Competition Law and Competition Audits of AI Systems

1. Introduction

Artificial Intelligence (AI) systems increasingly influence pricing, product ranking, recommendations, advertising, search, allocation of resources, procurement, credit, hiring, logistics, content distribution and customer access. As AI becomes embedded in commercial decision-making, competition-law risks may arise not only from traditional agreements between firms but also from the design, training data, inputs, outputs and deployment of algorithms.

A competition audit of an AI system is a structured legal, technical and economic examination designed to identify whether an AI system could:

• facilitate collusion or coordinated conduct;
• exchange or exploit competitors' competitively sensitive information;
• discriminate against competitors;
• exclude rivals from markets;
• self-preference the operator's products;
• create or strengthen market power;
• impose tying or interoperability restrictions;
• facilitate predatory or exclusionary pricing;
• foreclose access to important data or infrastructure;
• create barriers to entry or switching;
• undermine independent competitive decision-making.

The central principle is that automation does not remove competition-law responsibility. Conduct that would be unlawful when implemented by humans may remain unlawful when implemented through software or AI. The U.S. DOJ and FTC, for example, have expressly stated that competitors cannot evade antitrust law by using algorithms to coordinate prices.

2. Meaning of a Competition Audit of an AI System

A competition audit is broader than an ordinary technical AI audit.

A technical AI audit may ask:

Is the model accurate, secure, explainable and unbiased?

A competition audit asks:

Could the way this AI system is designed, trained, operated or integrated distort competition?

Thus, the audit examines the entire AI competition lifecycle:

Data → Training → Model → Deployment → Business Rules → Competitor Interaction → Market Effects → Monitoring

For example, suppose five competing landlords provide confidential rental information to a common AI pricing platform. The system processes the information and recommends higher rents to all five.

The competition audit would investigate:

1. What competitor information enters the system?
2. Is that information competitively sensitive?
3. Who controls the algorithm?
4. Are competitors aware of the information supplied by others?
5. Does the algorithm recommend aligned prices?
6. Can competitors independently determine prices?
7. Does the system contain rules encouraging price alignment?
8. Are outputs monitored by humans?
9. Does the system facilitate market-wide coordination?
10. Are there safeguards against such coordination?

This is closely connected to the current U.S. RealPage litigation and settlements concerning algorithmic rental pricing.

3. Legal Framework

A. Prohibition of Anti-Competitive Agreements

AI can facilitate agreements between competitors even where the competitors do not communicate directly with one another.

Potentially problematic conduct includes:

• common pricing algorithms;
• common AI recommendation systems;
• exchange of competitively sensitive data;
• coordinated output decisions;
• algorithms designed to match competitors' prices;
• AI systems programmed to respond to competitors' prices in a coordinated manner.

The relevant legal provisions depend upon the jurisdiction.

United States

Important provisions include:

• Sherman Act §1 — agreements restraining trade;
• Sherman Act §2 — monopolization and attempted monopolization;
• Clayton Act §7 — anti-competitive mergers;
• FTC Act §5 — unfair methods of competition.

European Union

Important provisions include:

• Article 101 TFEU — anti-competitive agreements and concerted practices;
• Article 102 TFEU — abuse of dominant position;
• EU merger-control rules;
• Digital Markets Act where applicable.

India

Relevant provisions include:

• Competition Act 2002, §3 — anti-competitive agreements;
• §4 — abuse of dominant position;
• §5 and §6 — combinations;
• §19 — investigation by the CCI;
• §26 — investigation procedure;
• §27 — orders against anti-competitive conduct;
• §32 — conduct occurring outside India but having an effect on competition in India.

4. Major Competition Risks Requiring an AI Competition Audit

4.1 Algorithmic Collusion

The most important risk is that competing businesses use a common or interconnected AI system that produces coordinated outcomes.

Traditional cartel:

Human A communicates with Human B → they agree on price.

Algorithmic cartel:

Competitor A and Competitor B supply data to the same system → AI generates aligned pricing recommendations.

The absence of a conventional meeting or telephone call does not necessarily eliminate antitrust risk.

The DOJ and FTC have specifically emphasized that competitors cannot use algorithms as a mechanism to accomplish conduct that would otherwise violate antitrust law.

