Competition Law and Future Competition Risks in Autonomous Systems

Introduction

Autonomous systems are technologies capable of making or implementing decisions with limited or no contemporaneous human intervention. They include autonomous vehicles, AI agents, algorithmic pricing systems, autonomous procurement systems, trading systems, robotic marketplaces, smart-grid systems, AI-driven logistics, and autonomous digital platforms.

From a competition-law perspective, the central difficulty is that traditional competition law is designed around human or corporate decision-making, whereas autonomous systems may independently observe markets, predict competitors' conduct, alter prices, allocate resources, select suppliers, or negotiate transactions.

The legal problem is therefore not merely whether an algorithm produces an anticompetitive outcome. It is whether competition law can attribute that conduct to an undertaking, establish an agreement or concerted practice, identify exclusionary conduct, and impose an effective remedy when the system itself continuously adapts.

Existing cases such as Eturas, Topkins, RealPage, Uber-related litigation, and platform-dominance cases provide important foundations, although none completely resolves the problem of genuinely autonomous AI systems. The CJEU's Eturas judgment, for example, specifically considered automatic restrictions implemented through a common computerised booking system.

I. Meaning of Autonomous Systems

An autonomous system generally contains:

1. Data inputs – market prices, consumer behaviour, competitor information, supply conditions;
2. Algorithmic processing – prediction, optimisation or machine learning;
3. Decision rules or objectives – profit maximisation, cost reduction, market-share objectives;
4. Automated execution – changing prices, purchasing inputs, allocating customers or refusing access;
5. Feedback mechanisms – learning from subsequent market results; and
6. Continuous adaptation – modifying future conduct without a new human instruction.

Examples include:

• autonomous pricing agents;
• AI procurement systems;
• autonomous trading systems;
• self-driving transportation platforms;
• AI advertising auctions;
• autonomous supply-chain management;
• AI-powered marketplaces;
• smart-energy management systems;
• autonomous contracting agents;
• algorithmic recommendation systems; and
• multi-agent AI economies.

The greater the autonomy, the more difficult it becomes to identify a conventional "decision-maker."

II. Existing Competition-Law Framework

Autonomous systems can potentially engage several established areas of competition law.

1. Anti-competitive agreements

In India, Section 3 of the Competition Act, 2002 addresses agreements, arrangements, understandings and concerted actions having an appreciable adverse effect on competition.

Under EU law, Article 101 TFEU covers agreements, decisions and concerted practices.

The principal difficulty is determining whether two autonomous systems that independently learn similar behaviour have actually agreed.

2. Abuse of dominance

An autonomous system operated by a dominant undertaking may:

• discriminate between customers;
• favour affiliated businesses;
• deny interoperability;
• restrict access;
• manipulate rankings;
• foreclose rivals;
• bundle products;
• exploit commercially sensitive data; or
• impose exclusionary conditions.

Article 102 TFEU and comparable national provisions can therefore remain relevant even where no cartel exists.

3. Merger control

Autonomous systems may increase the strategic importance of:

• AI foundation models;
• cloud infrastructure;
• datasets;
• computing capacity;
• autonomous-agent platforms;
• robotics;
• autonomous mobility networks; and
• AI operating systems.

Competition authorities may therefore have to consider whether acquisitions of apparently small AI businesses eliminate future competitors or important innovation paths.

III. Major Future Competition Risks

1. Algorithmic Collusion

The most obvious risk is that autonomous systems may coordinate prices without traditional human communication.

For example:

AI System A observes System B's price → increases its own price → B responds → A learns that retaliation is profitable → both systems converge on supra-competitive prices.

There may be no telephone call, email or meeting between executives.

This creates a distinction between:

human conspiracy → algorithmic implementation

and

independent algorithms → autonomous coordination.

The first is comparatively easier for existing competition law to address. The second creates the harder question of whether autonomous behaviour itself constitutes an infringement.

IV. Six Important Case Laws

1. Eturas UAB and Others v Lithuanian Competition Council

C-74/14, CJEU, 2016

This is one of the most important cases for autonomous and algorithmic competition law.

Several travel agencies used a common electronic booking system. The system administrator sent a message indicating that discounts available through the system would be restricted, and the system automatically implemented the restriction.

The CJEU considered whether participation in the common system and awareness of the restriction could constitute a concerted practice.

Significance

The case demonstrates that:

• electronic communication can constitute the mechanism for competition-law coordination;
• automated implementation does not necessarily remove legal responsibility;
• evidence concerning awareness and subsequent conduct is crucial;
• technology may act as the infrastructure through which competitors coordinate.

Eturas is particularly important for autonomous systems because it shows that competition law can examine the interaction between software architecture and undertaking behaviour, rather than merely traditional written agreements.

