Competition Law And Competition Concerns In Self-Learning Economies

1. Introduction

A self-learning economy may be understood as an economic environment in which important commercial decisions are increasingly made or influenced by artificial intelligence, machine-learning systems, automated platforms, recommendation engines, dynamic-pricing software and other algorithms that continuously learn from data.

Traditional businesses normally change prices, production levels or marketing strategies through human decisions. In a self-learning economy, an algorithm may observe millions of transactions, competitor prices and consumer responses and then automatically change commercial behaviour.

Examples include:

• dynamic pricing systems;
• online marketplaces;
• search engines;
• recommendation algorithms;
• automated advertising systems;
• algorithmic credit or risk assessment;
• digital platforms matching buyers and sellers;
• automated procurement systems;
• AI-based supply-chain management; and
• personalised pricing or commercial offers.

These technologies can substantially increase efficiency. They can reduce transaction costs, improve matching between buyers and sellers, predict demand and make services more personalised.

However, the same characteristics can create serious competition-law concerns. Algorithms may facilitate coordination, dominant platforms may control essential datasets, recommendation systems may favour affiliated products, and network effects may make entry increasingly difficult.

Therefore, competition law in a self-learning economy must distinguish innovation and legitimate algorithmic optimisation from conduct that restricts competition.

2. Competition Law Framework

In the European Union, two particularly important provisions are Articles 101 and 102 of the Treaty on the Functioning of the European Union.

Article 101 TFEU

Article 101 deals principally with agreements, decisions by associations of undertakings and concerted practices that restrict competition.

It may therefore become relevant where competitors coordinate through:

• common pricing software;
• shared algorithms;
• common data systems;
• information exchanges;
• algorithm providers;
• digital marketplaces; or
• automated communication mechanisms.

The difficult question is whether autonomous algorithms that independently learn to coordinate prices can satisfy the legal requirements of an agreement or concerted practice.

Article 102 TFEU

Article 102 addresses abuse of a dominant position.

It is particularly significant in self-learning markets because powerful digital firms may control:

• large datasets;
• algorithms;
• search systems;
• operating systems;
• digital ecosystems;
• marketplaces;
• advertising infrastructure;
• application stores; and
• access to consumers.

Dominance itself is not prohibited. Competition law becomes concerned where dominant power is used abusively to exclude competitors or otherwise distort competition.

3. Algorithmic Collusion

One of the most important concerns is algorithmic collusion.

Imagine several competing retailers using algorithms that continuously observe each other's prices.

Normally competition should encourage each retailer to reduce prices where doing so is profitable. But sufficiently sophisticated learning systems might discover through repeated interaction that aggressive price competition reduces everyone's profits.

Algorithms could therefore potentially learn that maintaining higher prices produces better long-term outcomes.

The concern becomes particularly important when:

• only a small number of firms operate in the market;
• prices are highly transparent;
• transactions occur frequently;
• algorithms can respond rapidly to competitors;
• market conditions are relatively stable; and
• competitors use similar algorithms or the same algorithm provider.

Economic research has demonstrated that certain reinforcement-learning algorithms can learn strategies producing prices above competitive levels in experimental environments.

The legal difficulty is attribution.

Traditional cartel law generally looks for some form of agreement or coordination between undertakings. If two genuinely independent algorithms separately learn parallel strategies without communication between the businesses, establishing an Article 101 concerted practice may be substantially harder.

4. Hub-and-Spoke Algorithmic Coordination

Another problem arises when several competitors use the same algorithm provider.

Suppose companies A, B, C and D independently sell the same product but all employ one external company to determine prices automatically.

The software provider effectively becomes the "hub", while the competing sellers are the "spokes".

If the arrangement is deliberately structured so that the central system coordinates competitive behaviour, ordinary cartel principles can potentially apply.

Competition authorities therefore need to examine:

• who designed the algorithm;
• who supplied its objectives;
• whether competitors knew how it operated;
• what information was exchanged;
• whether competitors knowingly accepted coordination; and
• whether the system merely optimised independent decisions or facilitated collective behaviour.

5. Autonomous Algorithmic Coordination

The harder situation involves genuinely autonomous systems.

Assume:

• Firm A uses Algorithm A;
• Firm B uses Algorithm B;
• there is no communication;
• there is no common software provider; and
• neither company instructs its system to coordinate.

After millions of repeated interactions, however, both algorithms learn that matching higher prices produces greater profits.

This creates a major enforcement problem.

Economically, the outcome could resemble tacit collusion. Legally, however, proving an agreement or concerted practice may be difficult because conscious communication between the undertakings may be absent.

