Evidentiary Challenges in Predicting AI Market Evolution

1. Introduction

Evidentiary challenges in predicting AI market evolution arise because competition authorities and courts are increasingly required to determine how artificial-intelligence markets are likely to develop before the relevant competitive changes have actually occurred.

This problem is particularly important in:

• AI foundation models;
• generative AI;
• cloud and compute infrastructure;
• GPU/accelerator markets;
• AI inference APIs;
• AI agents;
• data-access ecosystems;
• AI-enabled search and advertising;
• AI safety and evaluation services;
• model licensing;
• AI application marketplaces; and
• vertically integrated AI ecosystems.

Traditional competition evidence often relies on historical prices, market shares, customer behaviour, entry records and observed competitive effects. AI markets may provide very little reliable historical evidence. A model may become obsolete within months; a new architecture may radically reduce compute requirements; an open-source model may suddenly constrain a previously powerful proprietary model; or an incumbent may integrate a previously separate AI function into its ecosystem.

Consequently, the evidentiary question becomes:

How can a competition authority establish, to the required legal standard, what an AI market is likely to look like several years into the future?

The UK CMA's current merger guidance expressly recognises this difficulty. It states that merger assessment is forward-looking, that some future developments involve substantial uncertainty, and that rapidly developing technological markets may provide less historical evidence. It also recognises that authorities need not make precise predictions about exactly which innovation, price movement or service change will occur.

2. Why AI Market Evolution Creates an Evidentiary Problem

A. The evidence concerns an inherently uncertain future

A conventional competition investigation can ask:

• What was the firm's market share?
• What prices did it charge?
• How did customers respond?
• What happened when a competitor entered?

AI competition frequently requires different questions:

• Will the model improve substantially?
• Will compute costs fall?
• Will open-source models become commercially viable?
• Will customers switch from standalone models to integrated AI ecosystems?
• Will an AI application become a platform?
• Will a new architecture displace the incumbent?
• Will AI agents become substitutes for conventional software?
• Will vertical integration produce foreclosure?

These propositions are inherently predictive.

The evidentiary challenge is therefore not simply insufficient evidence. It is the distinction between evidence of present facts and evidence capable of supporting a reasonable inference about future competitive conditions.

3. Historical Evidence May Be a Poor Predictor

In conventional markets, past competitive behaviour can provide a reasonable indication of future behaviour.

AI markets may break that assumption.

For example, historical evidence showing that:

Firm A had 70% of the market for AI models in 2025

does not necessarily establish that Firm A will have comparable market power in 2028.

Between those dates there could be:

• a new open-weight model;
• a major reduction in inference costs;
• a new chip architecture;
• a breakthrough in model efficiency;
• new interoperability requirements;
• technological convergence;
• customer migration;
• regulatory intervention; or
• entry by a cloud provider.

The CMA consequently recognises that recent pre-merger evidence may be less predictive in nascent and dynamic markets, where internal documents, competitor plans, product characteristics and expected numbers of competitors may receive greater weight.

4. The Counterfactual Problem

One of the most significant evidentiary challenges is construction of the counterfactual.

Competition authorities must often compare:

AI market with the challenged conduct/merger

against

AI market without the challenged conduct/merger.

But determining what the AI market would have looked like without the transaction can be exceptionally difficult.

Suppose an established cloud provider acquires a promising AI-model developer.

The authority might argue:

Without the acquisition, the AI developer would have independently become a significant competitor.

The merging parties might respond:

Without the acquisition, the developer would have failed to obtain sufficient compute, capital or distribution and would have remained commercially insignificant.

Both propositions concern an unrealised future.

The evidence could include:

• funding documents;
• internal strategic plans;
• technical roadmaps;
• compute procurement agreements;
• customer contracts;
• recruitment plans;
• model-development milestones;
• investor presentations;
• competitor reactions;
• developer adoption;
• historical entry into adjacent markets.

But none necessarily proves the future outcome with certainty.

The CMA itself recognises that establishing the appropriate counterfactual is inherently uncertain and that evidence concerning future developments absent a merger may be difficult to obtain.

5. Case Law: Commission v Tetra Laval

Commission v Tetra Laval, Case C-12/03 P

This is one of the most important authorities for predictive competition evidence.

The European Commission had prohibited the Tetra Laval/Sidel merger partly on the basis of predicted future conduct and conglomerate effects.

