AI “Middleware of Reality” and Epistemic Control Layers

Detailed Explanation with At Least 6 Case Laws

Important legal qualification: “AI middleware of reality” and “epistemic control layer” are not established legal categories. They are useful analytical concepts for describing AI systems that increasingly sit between people and the information, services, products, institutions, and digital environments through which they understand and act upon the world. The case law below is therefore analogical, mainly from EU competition and digital-platform jurisprudence.

1. Meaning of “AI Middleware of Reality”

The expression AI middleware of reality describes an AI system that operates as an intermediary between a person and the underlying information or service.

Traditional internet:

User → Search Engine → Websites

AI-mediated environment:

User → AI Assistant → Model/Database/Tools → Information/Services

The AI layer may:

select information;

summarize information;

rank alternatives;

decide which sources to consult;

generate answers;

suppress information;

recommend products;

execute transactions;

interact with other software;

determine what the user sees first.

Therefore, the AI does not merely store information.

It may increasingly determine how information reaches the user.

2. Meaning of “Epistemic Control”

Epistemic control means control over the conditions through which people or organizations obtain, evaluate, organize, and act upon information.

In simple terms:

Who controls the information pipeline can influence what users know, what they consider relevant, and what choices they are presented with.

An AI platform may exercise epistemic influence through:

search ranking;

source selection;

retrieval;

summarization;

filtering;

recommendation;

personalization;

confidence scoring;

automated fact selection;

tool selection;

memory;

feedback loops.

This does not mean that an AI system automatically controls people's beliefs. The legal issue is more precise: whether the platform's control over an important information intermediary gives it market power or enables exclusionary or otherwise unlawful conduct.

3. The Basic Architecture

A useful model is:

Data Layer

↓

Retrieval / Search Layer

↓

AI Model

↓

Ranking / Filtering Layer

↓

Answer / Recommendation Layer

↓

User

↓

Transaction / Action

The most important layer may be the middle.

That is why the term middleware is useful.

The AI can become the gateway through which the user accesses:

search;

news;

shopping;

travel;

finance;

education;

healthcare information;

software;

government information;

entertainment;

professional services.

4. Why Competition Law Becomes Relevant

Suppose an AI assistant becomes the main gateway through which consumers discover products.

The platform owns:

the AI assistant;

the search engine;

the advertising system;

the shopping marketplace;

payment infrastructure.

It could potentially prefer:

its own service

over

competing services.

That creates an analogy with the modern platform cases involving search, interoperability, data and self-preferencing.

The important distinction is:

Being the most-used AI intermediary is not itself an infringement.

Competition law becomes relevant when legally relevant market power is combined with conduct that satisfies the elements of an infringement.

Article 102 TFEU, for example, addresses abuse of a dominant position; dominance itself is not prohibited. The Google Shopping litigation illustrates how preferential treatment by a dominant digital platform can be examined under Article 102. (curia)

5. The “Epistemic Bottleneck”

An epistemic bottleneck arises where a small number of systems control an important pathway through which information reaches users.

Example:

10,000 websites

↓

One dominant AI assistant

↓

One answer

The user may never see the underlying 10,000 sources.

This creates several possible competition questions:

Can competing information providers reach users?

Does the AI systematically prefer affiliated services?

Are independent providers demoted?

Can competing AI systems obtain necessary data?

Are APIs available on reasonable terms?

Can users easily switch?

Can publishers opt out?

Is the AI extracting content while limiting traffic back to publishers?

6. AI as a “Gatekeeper of Relevance”

Traditional search engines rank webpages.

AI systems may go further by deciding:

Which facts are relevant enough to appear in the answer.

This creates a distinction between:

Search control

Which results are displayed?

and

Epistemic control

Which information is selected, combined, summarized and presented as the answer?

The second can be more difficult to audit because the user may receive a single synthesized response rather than a list of competing sources.

7. Key Competition-Law Risks

A. Self-preferencing

The AI may favor its own:

shopping service;

travel service;

payment system;

cloud service;

advertising platform;

content;

applications.

This is closely analogous to Google Shopping.

B. Foreclosure

Independent providers may lose access to users because the AI becomes the principal discovery mechanism.

Potential chain:

AI dominance → traffic control → reduced competitor visibility → reduced scale → weaker competitor → greater AI dominance

C. Tying

An AI assistant could make one service conditional upon another.

For example:

"To use the AI assistant's advanced functionality, users must use the platform's own search, browser, cloud or payment system."

The legal assessment would depend on the precise structure and applicable law.