This demonstrates a possible gap between traditional competition concepts and increasingly autonomous markets.

6. Algorithmic Price Discrimination

Self-learning systems can analyse large quantities of consumer information.

An algorithm may consider:

• purchasing history;
• location;
• browsing patterns;
• previous searches;
• device characteristics;
• responsiveness to previous prices; and
• estimated willingness to pay.

Businesses can therefore potentially offer different consumers different prices.

Not every form of personalised pricing violates competition law. However, competition concerns may arise where such practices reinforce market power, discriminate in circumstances covered by competition rules, exclude competitors or exploit a dominant position.

Competition law must therefore distinguish ordinary commercial price differentiation from conduct associated with exclusionary or exploitative market power.

7. Data As A Competitive Advantage

Self-learning algorithms depend heavily on data.

More users can produce more data. More data can improve an algorithm. A better algorithm can attract additional users, which generates still more data.

The process can therefore become:

More users → more data → better algorithm → better service → more users

This feedback mechanism may generate substantial competitive advantages.

A new competitor might possess excellent technology but still struggle because it lacks sufficient training data.

Competition authorities consequently need to consider whether control over data creates substantial barriers to entry.

8. Network Effects

Many self-learning digital markets also display strong network effects.

A service may become more valuable as additional people use it.

For example, a platform attracting many buyers becomes attractive to sellers. More sellers then attract additional buyers.

This creates a feedback loop:

Users → data → improved service → more users → stronger network → additional data

Network effects are not inherently anticompetitive. They can make services more useful and efficient.

But when combined with switching costs, data advantages and ecosystem control, they may make successful entry substantially harder.

9. Self-Preferencing By Digital Platforms

A platform may simultaneously:

1. operate the marketplace or intermediary service; and
2. compete with businesses using that marketplace.

Its algorithms may determine:

• search rankings;
• recommendations;
• visibility;
• advertising placement;
• product rankings; and
• access to customers.

A competition concern can therefore arise if a dominant platform systematically gives preferential treatment to its own services in circumstances amounting to abusive conduct.

Algorithmic self-preferencing is especially difficult to identify because the ranking mechanism can involve thousands of signals and frequent automated modifications.

10. Recommendation-System Competition

Recommendation systems increasingly determine which products consumers discover.

Examples include systems recommending:

• products;
• applications;
• restaurants;
• videos;
• music;
• financial services; or
• professional services.

A dominant recommendation system may therefore become an important gateway between businesses and consumers.

Potential concerns include:

• discriminatory ranking;
• suppression of competitors;
• preferential treatment of affiliated products;
• manipulation of visibility;
• discriminatory access to recommendation data; and
• making businesses purchase additional services to obtain meaningful visibility.

The competitive significance of ranking algorithms therefore depends on both market power and the specific conduct involved.

11. Algorithmic Tying And Bundling

AI ecosystems frequently combine several services.

A company may provide:

• an operating system;
• an AI assistant;
• cloud infrastructure;
• an app marketplace;
• search;
• advertising; and
• data analytics.

Integration can produce genuine efficiencies.

However, where a dominant company conditions access to an important product on acceptance of another service, competition-law questions concerning tying or bundling may arise.

The central inquiry is whether the arrangement improperly forecloses competing suppliers rather than merely providing an integrated and better product.

12. Switching Costs And Algorithmic Lock-In

Self-learning services often become personalised over time.

The system learns:

• preferences;
• transaction history;
• workflows;
• contacts;
• purchasing patterns; and
• behavioural characteristics.

Consequently, leaving the service may mean losing years of accumulated personalisation.

This can create algorithmic lock-in.

Even where a competitor offers superior technology, users may hesitate to move because their accumulated data, history and personalised model cannot easily be transferred.

Interoperability and data portability can therefore become important competitive considerations.

13. Predatory Algorithmic Strategies

Algorithms can potentially implement exclusionary strategies much faster than human managers.

For example, a powerful platform could theoretically identify a new competitor and automatically react through aggressive pricing or other commercial responses.

Low prices are normally beneficial to consumers and are not automatically unlawful.

Competition concerns arise only when the relevant legal requirements for predatory or exclusionary conduct are satisfied.

Authorities therefore have to examine the economic context rather than treating automated price reductions themselves as evidence of abuse.

14. Information Exchange Through Algorithms

Competition can also be harmed where businesses feed commercially sensitive information into common systems.

Potentially sensitive information includes:

• future prices;
• planned output;
• inventories;
• capacity;
• future promotions;
• strategic plans; and
• customer-specific information.

An algorithmic intermediary could aggregate this information and indirectly influence participating businesses.

