Market Behavior Attribution in Distributed Intelligence Systems

Introduction

Market behavior attribution in distributed intelligence systems concerns the legal and economic problem of determining which actor, system, algorithm, platform, or decision-making layer is responsible for market conduct when commercial decisions are produced collectively by interconnected AI systems.

Traditional competition law generally assumes that a firm can be associated with a particular decision: a manager approves a price, a sales department allocates customers, or executives agree with competitors. Distributed intelligence systems complicate this model. Pricing, ranking, allocation, procurement, advertising, inventory, logistics, and investment decisions may emerge from several interacting components:

• a platform's central algorithm;
• autonomous AI agents;
• third-party software;
• cloud infrastructure;
• data providers;
• foundation models;
• merchants or suppliers;
• competing algorithms;
• reinforcement-learning systems;
• human supervisors; and
• feedback generated by the market itself.

The central legal question therefore becomes:

When market conduct emerges from a distributed intelligent system, whose conduct is it for competition-law purposes?

The answer cannot depend exclusively on identifying a human who manually selected the final action. Competition law increasingly focuses on economic responsibility, control, implementation, effects, and the foreseeable functioning of the system.

I. Meaning of Distributed Intelligence Systems

A distributed intelligence system is a technological environment in which decision-making intelligence is spread across several interconnected actors or computational components rather than concentrated in one algorithm.

Simplified structure

Data sources → AI/model → platform algorithm → autonomous agent → market interface → competitors' systems → market feedback → AI/model

For example, an online marketplace may use:

1. a demand-prediction model;
2. an automated pricing engine;
3. a seller-management algorithm;
4. a recommendation system;
5. a logistics algorithm; and
6. an external AI agent used by individual sellers.

No single component necessarily determines the final price.

Nevertheless, the combined system may produce:

• parallel pricing;
• market allocation;
• discriminatory access;
• exclusionary ranking;
• coordinated inventory reductions;
• common price responses; or
• foreclosure of rivals.

This creates an attribution gap between technological causation and legal responsibility.

II. Why Attribution Matters in Competition Law

Attribution determines:

• whether there is an agreement or concerted practice;
• whether a dominant undertaking abused its position;
• whether conduct is unilateral or coordinated;
• which undertaking is legally responsible;
• whether an algorithm's conduct can be attributed to its operator;
• whether a technology provider can itself be liable;
• whether a platform is responsible for third-party AI behaviour;
• whether parent and subsidiary conduct can be aggregated;
• whether evidence establishes intent or knowledge; and
• whether remedies should target the algorithm, the firm, or the market structure.

The issue is particularly important under Article 101 TFEU, Article 102 TFEU, UK Competition Act 1998 Chapters I and II, German GWB §§1, 19 and 19a, and comparable national regimes.

III. Traditional Attribution Model

Traditional competition law largely operates through the concept of the undertaking.

The undertaking is treated as the relevant economic unit rather than merely the natural person who physically performs the conduct.

Consequently, a company can be responsible for:

• employee decisions;
• sales-agent conduct;
• contractual implementation;
• automated systems;
• subsidiary conduct; and
• commercial policies implemented through technological systems.

The emergence of AI therefore does not automatically eliminate attribution.

The more difficult question is whether the distributed system's output can reasonably be treated as conduct of one or several undertakings.

IV. The Five Layers of Attribution

A useful framework is to divide attribution into five layers.

1. Design attribution

Who designed the system?

Questions include:

• Who selected the objective function?
• Who selected the optimization target?
• Who configured pricing constraints?
• Who determined permissible competitors?
• Who supplied training data?
• Who designed the reward function?

Design evidence can establish that particular market outcomes were programmed or structurally encouraged.

2. Deployment attribution

Who deployed the system commercially?

A firm may rely on an externally developed AI model but integrate it into its own pricing or ranking system.

The relevant question becomes:

Who placed the intelligence into the market environment?

Deployment can be more legally significant than authorship of the underlying model.

3. Control attribution

Who possesses meaningful control over the system?

Control may be:

• direct;
• contractual;
• technical;
• supervisory;
• economic; or
• architectural.

A company does not necessarily escape responsibility merely because an external AI provider technically operates part of the system.

4. Implementation attribution

Who actually implemented the resulting conduct?

Suppose an AI recommends a 20% price increase.

The merchant's system automatically adopts that recommendation.

The important question is whether the merchant:

• accepted the system's authority;
• established the automated rule;
• knew how it operated;
• could intervene; and
• continued using it after observing its effects.

Implementation creates a powerful connection between AI output and undertaking conduct.

5. Effect attribution

Who benefited from or caused the competitive effect?