D. Interoperability restrictions

The AI platform could make it difficult for competing systems to:

access data;

communicate with the assistant;

use APIs;

connect external tools;

access relevant functionality.

This brings Microsoft into the analysis.

E. Data advantage

The dominant AI intermediary may possess:

user queries;

click data;

browsing information;

transaction information;

feedback;

interaction histories.

More users generate more data.

More data can potentially improve the system.

Better performance attracts more users.

This creates a possible:

Users → Data → Better AI → More Users

feedback loop.

8. Case Law 1 — Google Shopping

Google LLC and Alphabet Inc. v European Commission

Case C-48/22 P
CJEU, 10 September 2024

This is one of the most important modern authorities for the concept.

The case concerned Google's general search service and its treatment of its own comparison-shopping service.

The Commission found that Google favored its own comparison-shopping results while competing services were demoted; the CJEU dismissed Google's appeal and upheld the General Court's judgment. (curia)

Principle

The case demonstrates that the operation of a dominant digital platform's ranking system can be examined under Article 102 where its design and operation may favor its own service and foreclose competition.

AI relevance

Imagine:

AI assistant → product question → generated recommendation

The assistant recommends:

"Buy Product A."

But Product A belongs to the AI platform's own marketplace.

If competing products are systematically disadvantaged, the Google Shopping reasoning becomes highly relevant by analogy.

Key concept

Search ranking → AI ranking → AI-generated recommendation

The technology changes, but the competition question can remain similar.

9. Case Law 2 — Microsoft

Microsoft Corp. v Commission

Case T-201/04
General Court, 17 September 2007

Microsoft concerned, among other things, refusal to provide interoperability information and the tying of Windows with Windows Media Player. (Infocuria)

Principle

The case is important for:

interoperability;

refusal to provide technical information;

tying;

technological ecosystems;

remedies.

AI relevance

Suppose a dominant AI assistant controls the central interaction layer.

Competing applications may need:

APIs;

tool interfaces;

identity systems;

model access;

contextual data;

interoperability protocols.

If access is restricted in circumstances meeting the applicable legal requirements, Microsoft provides an important analogy.

Example

Dominant AI assistant

↓

Controls tool-access layer

↓

Independent AI application cannot integrate

↓

Users remain inside dominant ecosystem

That could potentially raise interoperability concerns.

10. Case Law 3 — Eturas

Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba

Case C-74/14
CJEU, 21 January 2016

This case involved travel agencies using a common computerized booking system. The system administrator implemented an automatic restriction on online discounts, accompanied by a message concerning the restriction. The Court addressed whether the circumstances could establish a concerted practice and how the evidence should be assessed. (Infocuria)

Why this case matters for AI

Eturas is particularly useful because the competition problem occurred inside a computerized system.

Modern AI systems can similarly establish:

common recommendation rules;

automated pricing;

common algorithmic constraints;

automated communication;

system-generated coordination.

Important distinction

An AI system producing similar outcomes does not by itself prove an agreement or concerted practice.

The legal question remains whether the evidence establishes the elements required by Article 101.

Formula

Common AI system + communication/rule + participant awareness + conduct

↓

Potential Article 101 investigation

11. Case Law 4 — T-Mobile Netherlands

T-Mobile Netherlands BV and Others v Raad van bestuur van de Nederlandse Mededingingsautoriteit

Case C-8/08
CJEU, 4 June 2009

The case concerned the concept of a concerted practice and the significance of information exchanged between competitors. The CJEU held that, in the circumstances of the case, even a single meeting could constitute a concerted practice. (Infocuria)

AI relevance

Consider competing companies using a common AI system that provides information concerning:

future prices;

demand;

capacity;

customers;

strategic plans.

An AI intermediary could therefore become a communication and information-exchange layer.

The important competition-law question would be whether the information exchange and subsequent conduct satisfy Article 101 requirements.

Important distinction

Algorithmic interaction ≠ automatically cartel.

Evidence of coordination still matters.

12. Case Law 5 — Meta Platforms

Meta Platforms Ireland Ltd and Others v Bundeskartellamt

Case C-252/21
CJEU, 4 July 2023

The case concerned Facebook's processing and combination of user data from Facebook and other Meta services and third-party websites/apps.

The CJEU held that a national competition authority can, in examining abuse of dominance, take account of GDPR compliance issues, while respecting the roles of the competent data-protection authorities. (curia)

AI relevance

AI assistants can potentially combine:

user queries;

search history;

location;

browsing;

purchases;

communications;

third-party application information.