Competition authorities must therefore examine whether the technology merely processes legitimate information or becomes a mechanism for reducing strategic uncertainty between competitors.

15. Acquisition Of Innovative AI Competitors

Dominant technology companies frequently acquire smaller firms.

Many acquisitions are legitimate and can help innovations reach consumers.

However, competition authorities may examine whether an acquisition removes an emerging competitive threat.

This issue can be particularly important in self-learning markets because a start-up may have:

• few customers;
• limited current revenue; but
• valuable technology, data or intellectual property.

Traditional measures such as current revenue or market share may therefore provide an incomplete picture of future competitive significance.

16. Explainability And Competition Enforcement

Advanced machine-learning models can be difficult to interpret.

A company itself may not always be able to explain precisely why a complex system generated a particular outcome.

This creates problems for enforcement.

Authorities may need to determine whether:

• discriminatory ranking was deliberate;
• pricing coordination resulted from programming;
• an algorithm excluded competitors;
• particular data influenced the decision; or
• a system independently developed its behaviour.

Competition investigations may therefore increasingly depend on technical evidence such as:

• source code;
• model documentation;
• training objectives;
• datasets;
• algorithmic logs;
• testing records; and
• internal communications.

17. Case Laws Relevant To Self-Learning Economies

There is still limited case law dealing specifically with autonomous self-learning algorithms. Therefore, existing digital-market and competition cases provide the legal principles most relevant to future AI markets.

Case 1: Eturas UAB and Others v Lithuanian Competition Council

Case C-74/14, Court of Justice of the European Union, 2016

This is one of the most important cases for understanding technologically facilitated coordination.

Travel agencies participated in a common online booking system. A technical restriction was introduced concerning discounts available through the system.

The Court considered circumstances in which participating businesses could be regarded as taking part in a concerted practice.

Importance

Eturas demonstrates that competition law does not require competitors to meet physically in order for coordination to arise.

A common digital system can facilitate coordination.

However, participation cannot simply be presumed without appropriate evidence concerning awareness and conduct.

Relevance to self-learning economies

The case provides an important foundation for analysing:

• common pricing algorithms;
• shared AI platforms;
• common optimisation software; and
• algorithmic hub-and-spoke arrangements.

The central issue remains whether sufficient evidence establishes participation in coordinated conduct.

18. Case 2: VM Remonts and Others

Case C-542/14, Court of Justice of the European Union, 2016

The case concerned whether an undertaking could be responsible for anticompetitive conduct resulting from the activities of an independent service provider.

The Court explained circumstances in which the conduct of such an external provider may be attributed to the undertaking.

Importance

Modern companies increasingly outsource important decisions to:

• software developers;
• algorithm vendors;
• cloud providers;
• consultants; and
• AI service providers.

Simply outsourcing a function does not necessarily eliminate competition-law responsibility.

Application to AI

Suppose several retailers employ an external algorithmic pricing provider.

If that provider coordinates commercially sensitive information or pricing strategies, questions can arise regarding the responsibility of the participating businesses.

VM Remonts therefore provides an important conceptual foundation for determining responsibility where competitive decisions are delegated to technology providers.

19. Case 3: Google Shopping

Google and Alphabet v Commission, Case C-48/22 P

The European Commission found that Google had favoured its own comparison-shopping service in general search results while competing comparison-shopping services could be demoted through Google's ranking mechanisms.

The Court of Justice ultimately upheld the central finding concerning the abuse and the approximately €2.4 billion fine.

Importance

This case is particularly relevant because competition increasingly takes place through rankings rather than traditional shelf space.

In a self-learning economy, an algorithm can decide:

• which business appears first;
• which product receives visibility;
• which competitor is recommended; and
• which service effectively disappears from consumer attention.

Principle

A dominant platform's algorithmic architecture can therefore become relevant under competition law when its operation departs from competition on the merits and is capable of producing exclusionary effects.

20. Case 4: Google Android

Google and Alphabet v Commission, Case C-738/22 P

The Android proceedings concerned contractual restrictions connected with Google's mobile ecosystem.

Among other matters, the case involved pre-installation conditions involving Google Search, Chrome and Play Store, as well as restrictions affecting alternative versions of Android.

In July 2026, the Court of Justice dismissed Google's appeal and confirmed a fine of approximately €4.1 billion.

Importance

The case demonstrates how competition problems can arise from control over a technological ecosystem rather than from a single isolated product.

Application to self-learning economies

Future AI ecosystems may similarly combine:

• AI models;
• operating systems;
• app stores;
• cloud computing;
• search;
• browsers;
• personal assistants; and
• advertising.