This includes examining:

• price effects;
• foreclosure;
• exclusion;
• reduced innovation;
• discriminatory access;
• market sharing;
• reduced output; and
• increased concentration.

Effect attribution is especially important where multiple AI systems jointly generate the result.

V. Distributed Intelligence Does Not Necessarily Mean Distributed Liability

A crucial principle is:

Distributed decision-making does not necessarily produce distributed legal responsibility.

If a company deliberately deploys an autonomous system to make commercial decisions, it may remain responsible even if the individual decision was not manually reviewed.

For example, a platform cannot necessarily argue:

"The AI increased prices, not the company."

Competition law normally evaluates the undertaking's economic conduct, not simply the physical source of the keystroke.

VI. Intent Versus Responsibility

One of the most difficult questions is whether human intent is necessary.

For certain forms of coordinated conduct, proof of communication, knowledge, or conscious participation may be highly important.

But the absence of a human intention behind every individual AI action does not necessarily prevent liability.

There are two different questions:

Question 1: Did humans intend the specific outcome?

Possibly not.

Question 2: Did the undertaking deliberately operate a system capable of producing the conduct?

This may be much easier to establish.

This distinction is fundamental for algorithmic competition law.

VII. Case Law

1. Eturas v Lietuvos Respublikos konkurencijos taryba (CJEU)

Case C-74/14

This is one of the most important cases for digital attribution.

Several travel agencies used an electronic booking system. The system administrator sent a message imposing a restriction on discounts that could be offered through the platform.

The CJEU examined whether the agencies could be regarded as participating in a concerted practice.

Principle

The Court emphasized that participation cannot simply be presumed from the existence of an electronic communication. However, where participants became aware of information capable of influencing their conduct and continued operating in the relevant system without distancing themselves, the circumstances could support an inference of participation.

Importance for distributed intelligence

The case demonstrates that:

• technological architecture can facilitate coordination;
• an electronic system can transmit competitively significant information;
• actual human communication need not resemble a traditional meeting;
• continued participation may be evidentially significant.

For AI systems, system participation plus knowledge of coordinated parameters may become important evidence of attribution.

VIII. AC-Treuhand v Commission

Cases C-194/14 P and earlier AC-Treuhand jurisprudence

AC-Treuhand is important because it addresses the responsibility of an actor that facilitates anticompetitive conduct without itself being a conventional competitor in the affected market.

The Court accepted that an undertaking may fall within Article 101 where its conduct forms part of an anticompetitive arrangement and it contributes to its implementation.

Distributed-intelligence significance

This provides a conceptual basis for analysing:

• AI infrastructure providers;
• algorithm intermediaries;
• data intermediaries;
• coordination platforms; and
• third-party system administrators.

The key question is not necessarily:

"Did the technology provider sell the same product?"

It may instead be:

"Did its conduct intentionally or knowingly contribute to the functioning of the anticompetitive arrangement?"

IX. Commission v Anic Partecipazioni

Case C-49/92 P

Anic is a foundational case concerning participation in concerted practices.

The Court treated participation in a concerted practice broadly and emphasized that an undertaking's responsibility can arise from participation in coordinated market conduct rather than from a conventional written agreement.

Relevance

Distributed intelligence systems may create conduct without:

• a signed agreement;
• explicit price instructions;
• direct communications; or
• centralized management.

The Anic approach supports examining the economic substance of participation rather than demanding traditional contractual evidence.

X. T-Mobile Netherlands

Case C-8/08

In T-Mobile Netherlands, the CJEU dealt with a meeting involving competitors and the concept of concerted practice.

The Court stressed the importance of the exchange of competitively sensitive information and the reduction of uncertainty concerning competitors' future market behaviour.

AI significance

Distributed AI systems can reduce strategic uncertainty without conventional conversations.

For example:

• AI systems observe competitors' prices;
• models predict competitor responses;
• algorithms automatically adjust prices;
• repeated machine interaction creates convergence.

The critical competition-law question becomes whether the reduction in strategic uncertainty is the result of independent market observation or attributable coordination.

XI. Hoffmann-La Roche v Commission

Case 85/76

Hoffmann-La Roche is a leading authority on the concept of abuse of dominance and loyalty-inducing exclusionary arrangements.

The Court examined conduct capable of restricting competition by tying customers to a dominant undertaking.

Distributed-system significance

The case illustrates that responsibility can arise from the commercial strategy implemented by an undertaking, rather than from identifying the individual employee who made every commercial decision.

For modern platforms, similar reasoning may apply where AI systems automatically:

• downgrade competitors;
• prioritize the platform's own products;
• restrict access;
• personalize exclusionary offers; or
• impose loyalty-inducing conditions.