5. Case Laws

1. United States v. Topkins — Algorithmic Pricing on Amazon Marketplace

This is one of the foundational algorithmic-collusion cases.

An online poster retailer and competitors agreed to coordinate prices for products sold through an online marketplace. Pricing software was used to implement the agreement and automatically respond to competitors' prices.

Principle

The use of software did not transform a conventional price-fixing agreement into lawful independent pricing.

The case demonstrates:

An algorithm can be the instrument of a cartel; it is not a defence to cartel liability.

Audit implication

An AI competition audit should identify:

• competitor-price inputs;
• automatic price matching;
• price-following rules;
• communication between competing users;
• common algorithm providers;
• automatic price adjustments.

The DOJ has previously described the Topkins conduct as an example where differently programmed algorithms nevertheless operated to accomplish a coordinated pricing objective.

2. Eturas v Lietuvos Respublikos konkurencijos taryba, C-74/14

Facts

The case concerned an online travel-booking platform through which travel agencies received a common electronic message imposing a maximum discount on bookings.

The European Court of Justice examined when participants using a common electronic system could be regarded as participating in a concerted practice.

Principle

Electronic systems can be relevant evidence of concerted conduct.

A competition authority may examine:

• the electronic message;
• knowledge of participants;
• conduct following the message;
• participation in the system;
• evidence showing whether firms distanced themselves from the arrangement.

Importance for AI audits

An AI audit should preserve:

• system messages;
• prompts;
• model instructions;
• configuration changes;
• administrator communications;
• API calls;
• model-generated recommendations;
• records showing whether users accepted or rejected recommendations.

Audit lesson: digital coordination may leave an electronic evidence trail even when there is no traditional cartel meeting.

6. RealPage Algorithmic Pricing Litigation

3. United States v. RealPage Inc.

The RealPage litigation is particularly important for AI and algorithmic competition compliance.

The DOJ alleged that landlords supplied competitively sensitive information concerning rents, vacancies and lease terms to RealPage's pricing software and received pricing recommendations in return. The DOJ alleged violations of Sections 1 and 2 of the Sherman Act.

The later settlement concerning RealPage imposed restrictions concerning the sharing and use of competitively sensitive information and algorithmic pricing.

Competition principle

A company cannot necessarily avoid antitrust scrutiny by saying:

"The price was generated by software rather than decided by a human."

Audit requirements

An AI competition audit should therefore examine:

The DOJ's later compliance measures concerning RealPage illustrate the importance of data age, aggregation, monitoring and compliance controls in algorithmic-pricing systems.

7. Duffy v. Yardi Systems

4. Duffy v. Yardi Systems, Inc.

This litigation concerns allegations that competing landlords used a common algorithmic pricing system to coordinate rental prices.

A federal court allowed algorithmic price-fixing allegations to proceed at an important stage of the litigation.

The case is significant because the alleged coordination occurs through a software intermediary rather than a conventional direct cartel meeting.

Competition lesson

Competition auditors should distinguish between:

Independent algorithmic pricing

Each company independently develops and operates its own model.

and

Potentially problematic algorithmic coordination

Multiple competitors provide sensitive information to a common system that generates coordinated pricing recommendations.

The DOJ has identified Duffy as an important development in the developing jurisprudence concerning algorithmic price fixing.

8. Cornish-Adebiyi v. Caesars Entertainment

5. Cornish-Adebiyi v. Caesars Entertainment

This case concerns allegations involving hotel pricing algorithms.

The DOJ and FTC filed a statement of interest explaining that hotels cannot use an algorithm to accomplish conduct that would violate antitrust law if accomplished manually.

The court's treatment also demonstrates an important limitation:

Alleging the use of an algorithm is not automatically sufficient to establish an unlawful agreement.

There must still be legally sufficient evidence concerning the necessary elements of the antitrust claim.

Competition-audit significance

An AI audit should therefore avoid assuming:

"Algorithm + similar prices = cartel."

Instead, it should investigate:

• whether there is an agreement;
• communications between participants;
• information sharing;
• common instructions;
• common pricing parameters;
• knowledge of competitors' participation;
• actual algorithmic outputs;
• implementation of recommendations.

The DOJ specifically noted that courts have required adequate allegations concerning the necessary agreement among competitors.