A dominant ecosystem could potentially use control over one layer to strengthen its position in another.

Google Android therefore provides important principles for analysing tying, ecosystem foreclosure and restrictions on competing technological architectures.

21. Case 5: Meta Platforms v Bundeskartellamt

Case C-252/21, Court of Justice of the European Union, 2023

The case involved Meta's processing and combination of user data from different sources.

The Court held, among other things, that a national competition authority examining an abuse of dominance may consider compliance with data-protection rules where relevant to the competition assessment, while respecting the responsibilities of data-protection authorities.

Importance

The judgment illustrates the increasingly important relationship between:

data + privacy + market power + competition.

Application to self-learning systems

Machine-learning models become more effective when supplied with extensive data.

A dominant company capable of combining enormous datasets may obtain advantages unavailable to smaller competitors.

Therefore, data collection cannot always be viewed exclusively as a privacy matter. In appropriate circumstances, the manner in which data is accumulated or used can also form part of the assessment of market power and abusive conduct.

22. Case 6: Intel v European Commission

Case C-413/14 P, Court of Justice of the European Union, 2017

Intel concerned conditional rebates and alleged exclusion of competitors in the microprocessor market.

The Court emphasised the importance of analysing whether challenged practices were capable of restricting competition, particularly where the dominant undertaking submits evidence disputing their foreclosure capability.

Importance For Self-Learning Economies

AI markets may involve automated:

• rebates;
• loyalty programmes;
• personalised discounts;
• exclusivity incentives; and
• pricing schemes.

Intel shows why authorities should not automatically assume that sophisticated commercial incentives are exclusionary.

Their competitive capability and economic context must be properly assessed.

The same principle is important where such incentives are generated automatically by AI.

23. Combined Lessons From The Case Law

These cases collectively provide several principles relevant to self-learning economies.

First, technology does not create immunity from competition law. A restriction implemented through software can still amount to anticompetitive conduct.

Second, responsibility remains important. Authorities must identify the connection between an undertaking and the algorithmic conduct rather than automatically treating every algorithmic outcome as an agreement.

Third, dominant digital platforms may face competition scrutiny where algorithms, contractual restrictions or ecosystem rules disadvantage competitors.

Fourth, data can be strategically important to competition.

Fifth, competition authorities increasingly need effects-based economic and technological analysis.

24. Major Enforcement Difficulty: Who Is Responsible?

Suppose an algorithm develops an anticompetitive strategy that its programmer never expressly instructed it to adopt.

Possible actors include:

• the company deploying the system;
• software developers;
• the algorithm provider;
• the platform operating the infrastructure; and
• businesses participating in the system.

Competition law generally regulates undertakings rather than machines themselves.

The central legal question is therefore not whether the algorithm itself can be "guilty", but whether its conduct can legally and economically be attributed to an undertaking.

This issue will become increasingly significant as autonomous commercial systems become more sophisticated.

25. Human Instruction Versus Machine Learning

Three situations should be distinguished.

Situation A — Human-created cartel implemented by algorithm

Competitors agree to maintain particular prices and use algorithms to execute the agreement.

This is fundamentally an ordinary cartel implemented technologically.

Situation B — Common algorithm facilitates coordination

Competitors use a shared system that coordinates pricing or exchanges strategic information.

This can raise Article 101 concerns depending on knowledge, participation and the surrounding evidence.

Situation C — Independent algorithms autonomously learn parallel strategies

Competitors independently develop algorithms without communication, but the systems eventually learn that parallel higher prices maximise profits.

This is the most difficult situation because the economic result can resemble collusion while the traditional legal concept of agreement or concerted practice may be harder to establish.

26. Market Definition Problems

Traditional competition analysis defines relevant product and geographic markets.

Self-learning economies complicate this process.

A digital platform may simultaneously provide:

• free consumer services;
• paid advertising;
• data analytics;
• marketplace services; and
• AI infrastructure.

Consumers may pay no monetary price but provide attention or data.

Competition authorities therefore have to consider non-price dimensions such as:

• privacy;
• innovation;
• service quality;
• data access;
• interoperability;
• processing speed; and
• accuracy.

Zero monetary price does not mean that competition is economically unimportant.

27. Dynamic Competition

Self-learning markets can change rapidly.

Today's small AI company could potentially become an important competitor if its technology develops successfully.

Competition authorities therefore increasingly need to consider dynamic competition.

Instead of examining only existing market shares, they may also consider:

• innovation capabilities;
• intellectual property;
• access to computing resources;
• datasets;
• technical talent;
• potential entry; and
• ecosystem relationships.