XII. Google Shopping

Google Search (Shopping), Case T-612/17

The General Court examined Google's conduct involving the positioning and display of its comparison-shopping service.

The Court upheld the Commission's conclusion that Google's conduct constituted an abuse of dominance, although certain aspects of the Commission's legal analysis were scrutinized closely.

Distributed-intelligence significance

Modern ranking systems are algorithmic.

The relevant legal question cannot simply be:

"Which line of code caused the demotion?"

Instead, the analysis concerns the undertaking's overall ranking architecture and commercial implementation.

This is highly relevant where AI determines:

• ranking;
• recommendation;
• visibility;
• search results;
• access;
• advertising exposure.

XIII. Intel v Commission

Case C-413/14 P

Intel concerned conditional rebates and the assessment of exclusionary effects by a dominant undertaking.

The CJEU required appropriate consideration of the circumstances surrounding the rebate system and its potential ability to foreclose an equally efficient competitor.

Distributed-intelligence relevance

Intel demonstrates that competition analysis may require sophisticated economic examination rather than relying exclusively on formal labels.

For AI systems, authorities may therefore need to examine:

• algorithmic incentives;
• optimization objectives;
• foreclosure probabilities;
• switching costs;
• scale effects;
• dynamic feedback; and
• competitive counterfactuals.

Attribution and effects can consequently become deeply intertwined.

XIV. Google Android

Case T-604/18

The General Court considered Google's contractual and technological arrangements concerning Android and associated services.

The case illustrates the importance of analysing interlocking contractual and technological mechanisms within a digital ecosystem.

Distributed-intelligence significance

AI ecosystems similarly contain multiple layers:

Operating system → model → app store → API → data → recommendation → user interface

Attribution should therefore examine the ecosystem as a whole rather than isolating one algorithmic event.

XV. United States v Apple / Apple Antitrust Litigation

US antitrust litigation involving Apple provides another useful modern context for attribution in digital ecosystems.

The central issue in such cases is often not merely what a particular software component technically did, but how Apple's integrated technological and contractual architecture affected competition.

Relevance

This supports a broader analytical proposition:

Technological decentralisation within a product does not necessarily eliminate economic responsibility at the firm level.

XVI. The Problem of Autonomous AI Agents

The most difficult scenario arises when AI agents interact without direct human intervention.

Consider:

Agent A → observes Agent B → changes price → Agent B observes → changes price → Agent A responds

No human communicates.

Yet the market may converge on a supracompetitive price.

Three possible legal classifications

A. Independent parallel conduct

Each undertaking independently responds to market conditions.

No agreement or concerted practice necessarily exists.

B. Algorithmically facilitated concerted practice

The undertakings intentionally deploy systems designed to coordinate or knowingly participate in a common coordination mechanism.

This may create substantial Article 101/Chapter I risk.

C. Unilateral algorithmic conduct

A dominant undertaking's algorithm independently produces exclusionary effects.

The issue may instead fall under Article 102/Chapter II.

XVII. Causal Attribution

A sophisticated attribution test should ask:

Step 1 — Who supplied the objective?

Was the AI instructed to:

• maximize profit;
• maximize market share;
• eliminate rivals;
• stabilize prices;
• maximize engagement?

Step 2 — Who supplied the constraints?

For example:

• minimum margins;
• competitor exclusions;
• preferred sellers;
• geographic restrictions.

Step 3 — Who supplied the data?

Data may determine the AI's commercial perception of the market.

Step 4 — Who deployed the model?

The deploying undertaking usually has the strongest connection to market conduct.

Step 5 — Who retained intervention rights?

Human override capabilities can be evidentially important.

Step 6 — Who knew the outcome?

Actual knowledge is important, but constructive or inferred knowledge may also become relevant depending on the legal test.

Step 7 — Who continued the system?

Continued operation after discovering anticompetitive consequences may strengthen attribution.

XVIII. The "Black Box" Defence

An undertaking may argue:

"We cannot explain exactly why the AI reached this decision."

That defence should not automatically eliminate responsibility.

There is an important distinction between:

technical explainability and legal attribution.

An algorithm may be technically opaque while the undertaking's organizational responsibility remains clear.

For example:

• Company A chooses the model.
• Company A supplies the data.
• Company A establishes the commercial objective.
• Company A deploys the model.
• Company A earns the resulting revenue.

The inability to reconstruct every internal computational step does not necessarily sever attribution.

XIX. Foundation Models and Attribution

Foundation models create a further complication.

Suppose:

Foundation-model provider → cloud provider → platform → merchant → autonomous agent

Each layer may influence the final market behaviour.

