AI Literature Synthesis Systems and Knowledge Filtering Power

Introduction

AI literature-synthesis systems—such as AI research assistants, scholarly search engines, retrieval-augmented generation systems, citation-analysis platforms, and generative-AI tools that summarize academic literature—can occupy an important information-intermediation position.

Their competitive significance is different from that of an ordinary database. A traditional database may allow a researcher to retrieve 500 papers. An AI synthesis system may determine:

• which papers are retrieved;
• which papers are omitted;
• which sources are treated as authoritative;
• how conflicting findings are reconciled;
• which citations are displayed;
• how literature is summarized;
• what terminology or conceptual framework is presented;
• and ultimately what the user believes the literature says.

This creates a possible competition-law issue that can be described as knowledge-filtering power: the ability of an intermediary with substantial market power to influence access to, visibility of, ranking of, and interpretation of information.

The legal issue is not whether an AI system is "wrong" merely because it produces an imperfect synthesis. Competition law becomes relevant where control over information, data, ranking, distribution, interoperability, or access is used to exclude rivals, disadvantage dependent businesses, foreclose alternative sources, or leverage dominance from one market into another.

I. Meaning of AI Literature Synthesis Systems

An AI literature-synthesis system can perform several functions simultaneously:

1. Discovery – locating papers, books, datasets and research outputs.
2. Ranking – determining which sources appear first or receive greater prominence.
3. Filtering – excluding sources according to relevance, quality, commercial or technical criteria.
4. Summarisation – compressing large bodies of literature.
5. Synthesis – combining conclusions from multiple sources.
6. Citation generation – selecting sources that supposedly support propositions.
7. Interpretation – explaining disagreements or methodological differences.
8. Recommendation – suggesting what a researcher should read next.
9. Personalisation – modifying results according to user history or institutional context.
10. Knowledge graph construction – establishing relationships among authors, papers, concepts and institutions.

Consequently, the system can become a gatekeeper between the research ecosystem and the researcher.

II. What Is "Knowledge Filtering Power"?

Knowledge-filtering power is not itself a conventional statutory category of competition law.

It is better understood as a competitive theory of harm involving control over information flows.

A powerful AI literature platform could potentially influence competition through:

1. Ranking power

A system can determine which research appears prominently.

2. Inclusion/exclusion power

It can determine which publishers, repositories, journals or databases are incorporated into its corpus.

3. Citation power

The system can influence which researchers and publications receive citations and visibility.

4. Interpretation power

A synthesis model may transform a disputed research field into a seemingly settled conclusion.

5. Interface power

Researchers may increasingly interact with the synthesis layer rather than visiting underlying databases.

6. Data-access power

A platform controlling a large scholarly corpus may possess a dataset difficult for competitors to reproduce.

7. Distribution power

If an AI system becomes embedded into browsers, operating systems, university platforms or productivity software, competitors may find it difficult to reach users.

III. Relevant Competition-Law Framework

Several legal doctrines become relevant.

A. Article 102 TFEU / Abuse of Dominance

Potential theories include:

• discriminatory access;
• refusal to supply;
• exclusionary ranking;
• self-preferencing;
• tying and bundling;
• leveraging;
• discriminatory interoperability;
• exploitative contractual terms;
• exclusionary data practices.

The central question would normally be whether the undertaking possesses dominance in a properly defined relevant market and whether the conduct produces or is capable of producing exclusionary or other prohibited effects.

B. Essential-Facilities Principles

A literature corpus could theoretically become relevant where access to particular information resources is indispensable for competition.

However, European law imposes a demanding threshold for compulsory access.

The classic jurisprudence therefore provides an important limitation: mere usefulness or commercial importance is not enough.

C. Self-Preferencing

If an AI platform operates both:

• a literature-indexing/synthesis service, and
• competing research-content or publishing services,

competition concerns could arise if it systematically gives its own material preferential treatment.

This is particularly relevant where the platform controls the primary discovery interface.

D. Tying and Bundling

A dominant research platform might potentially condition access to:

• literature databases,
• AI summarisation,
• citation-management tools,
• institutional search,
• cloud storage,
• research analytics,

on purchasing or adopting another service.

E. Interoperability and Data Access

Competition concerns can also arise if competitors cannot obtain access to:

• citation metadata;
• bibliographic information;
• interoperability interfaces;
• search indexes;
• ranking information;
• user-authorised research histories.

