learn from market conditions and modify their behaviour over time.

Unlike traditional business systems, where a human expressly determines each competitive decision, a self-evolving system may:

• continuously collect market data;
• observe competitors and consumers;
• modify prices automatically;
• optimise rankings and recommendations;
• learn from previous outcomes;
• alter its commercial strategy without a fresh human instruction;
• create feedback loops that reinforce an existing market position; and
• potentially converge toward similar conduct among competing firms.

Competition law therefore faces a fundamental question:

When competitive behaviour is increasingly produced by adaptive technological systems, how should competition law distinguish legitimate optimisation from unlawful coordination, exclusion or exploitation?

The issue is not that algorithms are inherently anti-competitive. Algorithms can produce substantial efficiencies. The difficulty arises when the architecture, data, incentives or learning process of the system reduces independent competitive decision-making.

2. Meaning of a Self-Evolving Competitive System

A conventional competitive system can be represented as:

Human decision → Business strategy → Market response

A self-evolving system may operate as:

Data → Algorithm → Prediction → Market action → New data → Learning → Modified strategy → New market action

This creates a continuous competitive feedback loop.

For example:

Competitor A changes price → Algorithm observes A → Algorithm changes B's price → A observes B → A's algorithm changes again → repeated interaction → price convergence.

The legal difficulty is determining whether this is:

1. legitimate independent adaptation;
2. conscious parallelism;
3. information exchange;
4. facilitated coordination;
5. a hub-and-spoke arrangement;
6. an agreement or concerted practice; or
7. autonomous algorithmic collusion.

3. Why Self-Evolving Systems Create Competition-Law Problems

A. Algorithmic coordination

Algorithms can observe competitors much faster than humans.

A pricing system can monitor:

• competitor prices;
• inventory;
• demand;
• discounts;
• promotions;
• customer behaviour;
• market capacity.

This can make deviations from a coordinated outcome easier to detect.

The UK Competition and Markets Authority has specifically identified the possibility that algorithms can facilitate explicit coordination, operate through common intermediaries and, in more difficult scenarios, potentially produce autonomous tacit coordination.

B. Self-learning collusion

The most difficult category is where competitors do not expressly communicate, but their algorithms independently learn that maintaining higher prices is profitable.

The distinction is important:

Human cartel

Human agreement → algorithm implements agreement.

Self-learning system

Algorithm interacts with market → algorithm discovers profitable response → repeated interaction → possible coordinated outcome.

Existing competition statutes are generally much clearer about the first situation than the second.

There remains substantial legal debate about whether purely autonomous algorithmic convergence, without an identifiable agreement or communication attributable to the firms, satisfies traditional requirements of cartel liability.

4. The Principle of Independent Competitive Decision-Making

A central principle is that competitors must ordinarily make their commercial decisions independently.

This becomes particularly important where firms:

• use common pricing software;
• exchange data through a platform;
• use common AI models;
• rely upon the same market intelligence provider;
• share competitively sensitive information;
• allow a third party to determine commercially important variables.

The more a firm delegates its competitive decision-making to a system that incorporates competitors' confidential information, the greater the competition-law risk.

5. Algorithmic Information Exchange

Self-evolving systems require data.

That creates another competition concern.

Potentially sensitive information includes:

• current prices;
• future prices;
• discounts;
• capacity;
• inventory;
• costs;
• demand forecasts;
• customer-specific information;
• production plans;
• strategic intentions.

The problem becomes particularly serious when competitors feed such information into the same intermediary or algorithmic system.

The legal theory can resemble a hub-and-spoke arrangement:

Competitor A → Sensitive information → Algorithmic hub ← Sensitive information ← Competitor B

The hub processes the information and supplies recommendations to both competitors.

6. Feedback Loops and Market Entrenchment

Self-evolving systems can create a particularly important phenomenon:

Data → Scale → Better algorithm → Better service → More users → More data → Greater scale

This is a data-driven competitive feedback loop.

It can generate legitimate efficiencies, but it can also reinforce market power.