9. Gibson v. Cendyn Group

6. Gibson v. Cendyn Group

Another hotel algorithm case concerns allegations that competitors used a common revenue-management technology.

The litigation illustrates the distinction between:

Mere parallel conduct

Several competitors independently use similar technology and consequently produce similar outcomes.

and

Concerted conduct

Competitors knowingly participate in an arrangement facilitated by a common technology provider.

The DOJ has identified Gibson v. Cendyn Group alongside Cornish-Adebiyi as an important example of courts examining whether algorithmic pricing allegations adequately establish a "rim agreement" between competing firms.

Audit significance

The auditor should therefore examine both:

Hub

The AI/software provider.

Spokes

The individual competing businesses.

The key question becomes:

Is the AI merely a technological tool, or does the structure of its operation facilitate concerted conduct?

10. Google Shopping — Google Search (EU)

7. Google Search (Shopping)

The European Commission's Google Shopping case is not an AI case in the narrow sense, but it is highly relevant to AI competition audits because it establishes principles concerning self-preferencing and platform-controlled ranking systems.

Google was found to have favoured its own comparison-shopping service in search results over competing comparison-shopping services.

Competition principle

A dominant digital platform may face competition-law scrutiny when its control over an important platform infrastructure is used to disadvantage competing services.

AI audit relevance

Modern AI systems frequently determine:

• search results;
• recommendation rankings;
• product visibility;
• advertising placement;
• content recommendations;
• marketplace exposure.

An AI competition audit should therefore ask:

Does the model systematically favour the platform's own downstream products or services?

Potential indicators include:

• unexplained ranking advantages;
• preferential access to data;
• differential API access;
• preferential model integration;
• exclusion of rival services;
• manipulation of recommendation outputs.

11. Microsoft — Tying and Digital Ecosystems

8. Microsoft Corp. v. Commission / Microsoft Competition Cases

The Microsoft competition litigation provides important principles concerning tying, interoperability and exclusionary conduct.

Although the original cases predate generative AI, the principles are increasingly relevant to AI ecosystems.

For example, an AI provider might combine:

operating system + cloud + AI model + productivity software + search + advertising.

A competition audit must examine whether customers or rivals are effectively compelled to use one component to obtain another.

Questions for an AI competition audit

• Is the AI model technically tied to another product?
• Are APIs restricted?
• Are competing AI models prevented from functioning?
• Is interoperability deliberately degraded?
• Are customers prevented from switching?
• Is proprietary data unavailable to competitors?
• Are rivals denied access to essential technical interfaces?

12. Competition Audit Framework for AI Systems

A comprehensive audit can be divided into 10 stages.

Stage 1 — Identify the Relevant Market

Determine:

• product market;
• geographic market;
• upstream/downstream markets;
• AI model market;
• data market;
• infrastructure market;
• application market;
• distribution market.

For example:

Cloud infrastructure → Foundation model → AI API → Application → Consumer

Competition concerns can arise at several levels.

13. Stage 2 — Determine Market Power

The auditor should assess:

• market share;
• concentration;
• entry barriers;
• switching costs;
• network effects;
• data advantages;
• compute access;
• intellectual-property rights;
• interoperability;
• customer lock-in;
• economies of scale;
• ecosystem effects.

A company with a small market share may present relatively limited dominance concerns, whereas a platform controlling an essential input may require considerably greater scrutiny.

14. Stage 3 — Audit Training Data

Training data can itself generate competition concerns.

Questions

1. Where did the data originate?
2. Does it contain competitors' confidential information?
3. Was it obtained lawfully?
4. Is data access exclusive?
5. Does the company control a strategically important dataset?
6. Is the data unavailable to competitors?
7. Does the model reproduce competitively sensitive information?
8. Can the model infer confidential business information?

Competition risk

A dominant firm could potentially use exclusive control over strategically important data to create barriers to entry.

15. Stage 4 — Audit Competitor Information

This is particularly important for pricing AI.

The audit should identify whether the system receives:

• competitor prices;
• discounts;
• inventories;
• output levels;
• customer lists;
• production costs;
• future pricing plans;
• vacancies;
• capacity information;
• strategic forecasts.

Competitively sensitive information should generally receive heightened scrutiny.