The relevant question is whether the information is genuinely indispensable and whether withholding it forecloses effective competition.

IV. Major Case Laws

The following cases do not concern modern generative-AI literature-synthesis systems directly. They provide legal analogies and principles that can be applied to AI knowledge-filtering systems.

1. Magill TV Guide Cases

Joined Cases C-241/91 P and C-242/91 P, RTE and ITP v Commission

Facts

Television broadcasters controlled copyright-protected information concerning weekly programme schedules. Third parties sought to produce comprehensive television guides using that information.

The broadcasters refused to license the information.

Legal principle

The Court recognised that exercise of an intellectual-property right can, in exceptional circumstances, constitute an abuse of dominant position.

The case identified circumstances involving, among other factors:

• information indispensable for a new product;
• elimination of competition;
• refusal without objective justification;
• potential consumer demand for the new product.

The Court's case concerned programme information, but its conceptual importance extends to control over information resources.

Application to AI literature synthesis

A hypothetical AI literature platform might control a particularly valuable corpus of:

• scientific metadata;
• abstracts;
• citation relationships;
• research classifications;
• full-text material.

If competitors cannot effectively reproduce the service without access to that resource, Magill-type reasoning could become relevant.

Limitation

Ordinary possession of valuable academic information does not automatically create an obligation to license it.

The exceptional conditions remain important.

2. IMS Health v NDC Health

Case C-418/01, IMS Health GmbH & Co. OHG v NDC Health GmbH & Co. KG

Facts

IMS Health developed a particular "brick structure" for organising regional pharmaceutical-sales data.

A competing undertaking required access to the structure to compete effectively in supplying pharmaceutical-sales information.

Holding

The Court identified stringent conditions under which refusal to license intellectual-property-protected information can amount to abuse.

Among the important considerations were whether:

1. access was indispensable;
2. the refusal prevented a new product or service;
3. there was potential consumer demand;
4. the refusal lacked objective justification;
5. the refusal eliminated competition in the relevant downstream market. 

AI relevance

Imagine an AI research platform possessing an indispensable scholarly classification system used to organise an entire scientific literature market.

If competing AI systems genuinely cannot provide comparable services without access to that system, IMS Health could provide an analytical framework.

Key lesson

Data importance ≠ automatic essential facility.

The indispensability threshold remains high.

3. Microsoft v Commission

Case T-201/04, Microsoft Corp. v Commission

Facts

Microsoft controlled the dominant PC operating-system platform.

The European Commission found abuse relating, among other things, to Microsoft's refusal to provide competitors with interoperability information.

The General Court largely upheld the Commission's decision.

AI literature relevance

This case is particularly important for AI research ecosystems because literature platforms may depend on interoperability with:

• university repositories;
• citation-management systems;
• research databases;
• institutional authentication systems;
• bibliographic APIs;
• academic search services.

A dominant AI research platform could theoretically make interoperability unnecessarily difficult for competing systems.

Example

Suppose Platform A dominates AI-assisted scholarly search and controls the principal research index.

If Platform A deliberately prevents competing synthesis tools from interoperating with essential research infrastructure while using that infrastructure itself, interoperability foreclosure could become relevant.

Principle

Control over a technological interface can become a competition issue where the refusal or restriction substantially prevents effective competition.

4. Bronner v Mediaprint

Case C-7/97, Oscar Bronner GmbH & Co. KG v Mediaprint

Facts

A newspaper group controlled a nationwide newspaper home-delivery system.

A competing newspaper sought access to the distribution network.

Holding

The Court applied a strict test for compulsory access to infrastructure controlled by a dominant undertaking.

The facility needed to be genuinely indispensable rather than merely more convenient or economically advantageous.

AI application

This provides an important counterweight to arguments that every important AI corpus should be shared.

For example, a scholarly AI system might argue:

"Our enormous research corpus is essential for competitors."

That assertion alone would not establish an essential-facility obligation.

A competition authority would need to examine:

• alternative databases;
• public repositories;
• publisher APIs;
• institutional repositories;
• data portability;
• technical feasibility of duplication;
• economic feasibility;
• whether access is objectively indispensable.

Principle

A highly valuable database is not automatically an essential facility.