2. United States v Topkins

In United States v Topkins, an online seller and competitors used pricing software to implement an agreement concerning prices for products sold through an online marketplace.

The significance lies in the distinction between:

illegal agreement + algorithmic implementation

and merely:

independent algorithmic pricing.

The algorithm did not automatically absolve the participants from responsibility. Rather, software was used as an instrument for implementing the alleged price-fixing arrangement.

Competition-law principle

Technology cannot be used as a legal shield where human actors have already formed an anticompetitive agreement.

This principle is likely to remain important as autonomous systems become more sophisticated.

3. United States v RealPage Inc.

The RealPage litigation represents a major development concerning algorithmic pricing.

The allegations concerned rental-pricing software that used information supplied by competing landlords and generated pricing recommendations.

The competition concern was that competitors could effectively delegate important pricing decisions to a common algorithmic intermediary.

The case illustrates the potential hub-and-spoke problem:

Landlord A → algorithm/platform → pricing recommendation

Landlord B → algorithm/platform → pricing recommendation

If the system facilitates the exchange or use of competitively sensitive information, conventional competition-law concepts may become relevant even though competitors do not directly communicate with each other. Contemporary analysis of algorithmic pricing identifies RealPage alongside Topkins and Eturas as central examples of the problem.

Future significance

Autonomous procurement and pricing agents could create similar structures on a much larger scale.

4. Uber Technologies – Algorithmic Pricing Litigation

Uber-related litigation has raised questions about whether a common algorithm determining prices for drivers can facilitate coordination among otherwise competing drivers.

The central theoretical issue is:

If independent drivers delegate pricing decisions to the same algorithm, does the algorithm become a mechanism for horizontal coordination?

The question becomes particularly significant where drivers are legally independent undertakings rather than employees.

Academic analysis of the issue has specifically examined whether Uber's common pricing mechanism could potentially be characterised through a hub-and-spoke theory, while noting that no general EU infringement finding has established such a proposition.

Importance for autonomous systems

Future autonomous transport systems may remove even more human involvement:

• autonomous vehicle A chooses fare;
• autonomous vehicle B observes market response;
• both systems modify prices;
• AI agents learn from one another.

Competition authorities would then have to determine whether this is merely parallel adaptation or legally cognisable coordination.

5. Amazon Marketplace / Amazon Data-Use Investigation

The European Commission investigated Amazon's use of non-public seller data obtained through its marketplace.

The Commission's preliminary view was that Amazon's automated systems and employees systematically relied on non-public information concerning competitors and that this could distort competition by giving Amazon advantages unavailable to ordinary competitors.

Importance

This demonstrates another autonomous-system risk:

data-powered competitive self-preferencing.

An autonomous marketplace could continuously collect:

• rival prices;
• inventory;
• sales volumes;
• conversion rates;
• customer behaviour;
• product performance; and
• seller strategies.

Its AI could then automatically modify its own commercial decisions.

The competition problem therefore may not be an algorithm "colluding" with rivals. It may instead be a dominant algorithm using privileged data to compete against the businesses dependent upon its infrastructure.

6. Google Android / Google Search Competition Litigation

The Google Android litigation illustrates how digital ecosystems can combine several layers of market power.

The CJEU's 2026 judgment concerning Google's Android-related conduct addressed contractual restrictions, tying, payments connected with exclusive pre-installation, and restrictions affecting Android forks.

Relevance to autonomous systems

Future autonomous ecosystems could similarly control several layers:

AI operating system → agent marketplace → cloud → data → payments → applications → autonomous services.

If one undertaking controls several essential layers, autonomous agents may be technically capable of switching providers but commercially constrained from doing so.

This makes interoperability and switching costs increasingly important competition-law questions.

V. Autonomous Collusion Without Human Agreement

This is perhaps the most difficult future problem.

Consider three competing AI pricing agents:

• Agent A seeks maximum long-term profit;
• Agent B seeks maximum long-term profit;
• Agent C seeks maximum long-term profit.

They independently learn that aggressive price competition reduces their rewards.

Through repeated interactions, they discover:

"Maintaining a higher price produces higher long-term returns."

They begin maintaining prices above competitive levels.

No human communicates with another human.

Legal problem

Traditional cartel law asks:

Did the undertakings agree or coordinate?

Autonomous systems create another question:

Can an undertaking be responsible when its AI independently discovers an anti-competitive strategy that its designers neither expressly instructed nor anticipated?

This is one of the principal unresolved issues in future competition law. Contemporary research has specifically highlighted the difficulty of applying agreement-centred competition rules to autonomous reinforcement-learning systems.

VI. The "Black Box" Problem

Autonomous systems can make competition-law investigations difficult because regulators may not know:

• why a price increased;
• why a rival was excluded;
• why a customer received different terms;
• why an acquisition target was selected;
• why a supplier was rejected;
• why access was denied;
• why an algorithm changed strategy.