A dominant platform may acquire:

1. more users;
2. more data;
3. better predictions;
4. better recommendations;
5. greater visibility;
6. more users.

This can increase barriers to entry even without a conventional exclusionary contract.

Competition analysis therefore increasingly needs to examine dynamic competitive processes, rather than merely static market shares.

7. Self-Preferencing by Adaptive Systems

A self-evolving platform may continuously optimise its ranking algorithm.

Suppose a dominant platform controls:

• search;
• marketplace access;
• advertising;
• payments;
• logistics.

Its algorithm may learn that placing its own product prominently increases platform revenue.

The resulting system can progressively favour affiliated products.

The important question becomes whether the system is:

• competing on the merits; or
• exploiting control over an essential gateway to disadvantage rivals.

The Google Shopping litigation is particularly significant here.

8. Major Case Laws

1. Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba, Case C-74/14 (CJEU, 2016)

Facts

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

A system message announced a restriction on the discounts that agencies could provide. The system was technically modified so that the restriction became operational.

Legal issue

Could use of a common computerised system contribute to establishing a concerted practice even where the competitors did not directly communicate with one another?

Decision

The CJEU held that, where participating operators were aware of the system message and the technical implementation, they could be presumed to have participated in a concerted practice, subject to rebuttal.

They could rebut the presumption by, among other things:

• publicly distancing themselves;
• reporting the conduct; or
• demonstrating contrary market behaviour.

The Court also stressed that the mere sending of a message was not automatically sufficient proof that every undertaking knew of it.

Importance for self-evolving systems

Eturas demonstrates that technology can become the mechanism through which competitive coordination occurs.

The case is particularly relevant to:

• common platforms;
• automated restrictions;
• algorithmic communications;
• shared software;
• digital evidence.

It shows that competition law does not necessarily require competitors to sit together physically.

2. United States v. David Topkins, No. CR 15-00201 (N.D. Cal., 2015)

Facts

Online sellers of posters agreed to coordinate prices on Amazon Marketplace.

They subsequently used pricing algorithms to implement their agreement.

Decision

The conduct was prosecuted as horizontal price fixing under §1 of the Sherman Act.

The DOJ's charging document specifically alleged that the conspirators agreed to adopt particular pricing algorithms and that code was used to implement the pricing arrangement.

Importance

Topkins establishes a crucial principle:

Using software to implement an unlawful agreement does not make the agreement lawful.

The algorithm was essentially the mechanism of cartel implementation.

It therefore represents the relatively straightforward category:

Human agreement → Algorithm → Collusive prices

It does not, however, conclusively resolve the much harder question of purely autonomous AI collusion.

3. Trod Ltd and GB eye Ltd, CMA Case 50223 (UK, 2016)

Facts

Two online sellers of posters and frames competed on Amazon's UK website.

They agreed not to undercut one another.

The agreement was implemented through automated repricing software. The software was configured so that the competitors would not automatically undercut one another.

Outcome

The CMA found an infringement.

Trod received a fine of £163,371, while GB eye received immunity after reporting the cartel and cooperating with the investigation.

Importance

The case illustrates the algorithm-as-messenger model:

Human cartel → automated repricing → continuous enforcement of cartel.

It also demonstrates why competition authorities increasingly need to examine:

• software configuration;
• source rules;
• "ignore" lists;
• pricing parameters;
• communication between software users and providers;
• algorithmic audit trails.

4. Google LLC and Alphabet Inc. v European Commission (Google Shopping), Case C-48/22 P (CJEU, 2024)

Facts

Google operated a dominant general search service and also operated a specialised comparison-shopping service.

Its ranking and display mechanisms gave its own comparison-shopping results preferential visibility while competing services were disadvantaged.

Decision

The CJEU dismissed Google's appeal and upheld the finding of abuse of dominant position. The case concerned Google's use of its dominant general-search position to favour its own specialised comparison-shopping service.

Importance

This is highly significant for self-evolving systems because the competitive effect can be embedded in ranking architecture rather than a conventional contractual restriction.

