Alignment Service Provider Dominance and Dependency Risks

1. Introduction

Alignment service providers are entities that supply technological or professional services designed to make artificial-intelligence systems operate according to specified objectives, safety rules, human preferences, organisational policies, or regulatory requirements. The term can cover RLHF systems, reward-model services, constitutional-AI tooling, safety-evaluation platforms, red-teaming services, model monitoring, interpretability tools, alignment datasets, guardrail APIs, AI-audit services, and automated compliance systems.

From a competition-law perspective, the important issue arises where a small number of providers become indispensable to AI developers because they control specialised alignment data, evaluation benchmarks, safety infrastructure, proprietary APIs, technical interfaces, expert talent, or certification systems. Recent regulatory developments concerning AI and cloud markets already identify switching costs, technical barriers, access to inputs and interoperability as significant competition issues.

The central question is therefore:

When does dependence upon an alignment service provider become a competition-law problem rather than an ordinary commercial dependency?

The answer generally depends upon market power, indispensability, foreclosure, discriminatory access, tying, exclusivity, switching costs, interoperability, self-preferencing, and the effect on downstream AI competition.

2. Meaning of Alignment Service Provider Dominance

An alignment service provider may acquire market power where it controls a particularly important input into AI development.

For example:

AI Developer → Alignment Service → Safety certification → Deployment

If competing AI developers cannot realistically deploy their systems without the alignment service, the provider may occupy an important bottleneck position.

Dominance can arise from:

1. proprietary alignment datasets;
2. specialised human-feedback networks;
3. exclusive access to expert evaluators;
4. proprietary safety benchmarks;
5. model-monitoring infrastructure;
6. proprietary guardrail APIs;
7. technical interoperability standards;
8. regulatory or industry certifications;
9. accumulated evaluation data;
10. network effects;
11. high switching costs; and
12. integration with cloud, operating-system or AI-model ecosystems.

Dominance does not, by itself, constitute an infringement. Competition law normally becomes concerned when market power is used to exclude competitors, exploit dependent customers, restrict access, or extend power into adjacent markets.

3. Alignment Services as a Bottleneck Input

An alignment service can function as a bottleneck when the service is technically difficult to reproduce.

Consider a provider that possesses:

• millions of specialised human preference evaluations;
• proprietary safety-testing datasets;
• accumulated model-failure information;
• specialised red-team expertise;
• benchmark results across thousands of models;
• regulatory certifications; and
• infrastructure integrated directly into AI-development pipelines.

A new competitor might technically be able to establish another alignment service, but the economic question is whether doing so is commercially feasible within a reasonable period.

This is closely related to the competition-law concept of an essential facility or indispensable input.

The classic European cases include Bronner, IMS Health, and Microsoft, although the precise legal tests differ according to the nature of the facility and conduct.

4. Dependency Risks

A. Technical dependency

AI developers may build their systems around one provider's:

• API;
• SDK;
• evaluation format;
• safety taxonomy;
• monitoring architecture; or
• model-development workflow.

Once integration is complete, changing providers can require substantial redevelopment.

This produces technical switching costs.

B. Data dependency

An alignment provider may accumulate an enormous proprietary dataset consisting of:

• human preference data;
• safety incidents;
• red-team results;
• model evaluations;
• behavioural testing;
• prompt-response pairs; and
• model failure patterns.

The more data the provider accumulates, the greater the possibility of a data feedback loop:

more customers → more evaluation data → better alignment service → more customers → still more data.

This can produce durable entry barriers.

C. Certification dependency

Suppose regulators, insurers or major enterprises begin relying upon one company's AI-safety certification.

The certification provider could then become a gatekeeper between AI developers and customers.

A provider that also operates a competing AI system could potentially have incentives to make rival AI developers subject to more burdensome evaluation requirements.

That creates a possible vertical foreclosure problem.

5. Refusal to Provide Access

A dominant alignment provider could potentially restrict competitors by refusing access to:

• evaluation infrastructure;
• safety benchmarks;
• APIs;
• testing environments;
• interoperability information;
• certification systems;
• specialised datasets; or
• technical interfaces.

The traditional European approach requires a very high threshold before a dominant company is required to deal with competitors.

Bronner v Mediaprint

In Oscar Bronner GmbH & Co. KG v Mediaprint, the Court of Justice established stringent conditions for treating an infrastructure or facility as indispensable.

The relevant considerations include whether:

1. access is indispensable;
2. refusal would eliminate effective competition;
3. duplication is practically or economically impossible; and
4. there is no objective justification.