An AI system may produce:

Input → Model → Decision

without producing a legally intelligible explanation.

This creates an evidentiary problem.

Competition authorities traditionally examine:

• emails;
• contracts;
• board minutes;
• communications;
• pricing instructions;
• internal documents.

Autonomous systems may instead require examination of:

• model weights;
• training data;
• prompts;
• reward functions;
• system logs;
• API calls;
• model updates;
• agent-to-agent communications;
• decision histories; and
• simulation records.

VII. Autonomous Self-Preferencing

A dominant autonomous platform could automatically favour its own products.

For example:

An AI marketplace controls search ranking and simultaneously sells its own products.

The system may continuously optimise rankings according to commercial objectives.

Potential problems include:

1. ranking affiliated products more prominently;
2. suppressing competitors;
3. manipulating recommendations;
4. allocating scarce visibility;
5. controlling advertising access;
6. prioritising proprietary AI agents; and
7. using competitor data to improve its own products.

The challenge is determining whether the conduct constitutes legitimate optimisation or exclusionary abuse.

VIII. Autonomous Discrimination

AI systems can automatically discriminate between:

• customers;
• suppliers;
• distributors;
• geographic markets;
• competing businesses;
• advertisers; and
• platform participants.

Competition risks include:

A. Price discrimination

Different customers may receive different prices based on predicted willingness to pay.

B. Supplier discrimination

A platform may automatically provide preferred suppliers with better access.

C. Competitor discrimination

The platform may identify emerging competitors and automatically reduce their visibility.

D. Data discrimination

Some businesses may receive extensive data access while competitors receive restricted access.

IX. Autonomous Refusal to Deal

A future AI infrastructure provider could automatically determine:

"This competitor is sufficiently threatening; access should be restricted."

The decision could be generated without direct human instruction.

This creates important questions concerning:

• essential facilities;
• interoperability;
• access obligations;
• dominant platforms;
• cloud infrastructure;
• AI compute;
• data access;
• API access; and
• operating-system interfaces.

The essential-facility doctrine may therefore become increasingly relevant where autonomous ecosystems control infrastructure necessary for competing AI agents.

X. Autonomous Mergers and Acquisition Decisions

Autonomous systems could themselves identify potential acquisition targets.

A dominant company could deploy an AI system instructed to:

"Acquire technologies that could threaten our market position."

The AI could identify dozens of startups and recommend acquisitions.

Competition-law risks include:

• acquisition of nascent competitors;
• elimination of potential competition;
• acquisition of unique datasets;
• acquisition of specialised AI talent;
• killer acquisitions;
• consolidation of AI infrastructure; and
• accumulation of strategic technologies.

The traditional turnover-based merger threshold may therefore fail to capture certain strategically important acquisitions.

XI. Autonomous Procurement Cartels

AI procurement agents could independently communicate with supplier systems.

For example:

Buyer A's AI → Supplier platform

Buyer B's AI → Supplier platform

If the same intermediary obtains sensitive information about both buyers, it may inadvertently facilitate coordination.

Similar concerns could arise in:

• construction;
• pharmaceuticals;
• defence procurement;
• energy;
• logistics;
• agriculture;
• shipping; and
• government procurement.

The future question will increasingly be whether machine-to-machine communication constitutes competitively significant information exchange.

XII. Autonomous Vertical Foreclosure

An autonomous platform may simultaneously operate as:

• infrastructure provider;
• marketplace;
• competitor;
• advertiser;
• data intermediary; and
• payment provider.

Its AI could automatically optimise the ecosystem in favour of affiliated businesses.

This can produce:

input foreclosure + data advantage + ranking advantage + customer lock-in.

Competition authorities may therefore need to assess the entire ecosystem rather than a single algorithmic decision.

XIII. Network Effects Become Autonomous

Autonomous systems can strengthen network effects faster than traditional platforms.

More users produce:

→ more data
→ better AI
→ better predictions
→ more users
→ more data
→ greater accuracy.

This creates a self-reinforcing competitive advantage.

The resulting data-feedback loop can make market entry difficult even where the underlying software itself is replicable.

XIV. Autonomous Switching and Lock-In

AI agents may become deeply integrated with:

• personal data;
• calendars;
• payment systems;
• cloud accounts;
• enterprise databases;
• smart homes;
• vehicles;
• healthcare systems;
• financial accounts; and
• communication systems.

Switching to a rival AI system could therefore become expensive.

Future competition law may need to examine:

• data portability;
• interoperability;
• API portability;
• agent portability;
• model portability;
• identity portability; and
• transaction-history portability.

XV. Autonomous Systems and Market Definition

Traditional market definition becomes complicated where one AI system performs multiple functions.

For example, an autonomous agent could simultaneously provide:

• search;
• shopping;
• financial services;
• travel;
• advertising;
• personal assistance;
• procurement; and
• payment services.