The competition concern can arise through:

• ranking;
• visibility;
• default positioning;
• recommendation;
• search architecture;
• algorithmic demotion.

Thus:

Algorithmic design can itself become a competitive instrument.

5. United States v. RealPage, Inc.

Facts

The US Department of Justice alleged that RealPage's rental-pricing software used competitively sensitive information supplied by competing landlords to generate pricing recommendations.

The DOJ alleged that the system contributed to alignment of rental prices and also alleged monopolisation of the revenue-management software market.

In November 2025, the DOJ announced a proposed settlement requiring RealPage to stop using competitively sensitive information in the relevant pricing system and to end practices designed to align pricing among competing landlords.

Importance

RealPage is particularly important because it illustrates a more advanced form of algorithmic coordination:

Competitors → Sensitive data → Common algorithm → Pricing recommendations → Market outcomes

It therefore raises issues beyond traditional cartel meetings.

It demonstrates the importance of examining:

• data inputs;
• algorithmic rules;
• recommendation mechanisms;
• auto-accept functionality;
• information-sharing arrangements;
• feedback loops.

6. FTC and States v. Amazon.com, Inc.

Facts

The US Federal Trade Commission and state authorities sued Amazon alleging that Amazon used a collection of practices to maintain monopoly power and restrict competition.

The case includes allegations concerning Amazon's treatment of sellers and mechanisms affecting prices and competitive conditions on its marketplace.

Importance for self-evolving systems

The Amazon litigation demonstrates how competition analysis can extend beyond a single algorithm.

A platform can combine:

• search ranking;
• Buy Box mechanisms;
• seller policies;
• advertising;
• pricing systems;
• marketplace access;
• data collection;
• fulfilment.

The result can be an ecosystem-level competitive system in which several automated mechanisms interact.

Competition law therefore increasingly has to examine the combined architecture rather than isolated algorithmic decisions.

7. Samir Agarwal v. Competition Commission of India, (2021) 3 SCC 136

This Indian Supreme Court decision is relevant to the broader issue of platform-mediated coordination.

The case concerned allegations involving cab aggregators and the possibility of coordination facilitated through a common technological intermediary.

The Supreme Court considered the requirements for establishing an agreement or concerted arrangement under Indian competition law and rejected the idea that the mere presence of a common technological intermediary automatically establishes a cartel.

Importance

The case is particularly useful for self-evolving systems because it reinforces an important distinction:

Common technology or similar algorithmic outcomes do not automatically establish an unlawful agreement.

There must still be legally sufficient evidence connecting the technological arrangement with prohibited coordination.

This principle becomes critical where two independent AI systems happen to produce similar market outcomes.

9. Comparative Lessons from the Cases

10. Algorithmic Collusion: Four Important Models

Model 1 — Messenger Algorithm

Human agreement → Algorithm implements it

Example:

Topkins

This is the easiest category for traditional competition law.

Model 2 — Hub-and-Spoke Algorithm

Firm A → Data → Common algorithm ← Data ← Firm B

The intermediary processes information and provides recommendations to both.

Example:

RealPage

The critical issues are:

• what data is shared;
• whether the information is competitively sensitive;
• who controls the algorithm;
• whether firms know how the system operates;
• whether recommendations align competitive behaviour.

Model 3 — Predictable-Agent Algorithm

Each firm independently programs an algorithm to react to competitors.

Example:

"If competitor decreases price by 5%, immediately reduce our price by 5%."

There may be no direct communication.

The competition concern is whether repeated automated reactions substantially reduce competitive uncertainty.

Model 4 — Autonomous Self-Learning Algorithm

The system is not expressly programmed to collude.

Instead:

Observe → Experiment → Learn → Adjust → Observe → Learn

Eventually it may discover that aggressive competition produces lower profits while coordinated behaviour produces higher returns.

This is the most conceptually difficult category.

Current competition law has not yet produced a universally accepted rule that every purely autonomous machine-generated coordination outcome constitutes an unlawful agreement. The existing cases overwhelmingly provide stronger legal foundations where there is evidence of human communication, common systems, information sharing or deliberate algorithmic design.