That can create substantial information advantages.

Key principle

Competition law and data protection can interact.

Therefore:

AI dominance + extensive data processing

may require analysis under both:

competition law;

data-protection law.

But a data-protection issue does not automatically establish competition-law abuse.

13. Case Law 6 — RTE and ITP / Magill

RTE and ITP v Commission

Joined Cases C-241/91 P and C-242/91 P
CJEU, 6 April 1995

The cases concerned television programme information and copyright.

The Court dealt with circumstances in which refusal to license protected information could amount to an abuse of dominance. (Infocuria)

AI relevance

This becomes important when AI systems depend on information controlled by another undertaking.

Imagine a dominant AI provider needs access to:

databases;

copyrighted archives;

structured information;

specialized datasets.

The question could become:

Can control over information be used in a way that unlawfully restricts downstream competition?

Magill is important because it demonstrates that intellectual-property rights and information control can intersect with Article 102.

14. Case Law 7 — IMS Health

IMS Health GmbH & Co. OHG v NDC Health GmbH & Co. KG

Case C-418/01
CJEU, 29 April 2004

IMS Health concerned a pharmaceutical information structure and access to a system protected by intellectual-property rights.

Principle

The case developed the stringent circumstances under which refusal to license/access may become abusive.

AI relevance

Suppose an AI platform possesses a unique:

knowledge graph;

industry dataset;

user-information architecture;

specialized database.

Competitors may argue that they cannot compete without it.

IMS Health demonstrates why "important" and "valuable" are not necessarily the same as legally indispensable.

15. Case Law 8 — Post Danmark

Post Danmark A/S v Konkurrencerådet

Case C-209/10
CJEU, 27 March 2012

The case concerned selective low prices and possible exclusionary effects by a dominant undertaking. The Court considered actual or likely exclusion and objective justification. (Infocuria)

AI relevance

An AI platform might provide its AI service:

free of charge

while using the service to exclude competitors in another market.

The apparent "zero price" therefore does not necessarily end the competition analysis.

The investigation may need to consider:

quality;

access;

data extraction;

traffic;

advertising;

complementary markets;

exclusionary effects.

16. Case Comparison Table

17. The “Reality Middleware” Problem

Traditional digital intermediaries primarily connected users to other services.

AI middleware may increasingly interpret the world for the user.

Consider:

Old model

User:

"Which laptop should I buy?"

Search engine:

10,000 results.

AI middleware

User:

"Which laptop should I buy?"

AI:

"Buy Model X."

The AI has potentially compressed:

10,000 information points → 1 recommendation.

That compression creates efficiency, but it also creates a possible bottleneck.

18. Epistemic Ranking

The AI may perform several forms of ranking:

Source ranking

Which source is consulted?

Fact ranking

Which facts are considered important?

Answer ranking

Which facts are included?

Product ranking

Which products are recommended?

Action ranking

Which action is suggested?

Tool ranking

Which external service is called?

Therefore:

The AI may control not merely visibility but the sequence of decisions leading to action.

19. The “Answer Layer” as a Competitive Bottleneck

Imagine:

Publisher A
Publisher B
Publisher C
Publisher D

↓

AI assistant

↓

One synthesized answer

The publishers may become dependent on the AI for:

traffic;

visibility;

customers;

subscriptions;

commercial discovery.

If the AI changes its ranking or citation rules, traffic may shift dramatically.

This can create a new type of intermediation power.

20. AI Self-Preferencing

A particularly important hypothetical is:

User asks:
"Find me a hotel."

AI:

"Here are three hotels."

But all three are from the platform's own booking service.

Or:

User asks:
"Compare financial products."

The AI only displays products supplied through its affiliated financial marketplace.

Or:

User asks:
"Find legal research."

The AI preferentially uses its own legal database.

The Google Shopping case provides an important analytical analogy because the CJEU examined the competitive effects and causal relationship associated with preferential treatment in a dominant search platform. (Infocuria)

21. AI Hallucination and Competition Law

Hallucination itself is generally not automatically a competition-law infringement.

However, competition concerns could arise if a dominant platform systematically:

misrepresents competing products;

suppresses competitors;

gives inaccurate information about rivals;

promotes affiliated services through manipulated outputs.

The relevant distinction is:

Accidental AI error

versus

deliberate or systematically exclusionary platform conduct.

Competition law primarily asks whether the latter satisfies the applicable legal elements.

22. Epistemic Manipulation Through Personalization

AI can produce different answers for different users.