The case is particularly relevant to alignment services because an AI developer should ordinarily not be entitled to demand access merely because another provider is more expensive or less convenient.

The essential-facility doctrine is exceptional.

6. Six Major Case Laws

Case 1 — Commercial Solvents Corp. v Commission

Court: Court of Justice of the European Union
Competition principle: Refusal to supply an essential input

Commercial Solvents concerned a dominant supplier that controlled an important upstream input and restricted supply to a downstream competitor.

The case established an important principle: a dominant undertaking controlling an indispensable upstream input cannot necessarily use that position to eliminate downstream competition.

Relevance to alignment services

Suppose an alignment provider controls a critical safety-evaluation infrastructure and simultaneously operates an AI-development business.

If it selectively refuses alignment services to competing AI developers while supplying its own AI business, the conduct could raise concerns similar to the upstream-input/downstream-competition problem addressed in Commercial Solvents.

The key issue would be whether the alignment service is sufficiently important and whether the refusal produces exclusionary effects.

Case 2 — Bronner v Mediaprint

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

This is one of the most important authorities for essential facilities.

The Court rejected an overly broad obligation to provide competitors with access to a dominant firm's infrastructure.

The facility had to be genuinely indispensable; the existence of alternative methods of reaching customers was important.

Application

An alignment provider should not automatically be required to share:

• proprietary evaluation tools;
• proprietary safety models;
• specialised datasets; or
• expensive infrastructure.

The claimant would need to establish genuine indispensability, rather than simply showing that the provider's service is better, cheaper, faster or more popular.

Case 3 — IMS Health v NDC Health

Case: IMS Health GmbH & Co KG v NDC Health GmbH & Co KG
Case C-418/01

IMS Health concerned access to a copyrighted pharmaceutical-sales database structure.

The Court identified exceptional circumstances under which refusal to license intellectual property could constitute abuse.

Important considerations included:

1. indispensability;
2. elimination of effective competition;
3. prevention of a new product for which there is consumer demand; and
4. absence of objective justification.

Application to alignment data

Alignment datasets may possess intellectual-property protection or trade-secret characteristics.

Therefore, a competitor cannot ordinarily demand access merely because a dataset is commercially valuable.

However, where a proprietary alignment resource becomes indispensable and refusal prevents meaningful competition or innovative downstream products, IMS Health provides an important analytical framework.

Case 4 — Microsoft Corp. v Commission

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

Microsoft involved interoperability information and Microsoft's dominant position in operating systems.

The European Commission and General Court were concerned with Microsoft's refusal to provide interoperability information necessary for competing server products.

The case demonstrates that interoperability can itself become a competition-law issue.

The Court upheld the Commission's finding concerning Microsoft's refusal to provide interoperability information.

Application to AI alignment

Imagine that an alignment service provider develops a proprietary API through which AI systems must communicate with:

• safety-monitoring systems;
• evaluation platforms;
• deployment infrastructure;
• cloud services; or
• regulatory compliance systems.

If the provider controls the interface and selectively degrades access for competing AI developers, Microsoft provides an important analogy.

The issue would not merely be "access to software"; it would concern whether control over interoperability creates downstream exclusion.

Case 5 — Slovak Telekom v Commission

Case: Slovak Telekom a.s. v European Commission
Cases C-165/19 P and C-166/19 P

The case concerned access to telecommunications infrastructure and exclusionary conduct.

It is significant because the Court clarified the relationship between regulated access obligations and the strict Bronner conditions.

Where access is already governed by regulatory obligations, the exceptionally strict Bronner test does not necessarily operate in the same way.

Application to AI

This distinction is increasingly important as AI services become subject to:

• interoperability requirements;
• data-access obligations;
• cybersecurity rules;
• AI auditing;
• sector-specific regulation; and
• digital-platform regulation.

Where legislation specifically requires interoperability or access, the provider may have less freedom to invoke ordinary property or contractual rights as a justification for exclusion.

Case 6 — Alphabet and Others (Android Auto)

Case: Alphabet and Others v Autorità Garante della Concorrenza e del Mercato
Case C-233/23

This is particularly important for modern AI ecosystems.

In Android Auto, the CJEU addressed interoperability between a dominant digital platform and third-party applications.

The Court held that the strict Bronner indispensability requirement did not automatically apply where the platform was designed to accommodate third-party applications. A refusal to provide interoperability could potentially constitute abuse even though the technology was not indispensable in the traditional Bronner sense.