11. The Role of Market Definition

Self-evolving systems make traditional market definition more difficult.

Competition authorities may need to examine several dimensions simultaneously:

Product dimension

• physical product;
• digital service;
• platform;
• algorithmic service;
• data service.

User dimension

• consumers;
• sellers;
• advertisers;
• developers;
• suppliers.

Temporal dimension

A system that changes rapidly may make historical market shares less informative.

Ecosystem dimension

A platform may operate simultaneously in:

• search;
• advertising;
• payments;
• cloud;
• commerce;
• logistics;
• AI.

Therefore, ecosystem competition can become more significant than a single conventional product market.

12. Dynamic Market Power

Traditional analysis often asks:

What is the firm's market share today?

Self-evolving systems require additional questions:

• How quickly can competitors reproduce the system?
• Does the system improve with additional data?
• Are there network effects?
• Are switching costs increasing?
• Does the algorithm become better as its user base grows?
• Can rivals access equivalent data?
• Is interoperability available?
• Can users multi-home?
• Can competitors challenge the algorithm's output?

This shifts competition law from static market power toward dynamic market power.

13. Data as a Competitive Asset

Data can become an input into self-evolving competition.

The competition authority may therefore examine:

Data accumulation

Does a dominant firm accumulate substantially more data than competitors?

Data exclusivity

Are rivals prevented from obtaining equivalent information?

Data portability

Can users transfer their data?

Data interoperability

Can competing systems interact?

Data feedback

Does more usage automatically make the dominant algorithm better?

Data foreclosure

Does control over data prevent competitors from developing comparable systems?

14. Algorithmic Self-Preferencing

A dominant platform's algorithm may favour:

• its own products;
• affiliated services;
• its advertising inventory;
• its payment service;
• its logistics network;
• its AI assistant.

This can be problematic where the platform is simultaneously:

Market infrastructure + competitor + rule-maker + algorithm designer.

Google Shopping demonstrates how algorithmic positioning can become an important component of an abuse-of-dominance analysis.

15. Competition Law and Explainability

A self-evolving system can create a major evidentiary problem.

A regulator may ask:

"Why did the algorithm make this decision?"

The firm may respond:

"The model learned this behaviour from the data."

This cannot automatically end the inquiry.

Competition authorities may need access to:

• model documentation;
• training data;
• logs;
• input variables;
• output records;
• version histories;
• parameter changes;
• human overrides;
• system instructions;
• audit trails.

Thus, algorithmic accountability becomes an evidentiary component of competition enforcement.

16. Algorithmic Audit Trails

A major governance mechanism should be preservation of:

1. Input logs

What information entered the system?

2. Decision logs

What recommendation did the system produce?

3. Version logs

Which model was operating at the relevant time?

4. Configuration logs

Who changed the parameters?

5. Override logs

Did humans accept or reject the recommendation?

6. Data-source logs

Did competitor information enter the model?

7. Communication logs

Did firms communicate regarding the system?

These records can help distinguish:

independent optimisation

from

algorithmically facilitated coordination.

17. Compliance Governance for Self-Evolving Systems

Businesses should implement an algorithmic competition compliance framework.

A. Pre-deployment competition assessment

Before deploying an algorithm, assess:

• competitors' data;
• pricing functions;
• ranking functions;
• recommendation rules;
• information flows;
• market concentration;
• potential coordination mechanisms.

B. Sensitive-data controls

Competitively sensitive information should be segregated.

Particular caution is necessary where the system:

• receives competitor prices;
• receives future pricing intentions;
• processes competitor capacity;
• recommends prices for multiple competitors;
• uses non-public competitor information.

C. Human oversight

Important competitive decisions should have meaningful human oversight.

However, human involvement should not merely become a rubber stamp for an algorithmically generated anti-competitive decision.

D. Independent algorithmic audits

Periodic audits should investigate:

• unexplained price convergence;
• systematic exclusion;
• self-preferencing;
• discriminatory rankings;
• competitor-sensitive inputs;
• unusual market responses;
• repeated algorithmic reactions.