The relevant question may become whether markets should be defined by function, ecosystem, data, user relationship, or agent capability.

XVI. Autonomous Systems and Predatory Conduct

An AI system may calculate that temporarily losing money will eliminate a competitor.

It could then automatically:

1. lower prices;
2. increase subsidies;
3. increase advertising;
4. target the competitor's customers;
5. maintain losses;
6. wait for competitor exit; and
7. raise prices afterward.

The difficulty is establishing:

• intention;
• strategy;
• recoupment;
• attribution; and
• duration.

AI therefore potentially makes traditional predatory-pricing analysis more complex.

XVII. Autonomous Systems and Exploitative Conduct

Autonomous systems can also exploit consumers through:

• personalised pricing;
• behavioural targeting;
• automated contract terms;
• dynamic fees;
• subscription renewal;
• personalised offers;
• switching barriers; and
• algorithmically optimised consumer segmentation.

Competition law may increasingly intersect with consumer protection and data regulation.

XVIII. Liability and Attribution

A future autonomous-system framework may require a distinction among:

1. Developer liability

The developer designed the system.

2. Operator liability

The undertaking deployed the system.

3. User liability

The undertaking used the system commercially.

4. Platform liability

The infrastructure provider enabled interactions between competing systems.

5. Agent autonomy

The system itself generated behaviour not specifically programmed by its operator.

The most practical competition-law approach is likely to continue focusing on the undertaking that controls, deploys, benefits from, or materially enables the relevant conduct, while developing more sophisticated evidentiary rules for autonomous decision-making.

XIX. Compliance Requirements for Autonomous Systems

Businesses deploying autonomous systems should consider:

A. Competition-by-design

Competition compliance should be incorporated into system architecture.

B. Restricted-data controls

Competitively sensitive information should not automatically flow between competitors.

C. Audit logs

Systems should retain records of important decisions.

D. Human escalation

High-risk decisions should be capable of human review.

E. Competition constraints

Reward functions should not incentivise cartel-like behaviour.

F. Independent testing

Systems should be tested for:

• collusion;
• exclusion;
• discriminatory access;
• self-preferencing;
• predatory strategies; and
• information leakage.

G. Model governance

Businesses should document:

• objectives;
• training data;
• reward functions;
• constraints;
• updates;
• monitoring;
• intervention procedures.

XX. Future Regulatory Models

Future competition regulation could develop along several complementary lines.

1. Algorithmic audit obligations

High-risk systems could require periodic competition audits.

2. Explainability requirements

Dominant platforms may need to explain important competitive decisions.

3. Data-access safeguards

Competitively sensitive data could receive additional protections.

4. Interoperability obligations

Dominant autonomous ecosystems could be required to permit meaningful interoperability.

5. Agent-to-agent communication rules

Competition authorities could establish rules concerning automated communication between competing systems.

6. Competition impact assessments

Large autonomous systems could undergo competition-impact assessments before deployment.

7. Real-time monitoring

Competition authorities could develop technical capabilities for monitoring algorithmic markets continuously.

8. Structural remedies

In extreme cases involving ecosystem control, regulators could consider:

• separation;
• access obligations;
• data silos;
• interoperability;
• non-discrimination obligations; or
• divestiture.

XXI. Important Doctrinal Questions for the Future

XXII. Key Legal Principle Emerging

The most important future principle is likely to be:

Automation should not eliminate competition-law responsibility merely because the decision was made by software rather than a human.

At the same time, competition law must distinguish between:

legitimate autonomous optimisation

and

autonomous conduct that is legally attributable to an undertaking and produces an anticompetitive effect.

That distinction will be central to future cases.

XXIII. Conclusion

Autonomous systems create a fundamental challenge for competition law because they can transform markets from human-mediated decision-making environments into continuously adapting computational environments.

The existing jurisprudence already provides important building blocks. Eturas demonstrates that automated systems can form part of a concerted-practice analysis; Topkins illustrates that algorithms cannot immunise traditional price-fixing; RealPage demonstrates the competition concerns surrounding shared algorithmic pricing and competitively sensitive data; Uber-related litigation highlights the hub-and-spoke problem; Amazon demonstrates the competitive implications of automated exploitation of marketplace data; and the Google Android litigation illustrates the importance of ecosystem control and exclusionary restrictions.

The next generation of competition cases will likely move beyond the simple question "Did a human order the anticompetitive conduct?" toward more complex questions concerning control, attribution, data, algorithmic incentives, machine-to-machine coordination, explainability, interoperability and autonomous adaptation.

Ultimately, competition law will need to remain technologically neutral while developing sufficiently sophisticated tools to deal with markets in which the economically significant decisions are increasingly made by autonomous computational systems rather than directly by human decision-makers.