For example:

User A → Recommendation X

User B → Recommendation Y

because of:

purchasing history;

income;

location;

interests;

previous searches;

behavioural profile.

This creates possible concerns involving:

discrimination;

exploitation;

privacy;

consumer protection;

competition.

Meta demonstrates how data practices can intersect with competition analysis when a dominant undertaking combines information from different sources. (curia)

23. Feedback Loops

One of the most important characteristics of AI middleware is the feedback loop.

Stage 1

More users ask questions.

Stage 2

The AI receives more interaction data.

Stage 3

The system improves.

Stage 4

Users become more dependent.

Stage 5

Competitors receive less traffic.

Stage 6

The AI becomes even more important.

This can be represented as:

Users ↑ → Data ↑ → AI Quality ↑ → User Dependence ↑ → Users ↑

This is not automatically anticompetitive.

But it can create high entry barriers.

24. Epistemic Lock-In

Traditional lock-in:

"My files are stored on Platform A."

AI lock-in:

"The AI understands my preferences, history and workflows."

The user may have difficulty moving because the new AI lacks:

conversation history;

personal preferences;

contextual memory;

customized workflows;

connected applications;

learned interaction patterns.

Therefore:

Data portability + model portability + memory portability

could become important competitive issues.

25. Interoperability and AI Agents

Future AI systems may interact with each other.

For example:

User AI

↓

Bank AI

↓

Travel AI

↓

Insurance AI

↓

Payment AI

A dominant intermediary might control access between these systems.

This creates a new potential layer:

Agent-to-agent interoperability.

Microsoft's interoperability reasoning becomes conceptually relevant here, although the factual and legal context is very different. (Infocuria)

26. AI as a Vertical Integration Engine

A single AI company could potentially operate:

Foundation model

↓

AI assistant

↓

Search

↓

Advertising

↓

Shopping

↓

Payments

↓

Cloud

↓

Applications

The more layers controlled by one undertaking, the greater the possibility of leveraging market power from one layer into another.

Google Shopping provides a modern example of a digital platform leveraging an important position in general search into a related specialized-search service. (Infocuria)

27. Possible Article 102 Theories

Depending on the facts, an investigation could examine:

1. Self-preferencing

AI gives affiliated services preferential treatment.

2. Refusal to supply

Competitors are denied indispensable access.

3. Interoperability restriction

Competing AI tools cannot effectively connect.

4. Tying

Use of one service requires another.

5. Bundling

AI + cloud + search + advertising are combined.

6. Exclusive arrangements

Users or suppliers are prevented from using competitors.

7. Discriminatory access

Competitors receive worse technical or commercial terms.

8. Predatory pricing

AI is offered under conditions designed to eliminate competitors.

9. Margin squeeze

Upstream AI infrastructure is expensive while downstream services are offered cheaply.

28. Possible Article 101 Issues

AI middleware can also facilitate coordination between independent businesses.

For example:

Competitor A

↓

AI platform

↓

Competitor B

If the system allows competitors to receive or use sensitive information regarding:

future prices;

capacity;

inventory;

demand;

strategic plans,

Article 101 concerns may arise.

T-Mobile Netherlands is relevant to the legal treatment of concerted practices and information exchange. (Infocuria)

29. Evidence in an AI Competition Investigation

Traditional evidence may not be enough.

Investigators may need:

Technical evidence

model architecture;

system prompts;

ranking algorithms;

API logs;

retrieval logs.

Commercial evidence

contracts;

exclusivity clauses;

pricing;

rebates;

platform terms.

Data evidence

training datasets;

user data;

click data;

interaction histories.

Output evidence

recommendations;

citations;

rankings;

competitor visibility.

Economic evidence

traffic changes;

switching rates;

foreclosure;

entry barriers;

counterfactual analysis.

30. Counterfactual Analysis

A central question can be:

What would users have seen if the platform had not used the allegedly exclusionary mechanism?

For example:

Actual world

AI ranks affiliated service first.

Counterfactual

AI ranks all services using neutral criteria.

The authority may compare:

traffic;

conversions;

market shares;

consumer choice;

competitor viability.

Google Shopping expressly involved issues such as potential anticompetitive effects, causal link, counterfactual scenarios and foreclosure capability. (Infocuria)

31. Consumer Harm

Potential harms may include:

fewer choices;

reduced innovation;

higher prices;

lower quality;

reduced privacy;

reduced access to information;

lower diversity of suppliers.

But the mere fact that an AI gives one answer instead of ten does not establish consumer harm.