16. Stage 5 — Audit AI Inputs and Prompts

For generative AI, the audit should examine:

• system prompts;
• developer instructions;
• user prompts;
• retrieved data;
• APIs;
• external databases;
• competitor information;
• automated tools;
• agent instructions.

For example:

"Check competitors A, B and C and determine what price we should charge."

may require substantially greater scrutiny than:

"Calculate our internal cost-based price."

17. Stage 6 — Audit Model Outputs

The auditor should test whether the AI:

• recommends identical prices;
• systematically follows competitors;
• suppresses competitors;
• recommends exclusionary contracts;
• ranks affiliated products disproportionately;
• restricts interoperability;
• recommends tying;
• discriminates against rival distributors;
• allocates customers improperly.

This requires output testing, not merely document review.

18. Stage 7 — Audit Feedback Loops

AI systems learn from outcomes.

This creates a particular competition concern.

Example:

Competitor A raises price → AI observes it → AI raises Company's price → Competitor B's AI observes this → B raises price → A's AI observes B → prices converge

Even without direct human communication, repeated algorithmic interaction may create competition concerns depending upon the facts and applicable law.

Therefore the audit should examine:

• reinforcement learning;
• automated feedback;
• competitor monitoring;
• real-time price updates;
• automated reactions;
• common data feeds;
• model retraining.

19. Stage 8 — Audit Self-Preferencing

A platform-controlled AI system can potentially favour its own products.

For example:

AI shopping assistant → recommends platform's private-label products → competing sellers receive lower visibility.

Audit questions include:

• Are affiliated products treated differently?
• Are ranking criteria transparent?
• Are competitors disadvantaged?
• Does the AI have access to competitors' confidential data?
• Is proprietary data used to compete against marketplace sellers?

20. Stage 9 — Audit Tying and Bundling

AI ecosystems frequently operate through bundles:

Cloud + AI model + storage + enterprise software

The audit should examine whether:

• customers are forced to purchase AI with another service;
• rival AI systems are technically blocked;
• discounts are conditional;
• APIs are restricted;
• interoperability is deliberately impaired;
• switching is artificially difficult.

21. Stage 10 — Continuous Monitoring

An AI competition audit should not be a one-time exercise.

AI systems change through:

• model updates;
• retraining;
• new datasets;
• new APIs;
• new agents;
• new integrations;
• changed system prompts;
• acquisition of competitors;
• changes in market structure.

Accordingly, companies should establish:

Continuous AI Competition Compliance Monitoring

rather than relying exclusively upon an annual legal audit.

22. Competition Audit Checklist

23. Evidence Preservation

AI competition investigations may involve evidence that traditional compliance programmes overlook.

Companies should preserve:

• model versions;
• training datasets;
• data provenance records;
• prompts;
• system instructions;
• API calls;
• model outputs;
• logs;
• pricing recommendations;
• model-change histories;
• employee communications;
• contracts with AI providers;
• competitor-data agreements;
• governance approvals;
• model evaluation reports.

This is particularly important because AI decisions may otherwise be difficult to reconstruct after a model has been updated.

24. Role of Legal, Technical and Economic Teams

An effective competition audit should be interdisciplinary.

Legal team

Examines:

• §§ 3 and 4 Competition Act;
• Article 101/102 TFEU;
• Sherman Act;
• merger rules;
• sectoral regulation.

Economists

Examine:

• market definition;
• market power;
• concentration;
• pricing effects;
• foreclosure;
• efficiencies;
• consumer harm.

AI/technical team

Examines:

• model architecture;
• training data;
• prompts;
• algorithms;
• APIs;
• model outputs;
• feedback loops.

Compliance team

Examines:

• policies;
• employee training;
• approval processes;
• monitoring;
• incident response.

25. Red-Flag Indicators

The following should trigger heightened legal review:

Red Flag 1

A competitor supplies confidential pricing information to the same AI provider.

Red Flag 2

An AI automatically matches competitor prices.

Red Flag 3

The system recommends prices using real-time competitor data.

Red Flag 4

Several competitors use identical AI pricing instructions.

Red Flag 5

An AI platform controls access to a strategically important dataset.

Red Flag 6

A dominant platform's AI systematically ranks its own products higher.

Red Flag 7

Competitors are prohibited from accessing an important AI API.

