Competition Law and Antitrust Frameworks for Self-Learning Ecosystems

1. Introduction

A self-learning ecosystem is a digital or AI-based environment in which algorithms continuously learn from data, user behaviour, transactions, competitors, and previous outcomes, and then modify their future decisions. Examples include AI pricing systems, recommendation engines, digital marketplaces, advertising platforms, autonomous purchasing systems, and machine-learning-based allocation systems.

From a competition-law perspective, the important point is that automation does not remove legal responsibility. Existing antitrust principles concerning collusion, monopolization or abuse of dominance, exclusionary conduct, information exchange, tying, discrimination, and mergers can still apply when conduct is implemented or reinforced through machine-learning systems.

The issue is becoming particularly important because a self-learning system can create a feedback loop:

More users → more data → better algorithm → better service/predictions → more users → still more data.

Such a loop can be beneficial because it improves products. But where rivals cannot reasonably replicate the data, infrastructure, distribution, or user base involved, it can also contribute to durable market power.

A useful contemporary example is the U.S. litigation involving RealPage. The U.S. Department of Justice alleged that competing landlords supplied non-public information that was processed through RealPage's pricing software to generate pricing recommendations. The DOJ brought claims under Sections 1 and 2 of the Sherman Act. The case illustrates how conventional antitrust concepts can be applied to algorithm-supported markets.

2. Why Self-Learning Ecosystems Create Competition-Law Issues

Traditional competition law normally assumes that firms independently determine important commercial matters such as price, production, investment, product quality, and contractual terms.

Self-learning systems complicate this assumption because decisions may increasingly be generated through algorithms.

The central legal question therefore is not simply:

"Did an algorithm make the decision?"

The more useful questions are:

• Who designed or deployed the system?
• What information is supplied to it?
• Does it use competitors' confidential information?
• Do competing firms use the same intermediary or algorithm?
• Does learning make competitors' behaviour more predictable?
• Can firms genuinely reject algorithmic recommendations?
• Does control over training data create barriers to entry?
• Does a dominant platform favour its own services?
• Can rivals obtain access to important interfaces, infrastructure, or data?
• Does the system strengthen market power over time?

The answers determine which competition-law framework becomes relevant.

3. Algorithmic Collusion and Coordinated Behaviour

One of the most significant concerns is algorithm-assisted coordination.

Suppose several competing sellers independently use the same pricing platform. Each seller sends current confidential information to the platform. The platform learns from all participants and recommends prices back to them.

The arrangement can potentially reduce independent competitive decision-making.

Under U.S. law, Section 1 of the Sherman Act prohibits certain agreements that unreasonably restrain trade. In the EU, Article 101 TFEU prohibits agreements and concerted practices that restrict competition.

An algorithm therefore cannot automatically legalize conduct that would otherwise constitute unlawful coordination.

The RealPage litigation provides an important current illustration. The DOJ alleged that RealPage's software used competing landlords' non-public information in generating rental-price recommendations. It characterized the alleged conduct as both a coordination problem and, separately, a monopolization issue.

Importantly, these were allegations in litigation rather than simply a general judicial rule that every common pricing algorithm constitutes collusion.

4. Autonomous or Tacit Algorithmic Coordination

A more difficult situation occurs when competitors do not explicitly agree to coordinate.

Imagine two independently developed AI pricing systems. Each continuously observes publicly available prices and learns that immediately matching a competitor's price increases produces higher profits than aggressive price competition.

Eventually, the algorithms may reach parallel strategies without a conventional human agreement.

This creates a difficult distinction between:

unlawful coordination and lawful conscious parallelism or unilateral adaptation to market conditions.

Traditional competition law generally requires something more than merely showing that competitors behaved similarly when liability depends on an agreement or concerted practice.

Consequently, regulators may investigate evidence concerning communications, system architecture, common intermediaries, information exchange, instructions given to algorithms, and whether firms deliberately created mechanisms facilitating coordination.

5. Hub-and-Spoke Algorithmic Structures

Self-learning ecosystems can also resemble a hub-and-spoke arrangement.

The structure may look like:

Competitor A → AI platform ← Competitor B

Competitor C → AI platform ← Competitor D

The AI provider functions as the hub, while competing businesses constitute the spokes.

Competition concerns become stronger when competing firms provide current confidential information to the same platform and understand that the system processes information from other competitors.

This is particularly important for pricing, capacity, inventory, discounts, wages, bidding, and output.

The RealPage proceedings demonstrate why competition authorities are examining common algorithmic intermediaries closely. DOJ alleged that participating landlords supplied competitively sensitive information and received algorithmically generated recommendations.

6. Data-Driven Market Power

Self-learning ecosystems depend heavily on data.

Consider:

Platform A has 100 million users.

Their activity generates enormous training data. The platform improves its algorithm. Improved recommendations attract additional users, producing still more training information.

This can generate a self-reinforcing competitive advantage.

Competition authorities may therefore examine whether control over data creates:

• barriers to entry;
• switching costs;
• network effects;
• learning advantages;
• economies of scale;
• economies of scope;
• ecosystem dependence; or
• foreclosure opportunities.