AI aggregation can also produce genuine efficiencies:

lower search costs;

easier comparison;

personalized assistance;

faster information retrieval.

Competition analysis must therefore distinguish efficient intermediation from exclusionary conduct.

32. Epistemic Control and Freedom of Choice

There is a broader policy question:

If an AI becomes the principal interface between users and information, should users be able to see alternative sources?

Possible regulatory mechanisms include:

source transparency;

citation requirements;

user choice;

interoperability;

data portability;

ranking transparency;

auditing;

opt-out mechanisms.

These are policy questions that can interact with competition law but are not identical to Article 102 analysis.

33. The “Reality Stack”

A useful conceptual model is:

The greater the number of layers controlled by one company, the greater the potential for vertical leveraging.

34. AI Middleware vs Traditional Search

This difference explains why AI could become a more significant economic and informational intermediary.

35. Key Legal Distinction

It is important to avoid this equation:

AI control of information = illegal monopoly

That is legally incorrect.

The appropriate analysis is:

Relevant market

↓

Dominance

↓

Specific conduct

↓

Foreclosure / exploitation / competitive harm

↓

Causation

↓

Objective justification / efficiencies

↓

Legal conclusion

36. Major Risks

Risk 1 — Information foreclosure

Competitors become invisible.

Risk 2 — Self-preferencing

AI promotes affiliated products.

Risk 3 — Data concentration

One platform accumulates exceptional datasets.

Risk 4 — Interoperability barriers

Other AI systems cannot connect.

Risk 5 — User lock-in

Personal context becomes difficult to transfer.

Risk 6 — Algorithmic discrimination

Different competitors receive systematically different treatment.

Risk 7 — Vertical leveraging

Power in one market is used in another.

Risk 8 — Algorithmic coordination

AI systems facilitate coordination among competitors.

37. Important Case-Law Matrix

38. Exam Formula

AI Middleware of Reality

Data + AI Model + Retrieval + Ranking + Personalization + Recommendation + User Interface

↓

Control of Information Flow

↓

Potential Epistemic Bottleneck

↓

User Dependency + Data Feedback + Network Effects

↓

Potential Market Power

↓

Self-Preferencing / Tying / Refusal / Interoperability Restriction / Discrimination

↓

Potential Competition-Law Concern

↓

Dominance + Conduct + Effects + Causation + Justification

39. Ultra-Simple Explanation

In very simple language:

AI middleware of reality means AI becomes the middle layer through which people access information and services.

Epistemic control means the AI may influence which information becomes visible, relevant or actionable.

For competition law, the central concern is not simply that AI influences information.

The question is whether a powerful AI intermediary can use its position to:

exclude competitors;

favor its own services;

restrict interoperability;

control essential information;

exploit data advantages;

tie users to its ecosystem;

facilitate coordination.

40. Six Cases to Remember for Exams

Google Shopping — C-48/22 P
→ Digital self-preferencing / ranking.

Microsoft — T-201/04
→ Interoperability + tying.

Eturas — C-74/14
→ Computerized system + concerted practice.

T-Mobile Netherlands — C-8/08
→ Information exchange + concerted practice.

Meta Platforms — C-252/21
→ Data processing + competition law.

Magill — C-241/91 P & C-242/91 P
→ Information control + refusal to license.

Memory formula:

Google = Ranking
Microsoft = Interoperability
Eturas = Algorithmic system
T-Mobile = Information
Meta = Data
Magill = Knowledge

Conclusion

AI “middleware of reality” describes the growing role of AI as an intermediary between users and the underlying information, markets and digital services they access. Epistemic control layers describe the mechanisms—retrieval, ranking, filtering, summarization and recommendation—through which an AI system can influence what information reaches the user.

From a competition-law perspective, the most important development is the movement from:

"AI helps me find information"

toward:

"AI determines what information, products or services I encounter and may act upon."

The existing case law does not establish a standalone doctrine of "epistemic control." Instead, cases such as Google Shopping, Microsoft, Eturas, T-Mobile, Meta, Magill and IMS Health provide different legal building blocks for analysing self-preferencing, interoperability, information exchange, data concentration and control over important information resources. (Infocuria)

The central formula is:

AI Middleware + Information Bottleneck + Network Effects + Data Advantage + Market Power + Exclusionary Conduct = Potential Competition-Law Problem

But:

Information influence alone ≠ dominance, and dominance alone ≠ abuse.

The specific market, conduct, effects, evidence and possible objective justifications must be examined in each case.