Attribution should distinguish:

Model creator

Created general-purpose intelligence.

Infrastructure provider

Provided computational infrastructure.

Integrator

Connected the model to a commercial decision system.

Operator

Used the system to make market decisions.

Beneficiary

Received economic benefits from the conduct.

Usually, mere technological causation should not equal competition-law liability.

A cloud provider should not automatically become responsible for every anticompetitive act conducted by a customer using its servers.

The stronger case arises where the provider itself:

• designs the coordination mechanism;
• actively facilitates the conduct;
• knows the commercial purpose;
• imposes competitively restrictive rules; or
• participates in the arrangement.

XX. Distributed Attribution Matrix

XXI. Attribution Under Article 101 TFEU

Article 101 analysis should ask:

1. Who are the undertakings?
2. What communication or coordination occurred?
3. Was there an agreement, decision, or concerted practice?
4. Which undertaking participated?
5. Did the AI merely execute an independent strategy?
6. Did the system facilitate common understanding?
7. Was competitively sensitive information exchanged?
8. Could participation be inferred from continued conduct?
9. Was the AI architecture intentionally designed to coordinate?

The critical distinction is:

algorithmic parallelism ≠ automatically unlawful coordination.

Parallel outcomes alone are not sufficient in every case.

XXII. Attribution Under Article 102 TFEU

Article 102 becomes especially important where one undertaking controls the distributed intelligence architecture.

Potential abuses include:

• self-preferencing;
• discriminatory ranking;
• exclusionary recommendation;
• interoperability restrictions;
• tying;
• exploitative personalization;
• discriminatory access;
• data foreclosure;
• algorithmic degradation of rivals.

The question becomes:

Can the conduct of the AI system be attributed to the dominant undertaking?

Where the dominant undertaking designed, deployed and commercially controlled the system, attribution will generally be much easier to establish than where the system is genuinely independent.

XXIII. UK Competition Law

Under the Competition Act 1998, distributed intelligence can create both Chapter I and Chapter II issues.

Chapter I

Relevant where multiple firms' systems contribute to:

• price coordination;
• market allocation;
• bid coordination;
• output restrictions;
• information exchange.

Chapter II

Relevant where a dominant digital undertaking uses AI to:

• exclude competitors;
• manipulate rankings;
• discriminate;
• restrict interoperability;
• exploit data advantages.

The UK's Digital Markets, Competition and Consumers Act 2024 further strengthens the importance of identifying responsibility within complex digital ecosystems because designated firms may be subject to conduct requirements and pro-competition interventions.

XXIV. German Competition Law

Germany provides a particularly important framework through GWB §§1, 19 and 19a.

Section 19a is significant for undertakings of paramount significance across markets.

A distributed AI ecosystem may involve:

• search;
• cloud;
• advertising;
• operating systems;
• app stores;
• data;
• AI models;
• marketplaces.

Where one undertaking controls several layers, attribution becomes an institutional question:

Which conduct should be attributed to the undertaking's economic power, even when technically generated by separate intelligent components?

The Bundeskartellamt's digital-market approach makes technological architecture particularly relevant to dominance and ecosystem analysis.

XXV. Evidence in Distributed Intelligence Cases

Competition authorities may need evidence from:

• source code;
• model cards;
• system prompts;
• API logs;
• training records;
• deployment records;
• model-version histories;
• reward functions;
• pricing logs;
• decision trees;
• audit trails;
• internal emails;
• developer instructions;
• human override records;
• A/B tests;
• incident reports.

A major evidentiary principle is:

The absence of a human-readable decision does not mean the absence of evidence.

The system itself creates an extensive digital record.

XXVI. Explainability as an Attribution Mechanism

Explainability should not be treated solely as an AI-governance requirement.

It can become a competition-law evidence mechanism.

An authority may ask:

• Why did the algorithm select this price?
• Which variables affected the decision?
• Was a competitor's price used?
• Did the system optimize against rivals?
• Was the model trained using competitors' confidential information?
• Did the model receive instructions to stabilize prices?
• Was the system repeatedly producing the same exclusionary outcome?

Therefore:

Explainability → causation → attribution → liability

XXVII. Human-in-the-Loop Does Not Automatically Solve Attribution

A company may claim:

"A human approved every AI recommendation."

But the relevant question is whether the human exercised meaningful independent judgment.

If the system automatically generates:

• the price;
• the recommended supplier;
• the ranking;
• the discount;

and the human merely clicks "approve," genuine human control may be questionable.

Thus, the law may increasingly distinguish:

Formal human involvement

A human technically approves the decision.