Red Flag 8

AI services are technically tied to another dominant product.

Red Flag 9

AI outputs automatically exclude competing products.

Red Flag 10

Employees communicate with competitors about AI parameters, pricing models or commercially sensitive inputs.

26. Safe-Harbour-Oriented Controls

Companies can reduce competition risks through:

1. Competitively sensitive data controls
2. Data aggregation and anonymisation
3. Restrictions on real-time competitor information
4. Independent pricing authority
5. Human approval for high-risk AI recommendations
6. Competition-law review of common AI platforms
7. API-access policies
8. Non-discrimination testing
9. Periodic self-preferencing tests
10. Model-change approval procedures
11. Competition compliance training
12. Automated red-flag monitoring
13. Independent AI competition audits
14. Whistleblower mechanisms
15. Legal review before deploying AI across competitor-facing markets.

27. AI Competition Audit — Three-Level Risk Model

Level I — Low Risk

AI used for:

• internal document classification;
• administrative automation;
• cybersecurity;
• ordinary forecasting;
• internal productivity.

Competition-law review remains appropriate but generally focuses on ordinary business conduct.

Level II — Medium Risk

AI used for:

• pricing recommendations;
• customer allocation;
• marketplace ranking;
• supplier selection;
• advertising;
• competitor monitoring.

Periodic competition audits should be undertaken.

Level III — High Risk

AI:

• receives competitors' confidential data;
• determines market prices;
• coordinates multiple competitors;
• controls an essential platform;
• ranks competitors;
• restricts interoperability;
• is operated by a dominant firm;
• is integrated into several downstream markets.

Such systems warrant continuous legal, economic and technical monitoring.

28. Important Legal Distinction: Parallel Pricing Is Not Automatically Collusion

This distinction is essential.

Suppose three competitors independently deploy AI systems and all three increase prices after observing market conditions.

That fact alone does not automatically establish an unlawful agreement.

The audit must investigate:

• communications;
• information exchanges;
• common providers;
• contractual arrangements;
• algorithmic instructions;
• knowledge;
• intent where legally relevant;
• implementation;
• market structure;
• economic evidence.

This is illustrated by the developing litigation concerning hotel and rental-pricing algorithms, where courts have considered whether allegations sufficiently establish an agreement among competitors.

29. Relationship Between AI Audit and Competition Impact Assessment

A mature AI governance programme should therefore contain a separate:

Competition Impact Assessment (CIA)

before deployment.

It should answer:

A. Market

What competitive market does the AI affect?

B. Power

Does the company have substantial market power?

C. Data

Does the AI use competitors' sensitive information?

D. Conduct

Could the AI facilitate coordination or exclusion?

E. Effects

Could competitors be foreclosed?

F. Consumers

Could prices rise, quality decline or choice decrease?

G. Remedies

Can technical safeguards reduce the risk?

H. Monitoring

Who will continuously supervise the system?

30. Six+ Key Cases at a Glance

The DOJ's recent enforcement activity demonstrates that algorithmic coordination remains an active competition-enforcement area, including proposed settlements involving landlords using common pricing systems and competitively sensitive information.

31. Conclusion

Competition audits of AI systems represent the intersection of antitrust law, data governance, algorithmic accountability, economics and technology governance.

The central principle is:

An AI system does not operate outside competition law merely because the competitive decision is automated.

A comprehensive competition audit should therefore examine the entire chain:

Data → Training → Algorithm → Model → Instructions → Output → Competitor Interaction → Market Effects → Monitoring

The most important areas are algorithmic collusion, competitively sensitive data, common AI providers, pricing algorithms, self-preferencing, exclusionary ranking, tying, interoperability, data foreclosure and AI-driven market concentration.

The emerging cases involving Topkins, Eturas, RealPage, Yardi, Caesars and Cendyn demonstrate that competition authorities and courts are increasingly examining the technological mechanism through which coordination occurs rather than limiting antitrust analysis to traditional human-to-human communications. The DOJ has expressly emphasized that competitors cannot use algorithms to accomplish conduct that would be unlawful if performed directly by humans.

Accordingly, an effective corporate AI competition programme should combine legal review + economic analysis + technical model auditing + continuous monitoring, with heightened controls whenever an AI system influences prices, competitor interactions, access, rankings or strategically important data.