5. Google Android

Case T-604/18, Google and Alphabet v Commission

Facts

The case concerned Google's Android ecosystem and contractual arrangements involving:

• Android;
• Google Play;
• Google Search;
• Chrome;
• device manufacturers;
• mobile network operators.

The General Court largely upheld the Commission's findings concerning restrictions used to strengthen Google's search position.

AI literature-synthesis relevance

This case illustrates ecosystem leveraging.

An AI company could theoretically operate:

operating system → browser → search → scholarly discovery → AI synthesis → research assistant.

If dominance in one layer makes it difficult for competing AI research services to reach users, competition authorities could examine whether ecosystem control is being used to reinforce another market.

Potential concern

A dominant search provider could theoretically give its own literature-synthesis system:

• preferred placement;
• default status;
• privileged access to data;
• exclusive distribution;
• technical advantages unavailable to rivals.

That would raise questions similar to the ecosystem concerns examined in Google Android.

6. Google Search / Search Monopoly Litigation

United States v Google LLC

The U.S. litigation concerning Google's general search services provides an especially important modern analogy.

The court found Google liable under Section 2 of the Sherman Act for maintaining monopolies in general search and general search text advertising through exclusionary distribution practices.

The 2025 remedies included restrictions on exclusive distribution arrangements and requirements concerning certain search-index and user-interaction data and search syndication.

AI literature relevance

The analogy is powerful because an AI literature-synthesis system can become an information-distribution interface.

The competitive concern is not merely:

"Who owns the best AI model?"

It can instead become:

"Who controls the gateway through which researchers discover knowledge?"

If a dominant platform controls the gateway, exclusionary distribution agreements could prevent competing synthesis systems from obtaining sufficient scale.

AI-specific significance

The U.S. proceedings also expressly recognised that generative AI is changing search and considered how AI could interact with existing search-market power. The DOJ litigation materials noted that AI does not eliminate the importance of crawling, indexing and ranking as foundational search functions.

7. Google Shopping

Google Search (Shopping) is another important analogy for AI knowledge filtering.

The European Commission's Google Shopping case concerned preferential treatment of Google's own comparison-shopping service within general search.

AI relevance

The analogous AI scenario would be:

General literature search → AI synthesis interface → proprietary research product

If a dominant platform systematically privileges its own research database, publisher network, citation product or AI-generated literature summaries over competing services, the legal analysis could involve:

• self-preferencing;
• discrimination;
• leveraging;
• foreclosure;
• ranking manipulation.

The critical distinction is that not every ranking decision is anticompetitive. Search quality, relevance, reliability and safety may constitute legitimate product-design objectives.

The competition question is whether ranking is being used as a mechanism for exclusion.

V. Emerging UK Approach: AI Search and Publisher Content

The UK's current digital-markets framework is especially relevant to this topic.

The CMA designated Google as having strategic market status in general search and search advertising in 2025. Its scope includes AI-based search features such as AI Overviews and AI Mode, while Gemini as a standalone assistant was initially outside the designation.

In June 2026, the CMA imposed a fair-ranking conduct requirement requiring organic search results to be ranked according to objective and non-discriminatory criteria, expressly including generative-AI search features.

It also imposed a publisher conduct requirement concerning publishers' control over how their content is used in AI features.

This is highly relevant to AI literature synthesis because scholarly publishers face a similar structural question:

Who controls the terms on which copyrighted or commercially valuable research content becomes input for an AI synthesis layer?

VI. Major Competition Concerns

1. Source-selection discrimination

An AI system may systematically prefer:

• affiliated publishers;
• affiliated databases;
• commercially integrated journals;
• sources generating greater platform revenue.

This can produce algorithmic self-preferencing.

2. Citation suppression

Suppose two studies reach similar conclusions but the system consistently cites one publisher and omits the other.

Over time, this can affect:

• academic visibility;
• citation counts;
• research funding;
• institutional reputation;
• downstream discovery.

The competitive effect can therefore occur without an explicit exclusionary contract.

3. Publisher foreclosure

If an AI platform becomes the principal research-discovery interface, publishers may become dependent upon it for traffic.

The platform could potentially acquire substantial bargaining power over:

• licensing;
• attribution;
• access;
• indexing;
• payment;
• crawling.