Importance for alignment providers

This is highly relevant to an alignment platform intentionally designed as an ecosystem for third-party AI developers.

For example:

Alignment Platform → Third-party AI models → Safety testing → Deployment

If the platform is expressly designed to accept third-party AI systems, a provider may face greater scrutiny if it selectively denies interoperability to rival AI developers while maintaining equivalent functionality for its own AI products.

7. Additional Important Authority — Magill

RTE and ITP v Commission — Magill

Cases: C-241/91 P and C-242/91 P

Magill concerned refusal to license copyright-protected television programme information.

It helped develop the exceptional-circumstances framework for compulsory access to intellectual property.

Relevance

An alignment provider may claim that:

• its evaluation methodology is proprietary;
• its safety dataset is copyrighted;
• its benchmarks constitute trade secrets; or
• its alignment technology is protected by patents.

Competition law does not automatically override those rights.

However, IP protection cannot necessarily be used as an absolute shield against exclusionary abuse when the exceptional conditions identified in the case law are satisfied.

8. Self-Preferencing Risks

A particularly important problem arises when an alignment provider operates at multiple levels.

For example:

Alignment Service Provider

↓
provides alignment infrastructure

Competing AI developers

but also

↓
develops its own AI model.

The provider could potentially:

• give its own AI models faster evaluation;
• provide competitors with slower API access;
• give its own models preferential safety scores;
• apply stricter evaluation thresholds to competitors;
• reserve the best evaluators for internal products;
• prioritise internal computing resources; or
• use customer alignment data to improve its own competing products.

This creates a vertical self-preferencing problem.

9. Tying and Bundling

An alignment provider may possess dominance in one market and condition access to another service.

For example:

"Access to our alignment platform is available only if you purchase our cloud-computing service."

Or:

"AI safety certification is available only to customers using our proprietary model-monitoring platform."

This can create concerns involving tying and bundling.

The competitive theory is:

Dominant service A → conditions access → complementary service B → competitors weakened in B.

The stronger the market position of A and the weaker the availability of substitutes, the greater the potential concern.

10. Exclusive Dealing

Alignment providers could potentially enter agreements requiring AI developers to use only their:

• evaluation platform;
• red-team service;
• safety dataset;
• model-monitoring software; or
• certification system.

Exclusivity becomes particularly significant when the provider has substantial market coverage.

For example:

"Customers receiving our highest-level safety certification may not use competing alignment providers."

If many important AI developers are bound by such agreements, competitors may be unable to obtain sufficient scale to enter the market.

11. Switching Costs

One of the most important dependency mechanisms is lock-in.

A developer might initially use Provider A because it is inexpensive.

Over time, the developer builds:

• proprietary workflows;
• datasets;
• monitoring systems;
• APIs;
• evaluation histories;
• compliance documentation; and
• internal expertise

around Provider A.

Switching to Provider B may therefore require:

migration + retraining + revalidation + recertification + technical redevelopment.

The provider may then increase prices after customers become dependent.

Current EU digital regulation expressly recognises interoperability and switching barriers as competition concerns in cloud markets. The Data Act, for example, addresses switching between data-processing providers and requires measures facilitating interoperability and portability.

12. Data Lock-In

Alignment services create a particularly interesting form of lock-in because the customer may continuously generate valuable data while using the service.

For example:

Customer uses alignment platform:

10,000 evaluations → 100,000 evaluations → 1 million evaluations

The provider accumulates:

• model behaviour data;
• failure patterns;
• preference signals;
• adversarial prompts;
• safety vulnerabilities;
• benchmark performance.

This can create a data-based competitive advantage.

The provider may become better precisely because customers are dependent upon it.

13. Information Asymmetry

Alignment providers often possess more information about model behaviour than their customers.

This creates an unusual competition problem.

The provider may know:

• which models perform poorly;
• which customers use competing providers;
• which safety techniques work;
• which prompts expose vulnerabilities;
• which models are likely to enter particular markets.

If the provider also operates competing AI products, customer information could potentially provide it with a strategic advantage.

The FTC's investigation into major AI/cloud partnerships has specifically identified access to sensitive technical and business information as a potential competition issue.

14. AI Alignment Services and Cloud Dependency

Alignment services increasingly depend upon cloud infrastructure.

This creates a potentially layered dependency:

Cloud provider
↓
AI model infrastructure
↓
Alignment service
↓
AI developer
↓
End users

If one company controls several layers, it may have opportunities to leverage market power vertically.