Substantive human involvement

The human genuinely evaluates and can reject the AI's recommendation.

The latter provides stronger evidence of independent decision-making.

XXVIII. Autonomous Agents and the "Electronic Employee" Analogy

A useful conceptual model is to treat an AI agent as functionally similar to an automated commercial instrument.

An employee may act without seeking permission for every transaction.

Likewise, an AI system may:

• negotiate;
• price;
• purchase;
• advertise;
• allocate inventory.

The absence of continuous human supervision does not necessarily remove the undertaking's responsibility for its commercial machinery.

However, the analogy has limits because autonomous AI may interact with other autonomous systems, producing emergent conduct that no single undertaking expressly planned.

XXIX. Emergent Conduct

Emergent conduct is perhaps the most difficult attribution category.

Imagine five competing AI pricing systems.

None is programmed to collude.

Nevertheless:

1. each observes competitors;
2. each predicts responses;
3. each learns that aggressive undercutting is unprofitable;
4. each raises prices;
5. each learns from the resulting market;
6. prices converge.

There may be a collective market outcome without explicit coordination.

Competition law must therefore distinguish:

conscious coordination

from

machine-generated convergence.

This is one of the major unresolved problems of AI competition law.

XXX. Proposed Attribution Test

A practical legal test can be expressed as:

Distributed Intelligence Attribution Test

A. Ownership
Who owns or controls the commercial system?

B. Design
Who selected its objectives?

C. Deployment
Who placed it into the market?

D. Direction
Who instructed the system?

E. Data
Who supplied competitively significant information?

F. Knowledge
Who knew or should reasonably have understood its operation and consequences?

G. Intervention
Who could modify or stop it?

H. Continuation
Who continued operating it after learning of problematic effects?

I. Benefit
Who obtained the commercial benefit?

J. Contribution
Did another actor knowingly contribute to the conduct?

The greater the accumulation of these factors, the stronger the attribution case.

XXXI. Attribution Spectrum

A useful spectrum is:

Independent AI output

↓

AI-assisted human decision

↓

AI-directed human decision

↓

AI-autonomous decision under undertaking control

↓

Multiple firms using interoperable coordination algorithms

↓

Deliberately coordinated autonomous agents

The further the system moves toward deliberate inter-firm coordination, the greater the potential Article 101/Chapter I risk.

XXXII. Remedies

If unlawful conduct is established, remedies should address the relevant attribution layer.

Possible remedies include:

Structural remedies

• separation of platform and AI services;
• divestiture;
• data separation.

Behavioural remedies

• non-discrimination;
• interoperability;
• access obligations;
• anti-steering requirements.

Algorithmic remedies

• independent audits;
• logging;
• model testing;
• explainability requirements;
• human override;
• restrictions on certain optimization objectives.

Governance remedies

• compliance officers;
• board-level AI oversight;
• incident reporting;
• algorithmic impact assessments.

XXXIII. Six Major Legal Principles Derived From the Case Law

XXXIV. Critical Legal Problem

The central challenge is that causation is becoming distributed while legal responsibility remains largely undertaking-based.

Traditional model:

Human decision → Firm → Market effect

Distributed-intelligence model:

Data → Model → Agent A → Platform → Agent B → Competitor algorithm → Market feedback → Market effect

The legal system must determine where along this chain responsibility attaches.

The answer should not be:

"Nobody is responsible because no human personally made the final decision."

Nor should it automatically be:

"Every technological participant is responsible."

Instead, competition law needs a functional attribution analysis based on control, contribution, knowledge, deployment, and economic responsibility.

Conclusion

Market behavior attribution in distributed intelligence systems represents a major evolution in competition law.

The decisive legal issue is increasingly not who physically made the decision, but:

Who designed, controlled, deployed, enabled, participated in, or knowingly benefited from the intelligent system that produced the market behavior?

The existing jurisprudence—particularly Eturas, AC-Treuhand, Anic, T-Mobile Netherlands, Hoffmann-La Roche, Intel, Google Shopping and Google Android—already provides important foundations.

The emerging doctrine should therefore recognize three propositions:

1. Autonomous execution does not automatically break attribution to the undertaking.
2. Technological causation alone does not automatically create competition-law liability.
3. Where multiple intelligent systems interact, attribution must distinguish independent adaptation from coordinated conduct.

For AI-driven markets, algorithmic logs, system architecture, objective functions, data flows, deployment decisions and human intervention records may become as legally important as traditional emails and contracts.

The future competition-law question is therefore moving from "Who agreed?" toward a more complex inquiry: "Who configured, controlled, contributed to, and knowingly participated in the system through which market behaviour emerged?"