E. Kill-switch mechanisms

High-risk systems should have mechanisms allowing immediate suspension where the system begins producing potentially unlawful outcomes.

18. Remedies

Competition authorities may employ several remedies.

Structural remedies

• divestiture;
• separation of business units;
• functional separation.

Behavioural remedies

• non-discrimination;
• interoperability;
• data access;
• prohibition of sensitive-data use;
• restrictions on self-preferencing.

Algorithmic remedies

• model modification;
• deletion of prohibited variables;
• independent auditing;
• certification;
• monitoring;
• source-data restrictions.

The RealPage settlement illustrates the emergence of remedies directed specifically at algorithmic coordination, including restrictions on competitively sensitive information and certain pricing-alignment practices.

19. Competition Law in India

For India, the principal framework is the Competition Act, 2002.

Self-evolving systems can potentially engage several provisions, depending upon the facts.

Section 3

Relevant where algorithmic systems facilitate:

• price fixing;
• market sharing;
• limitation of production;
• information exchange;
• other anti-competitive agreements.

Section 4

Relevant where a dominant digital platform uses an adaptive system to:

• unfairly discriminate;
• restrict market access;
• leverage dominance;
• foreclose competitors;
• favour affiliated services.

Sections 5 and 6

Particularly relevant where self-evolving ecosystems become involved in mergers and acquisitions.

An acquisition may strengthen a data or algorithmic feedback loop even where the acquired company has relatively modest current revenues.

20. Merger Control and Self-Evolving Systems

Traditional merger analysis may focus heavily upon:

• turnover;
• assets;
• market shares;
• current competitive overlap.

But AI-driven systems require examination of:

• datasets;
• training infrastructure;
• computing capacity;
• proprietary models;
• developer ecosystems;
• user networks;
• interoperability;
• switching costs;
• access to distribution.

A small AI firm may therefore possess strategically important assets even when conventional revenue-based measurements appear modest.

21. Essential-Facility and Interoperability Issues

A self-evolving ecosystem may become difficult for rivals to challenge where the dominant firm controls:

• an API;
• operating system;
• app store;
• cloud infrastructure;
• payment infrastructure;
• identity system;
• data access;
• search distribution.

The Google Shopping litigation demonstrates the importance of analysing access and leveraging where a dominant infrastructure is used to advantage a downstream service.

22. The Special Problem of Continuous Learning

Traditional competition enforcement often analyses a completed act:

Agreement → Conduct → Effect.

A self-evolving system instead produces:

Conduct₁ → Learning₁ → Conduct₂ → Learning₂ → Conduct₃ → Learning₃

This means that the relevant anti-competitive conduct may not exist as a single decision.

It may emerge from thousands of small automated adjustments.

Therefore, enforcement methodology must increasingly consider:

• longitudinal data;
• model evolution;
• repeated interactions;
• behavioural trajectories;
• feedback effects.

23. Causation

A particularly difficult issue is proving causation.

Suppose prices rise simultaneously across five firms.

That does not automatically prove algorithmic collusion.

The increase could result from:

• increased demand;
• common costs;
• supply shortages;
• inflation;
• independent algorithmic optimisation;
• common public information;
• or unlawful coordination.

Accordingly, competition authorities need to distinguish correlation from causation.

The Google Shopping judgment illustrates the importance of analysing causal links between algorithmic conduct and competitive effects rather than treating the existence of an algorithm as sufficient proof of harm.

24. Standard of Proof

A useful evidentiary framework is:

Level 1 — Algorithmic similarity

Two firms use similar algorithms.

Normally insufficient by itself.

Level 2 — Similar outcomes

Both algorithms produce similar prices.

Still not necessarily sufficient.

Level 3 — Common data

Both algorithms receive sensitive competitor information.

Greater competition concern.

Level 4 — Common intermediary

Both firms use the same intermediary to generate recommendations.

Potential hub-and-spoke issue.