The FTC's study of major AI/cloud partnerships identified possible concerns involving:

• access to computing resources;
• switching costs;
• technical barriers to migration;
• access to sensitive information; and
• exclusivity or control rights. 

The EU has likewise investigated cloud competition, including interoperability, data access, tying/bundling and contractual conditions.

15. Current Regulatory Development

The importance of interoperability is becoming especially clear in AI markets.

In July 2026, the European Commission adopted binding measures under the DMA concerning Google's Android interoperability obligations, specifically addressing access for competing AI services. The measures seek to ensure that rival AI services can access relevant Android capabilities on terms that allow them to compete with Google's own AI services.

This illustrates a broader movement from purely ex-post antitrust enforcement toward ex-ante interoperability obligations for strategically important digital ecosystems.

16. Competition Theories Applicable to Alignment Service Providers

17. Essential Facility Analysis

A useful analytical framework is:

Step 1 — Define the market

Possible markets include:

• AI alignment services;
• AI safety evaluation;
• AI red-teaming;
• model-monitoring services;
• alignment datasets;
• AI certification;
• AI compliance tooling.

Step 2 — Establish market power

Consider:

• market share;
• entry barriers;
• network effects;
• switching costs;
• data advantages;
• technical expertise;
• customer dependency.

Step 3 — Identify the bottleneck

Ask whether the provider controls an input that rivals cannot reasonably reproduce.

Step 4 — Examine alternatives

Can customers realistically switch to:

• another evaluator;
• an open-source system;
• an internal alignment team;
• another dataset;
• another certification provider?

Step 5 — Examine conduct

Was there:

• refusal;
• discrimination;
• degradation;
• tying;
• exclusivity;
• self-preferencing?

Step 6 — Examine effects

Does the conduct:

• raise rivals' costs;
• prevent entry;
• reduce innovation;
• reduce consumer choice;
• increase prices;
• reduce quality?

Step 7 — Consider justification

Possible legitimate justifications include:

• cybersecurity;
• privacy;
• intellectual-property protection;
• safety;
• technical limitations;
• protection against model abuse.

The key issue is whether the justification is objective, proportionate and genuinely connected to the restriction.

18. Important Distinction: Dependency Is Not Automatically Abuse

This distinction is fundamental.

A customer can become highly dependent upon a supplier without competition law being violated.

For example:

Provider A has the best alignment technology → customers voluntarily choose it → competitors remain available.

That alone does not establish unlawful dominance.

The competition concern becomes stronger where:

Provider A becomes dominant → customers become locked in → Provider A exploits the dependency to exclude rivals or impose discriminatory conditions.

Therefore:

Commercial dependency ≠ automatically anticompetitive conduct.

19. Potential Remedies

Competition authorities could theoretically consider remedies such as:

Structural remedies

• divestiture;
• separation of alignment and competing AI businesses.

Behavioural remedies

• non-discriminatory access;
• interoperability;
• API access;
• data portability;
• prohibition of exclusivity;
• transparent evaluation criteria.

Technical remedies

• open interfaces;
• export formats;
• portability of evaluation histories;
• interoperable safety standards.

Data remedies

• restrictions on use of customer data;
• data portability;
• separation of customer information from competing AI operations.

Contractual remedies

• reasonable termination rights;
• restrictions on excessive switching fees;
• limits on long-term exclusivity.

20. Six-Case Synthesis

The Android Auto decision is particularly significant for modern digital ecosystems because the CJEU recognised that where infrastructure is designed for third-party use, refusal of interoperability may be problematic even without satisfying the traditional indispensability requirement.

21. Conclusion

Alignment service provider dominance and dependency risks arise when control over AI-safety infrastructure, evaluation systems, alignment data, expert networks, APIs or certification becomes sufficiently concentrated that AI developers cannot effectively compete without access to that ecosystem.

The principal competition-law risks are:

1. essential-facility dependence;
2. refusal to supply;
3. discriminatory access;
4. interoperability restrictions;
5. self-preferencing;
6. tying and bundling;
7. exclusive dealing;
8. excessive switching costs;
9. data lock-in;
10. leveraging into downstream AI markets;
11. exploitation of confidential customer information; and
12. acquisition of emerging alignment competitors.

The central legal lesson from Commercial Solvents, Bronner, Magill, IMS Health, Microsoft, Slovak Telekom and Android Auto is that competition law must balance two competing considerations: the right of an innovating firm to control and monetise infrastructure that it has developed, and the need to prevent control of an indispensable or strategically important bottleneck from being used to foreclose effective competition.