This is particularly important in merger analysis.

28. Competition Versus Innovation

Competition regulation must avoid treating technological success itself as unlawful.

Large datasets, efficient algorithms and integrated ecosystems can create major consumer benefits.

Overly aggressive intervention could discourage:

• investment;
• research;
• experimentation;
• innovation; and
• development of new AI systems.

But insufficient enforcement can allow established companies to protect their positions by excluding emerging competitors.

The objective is therefore to protect the competitive process, rather than protecting individual competitors from legitimate competition.

29. Possible Regulatory Responses

Authorities dealing with self-learning economies may increasingly rely on several approaches.

These can include:

Algorithmic auditing: examining how important commercial algorithms operate.

Data portability: allowing users to transfer relevant data between competing services.

Interoperability: reducing technological barriers between ecosystems where legally justified.

Merger scrutiny: examining acquisitions of emerging AI competitors.

Information-exchange enforcement: preventing competitors from using algorithms as indirect channels for strategic coordination.

Technical expertise: employing economists, computer scientists and data specialists alongside lawyers.

Documentation requirements: maintaining records concerning important automated commercial decisions.

However, regulation must remain proportionate because extensive intervention can itself reduce incentives to innovate.

30. Special Problem Of Reinforcement Learning

Reinforcement-learning systems deserve particular attention.

Such systems learn by:

1. taking an action;
2. observing the outcome;
3. receiving a reward;
4. adjusting future behaviour; and
5. repeating the process.

In pricing environments, an algorithm could experiment with prices and observe competitors' reactions.

Over many interactions, it might learn that aggressive competition produces lower profits while stable parallel pricing produces higher rewards.

Experimental economic research has shown that certain reinforcement-learning systems can learn forms of tacitly cooperative pricing.

This does not mean every reinforcement-learning pricing system will collude.

It means that competition authorities cannot assume that independent automated pricing will necessarily reproduce traditional competitive behaviour.

31. Consumer Harm

Competition problems in self-learning economies may affect consumers through more than higher prices.

Potential harm includes:

• reduced product choice;
• lower innovation;
• weaker privacy protection;
• discriminatory commercial treatment;
• reduced service quality;
• restricted interoperability;
• higher switching costs;
• reduced visibility of competing products; and
• dependence on a small number of digital ecosystems.

Therefore, competition analysis must increasingly examine quality, innovation and choice alongside monetary prices.

32. Business Compliance

Businesses using AI should treat competition compliance as part of algorithmic governance.

Important questions include:

• What objective is the algorithm optimising?
• What competitor information can it access?
• Does it exchange information with external systems?
• Who controls its parameters?
• Can humans override its decisions?
• Are important decisions logged?
• Does a third-party provider serve competing businesses?
• Could the algorithm discriminate against competitors?
• Does it automatically impose exclusivity?
• Can potentially problematic behaviour be detected?

Compliance programmes should therefore involve lawyers, economists, engineers and data scientists.

33. Conclusion

Self-learning economies represent an important new stage in competition law.

Artificial intelligence and machine-learning systems can make markets faster, cheaper and more efficient. At the same time, they can strengthen existing market power, increase barriers to entry, facilitate coordination and make exclusionary behaviour more difficult to identify.

The most important competition concerns include:

1. algorithmic collusion;
2. hub-and-spoke coordination;
3. autonomous tacit coordination;
4. concentration of economically valuable data;
5. network effects;
6. algorithmic self-preferencing;
7. ecosystem tying and bundling;
8. algorithmic lock-in;
9. discriminatory ranking;
10. exclusionary automated incentives;
11. anticompetitive information exchange; and
12. acquisitions of emerging AI competitors.

Cases such as Eturas, VM Remonts, Google Shopping, Google Android, Meta Platforms v Bundeskartellamt, and Intel v Commission provide important legal principles even though most were not specifically about autonomous self-learning AI.

The fundamental challenge is that traditional competition law was developed around decisions made by human-controlled undertakings, whereas future markets may increasingly involve systems capable of adapting their strategies independently.

Competition law therefore does not necessarily need to abandon its traditional principles. Instead, concepts such as agreement, concerted practice, dominance, foreclosure, attribution and competitive effects must be applied carefully to markets in which algorithms increasingly determine how firms compete.

Ultimately, the central objective remains unchanged: to preserve effective competition while allowing legitimate technological innovation and efficiency to develop.

The six cases and current legal status above were checked against CJEU materials; notably, Google Shopping was upheld by the Court of Justice in 2024, and the Google Android appeal was dismissed on 2 July 2026, confirming the roughly €4.1 billion fine.