Level 5 — Communication or knowledge

Evidence shows firms knew about the coordinated system.

Much stronger basis for enforcement.

Level 6 — Intentional design

Firms deliberately configured algorithms to coordinate.

Strong evidence of algorithm-enabled collusion.

Topkins and Trod illustrate the significance of evidence showing deliberate configuration of algorithms to implement a pre-existing agreement.

25. Governance Model

A comprehensive regulatory framework for self-evolving competitive systems can be represented as:

Data governance
↓
Algorithmic design review
↓
Competition-risk assessment
↓
Deployment
↓
Continuous monitoring
↓
Audit logs
↓
Competition-impact testing
↓
Human intervention where necessary
↓
Regulatory reporting
↓
Corrective measures

The key change is that competition compliance cannot be treated as a one-time pre-deployment exercise.

It must become continuous.

26. Key Legal Principles Emerging from the Case Law

Principle 1

Code does not replace competition law.

Topkins demonstrates that an algorithm cannot legitimise a cartel.

Principle 2

A common technological infrastructure can facilitate coordination.

Eturas demonstrates the relevance of a common computerised system.

Principle 3

Automated repricing can constitute implementation of a cartel.

Trod/GB eye demonstrates this clearly.

Principle 4

Algorithmic ranking can produce exclusionary effects.

Google Shopping establishes the importance of self-preferencing through ranking and display mechanisms.

Principle 5

Competitively sensitive data can make algorithmic systems substantially more problematic.

RealPage demonstrates the importance of examining the data supplied to pricing systems as well as the resulting output.

Principle 6

Similar technological outcomes do not automatically prove an agreement.

The legal analysis must still establish the necessary elements of the relevant competition-law infringement.

Principle 7

The competitive effects of adaptive systems must be assessed dynamically.

The relevant question may concern how the system evolves over time rather than merely what it did on a particular day.

27. Challenges for Competition Authorities

Competition authorities face several difficulties:

1. Black-box decision-making
2. Rapid model evolution
3. Massive quantities of data
4. Difficulty identifying human intent
5. Cross-border deployment
6. Common third-party algorithms
7. Complex feedback loops
8. Difficulty reproducing algorithmic outcomes
9. Rapidly changing markets
10. Potentially simultaneous legitimate and anti-competitive effects

These challenges make conventional evidence-gathering methods less effective.

28. Future Direction of Competition Law

The future regulatory model is likely to move toward continuous algorithmic governance.

Important tools may include:

• algorithmic audits;
• regulatory sandboxes;
• model documentation;
• mandatory data governance;
• independent technical monitoring;
• competition impact assessments;
• interoperability obligations;
• data portability;
• real-time market monitoring;
• algorithmic transparency;
• competition-by-design requirements.

The CMA's recent work specifically recognises that AI and increasingly sophisticated algorithms are reshaping pricing and other competitive decisions, while also creating new enforcement and compliance challenges.

29. Conclusion

Self-evolving competitive systems represent a transition from competition conducted primarily by humans to competition increasingly mediated by continuously adapting computational systems.

The existing case law already provides several important foundations.

Eturas demonstrates the significance of common digital infrastructure.
Topkins shows that algorithms cannot lawfully implement human cartels.
Trod/GB eye demonstrates automated enforcement of price coordination.
Google Shopping establishes the competition significance of algorithmic self-preferencing and ranking.
RealPage demonstrates the risks created when competing firms' sensitive information is processed through a common pricing system.
Amazon illustrates how several platform mechanisms may combine into an ecosystem-level competition problem.
Samir Agarwal provides an important Indian reference point concerning the evidentiary requirements for establishing technologically facilitated coordination.

The central regulatory challenge is therefore not simply "whether AI is anti-competitive." It is to determine who controls the system, what information it receives, how it learns, what incentives shape its behaviour, how its decisions affect rivals, and whether its evolution reduces independent competitive decision-making.

In this sense, the future of competition law will increasingly require a combination of traditional antitrust doctrine + data governance + algorithmic auditing + dynamic market analysis + continuous regulatory oversight.