4. Self-preferencing

A platform operating:

• a research database,
• an AI synthesis engine,
• a citation manager,

could theoretically place its own database or tools ahead of competing products.

This creates a potential Google Shopping-type theory.

VII. Knowledge Filtering as a New Competitive Bottleneck

Traditional search involves:

Web → Index → Ranking → User

AI synthesis adds another layer:

Research universe → Retrieval → Filtering → Weighting → Synthesis → Citation → User belief

That additional layer is economically significant.

The system may no longer merely determine:

"What information can the user find?"

It may increasingly determine:

"What information is considered relevant enough to become part of the user's answer?"

This produces a new form of intermediation power.

VIII. Algorithmic Neutrality and Competition

A central regulatory problem is distinguishing legitimate AI optimisation from anticompetitive discrimination.

Legitimate optimisation may include:

• relevance;
• methodological quality;
• recency;
• citation reliability;
• peer-review status;
• duplicate removal;
• fraud detection;
• source credibility;
• user-selected preferences.

Potentially problematic conduct could include:

• deliberately suppressing competitors;
• preferential treatment of affiliated sources;
• discriminatory API access;
• exclusionary licensing arrangements;
• manipulation of rankings to foreclose rivals;
• preventing competing systems from accessing indispensable data;
• tying literature access to another product;
• using monopoly profits to lock up distribution.

The distinction is fact-specific.

IX. Data as a Competitive Asset

AI literature platforms may possess several layers of data:

The competitive advantage may therefore arise not simply from model parameters but from continuous accumulation of proprietary information and feedback.

X. Network Effects

AI literature platforms may benefit from several reinforcing loops:

More users
↓
More queries and feedback
↓
Better retrieval and ranking
↓
Better synthesis
↓
More users

At the same time:

More publishers indexed
↓
More comprehensive coverage
↓
More researchers rely on platform
↓
More publishers feel compelled to participate

This can create two-sided or multi-sided network effects.

XI. Switching Costs and Researcher Lock-In

Researchers may accumulate:

• saved searches;
• citation libraries;
• personal profiles;
• annotations;
• institutional integrations;
• research histories;
• customised prompts;
• knowledge graphs.

These can create switching costs.

A competing AI literature system may therefore face difficulty even if its underlying model is technically competitive.

Competition analysis should consequently examine portability, not merely price.

XII. Hallucination and Competition Law

Hallucinations are ordinarily an accuracy/product-quality issue, not automatically an antitrust violation.

However, they can acquire competition significance where a dominant platform:

1. systematically misrepresents rival research;
2. suppresses competing sources;
3. falsely characterises competitors' products;
4. ranks its own material using demonstrably discriminatory criteria;
5. prevents users from accessing alternative sources.

The important distinction is between:

ordinary AI error
and
strategically induced information distortion.

XIII. Remedies

Potential competition remedies could include:

1. Non-discriminatory ranking

Require objective criteria for source ranking.

2. Data portability

Allow users to transfer:

• search histories;
• citation libraries;
• annotations;
• research profiles.

3. API access

Provide competitors with appropriate technical access where legally justified.

4. Publisher choice

Allow publishers to control or negotiate the use of their content in AI synthesis.

5. Transparency

Require meaningful information concerning:

• source selection;
• ranking;
• citation;
• major algorithmic changes.

6. Interoperability

Prevent technical restrictions designed to exclude competing research tools.

7. Anti-self-preferencing obligations

Where legally justified, prevent systematic preference for affiliated research services.

XIV. Key Doctrinal Synthesis

XV. Conclusion

AI literature synthesis systems can transform information control into a potential source of competitive power. Their significance lies not merely in generating summaries but in controlling the intermediary layer between an enormous research universe and the final knowledge presented to users.

The most relevant established competition-law principles come from cases such as Magill, IMS Health, Bronner, Microsoft, Google Android, and Google Search. They collectively show that competition law can address circumstances where a powerful undertaking controls an important information, technological, distribution, or interoperability bottleneck—but they also establish substantial limits before access obligations or intervention are justified.

The emerging UK approach is particularly significant because its 2026 search conduct requirements expressly address fair ranking in generative-AI search features and publisher control over use of content in AI features.

Accordingly, the future competition-law question may evolve from "Who controls the database?" to "Who controls the algorithmic gateway through which knowledge becomes visible, comparable and synthesised?"