The European courts considered the evidence insufficiently convincing. The Court of Justice emphasised the need for sufficiently robust evidence when a competition authority relies upon predictions of future conduct. The underlying General Court judgment had found that the Commission's evidence concerning future pricing and innovation incentives did not establish the predicted effects to the requisite legal standard.

Importance for AI

The principle is highly relevant to AI because authorities may be tempted to construct elaborate chains such as:

acquisition → increased compute access → improved model → ecosystem integration → customer foreclosure → reduced innovation → higher future market power.

Each additional inferential step increases evidentiary vulnerability.

The lesson from Tetra Laval is not that future effects cannot be established. Rather, the authority must demonstrate that the predictive chain rests upon reliable, consistent and sufficiently robust evidence.

6. Case Law: Airtours v Commission

Airtours v Commission, Case T-342/99

The Commission prohibited the Airtours/First Choice merger on the basis of anticipated coordinated effects.

The General Court annulled the decision, finding that the Commission had failed adequately to establish the necessary conditions for the predicted competitive harm.

Relevance to AI

AI markets may generate similar predictive problems where authorities argue that several major AI firms will become mutually dependent.

For example:

Similar models + common cloud infrastructure + algorithmic pricing + transparent API information = increased probability of coordinated behaviour.

That proposition cannot simply be assumed.

The authority must establish the relevant causal conditions.

In AI markets, evidence may need to demonstrate:

• monitoring capability;
• transparency;
• predictability of rivals' conduct;
• ability to punish deviation;
• incentives to coordinate;
• stability of the market structure;
• technological similarity; and
• absence of disruptive innovation.

Airtours therefore illustrates the danger of moving from an economically plausible theory to a legally sufficient evidentiary conclusion.

7. Case Law: Ryanair v Commission / Ryanair–Aer Lingus

Ryanair Holdings plc v Commission, Case T-342/07

The Ryanair/Aer Lingus litigation illustrates the difficulty of predicting competitive relationships using market shares and other evidence when the competitive relationship between firms is evolving.

The General Court considered arguments concerning:

• market shares;
• closeness of competition;
• econometric evidence;
• customer surveys;
• competitive constraints;
• non-technical evidence; and
• the significance of the parties' positions in the relevant market. 

Relevance to AI

AI markets frequently involve multi-dimensional competitive relationships.

A company might simultaneously compete with another firm in:

• foundation models;
• cloud computing;
• AI chips;
• enterprise software;
• advertising;
• search;
• cybersecurity; and
• AI applications.

Consequently, today's market share may not adequately demonstrate tomorrow's competitive constraint.

Evidence concerning closeness of competition becomes especially important.

8. Case Law: CK Telecoms UK Investments v CMA

CK Telecoms UK Investments Ltd v CMA

The UK litigation concerning the proposed Vodafone/Three transaction provides an important modern example of the difficulties surrounding evidence of future competitive conditions.

The litigation generated substantial discussion concerning the evidentiary basis for prospective theories of harm and the degree of confidence required before a competition authority can conclude that a merger is likely to produce an SLC.

The case is particularly important because UK merger control is inherently prospective: the authority must assess what competition is likely to look like after the transaction.

AI relevance

AI merger cases may involve even greater uncertainty than telecommunications because technological change can be much faster.

An AI authority may therefore need to distinguish between:

Evidence of probability

and

speculation about technological possibility.

For example:

"Model X could theoretically become the leading foundation model"

is substantially weaker than:

"Model X has secured the compute, talent, financing, customers and distribution necessary to make substantial expansion highly likely."

The latter provides a considerably stronger evidentiary foundation.

9. Case Law: Bertelsmann and Sony v Impala

Bertelsmann AG and Sony Corporation of America v Impala, Case C-413/06 P

This case concerned the Sony/BMG merger and the assessment of coordinated effects.

The Court of Justice considered the evidentiary requirements for establishing future competitive effects and the need for the Commission's analysis to be sufficiently supported.

AI relevance

AI markets may involve claims that algorithmic transparency, common data sources, similar model architectures and repeated interactions could facilitate coordination.

But similarities alone are not necessarily sufficient.

An authority would ideally need evidence concerning:

1. market transparency;
2. ability to observe rivals;
3. ability to detect deviations;
4. incentives to coordinate;
5. ability to retaliate;
6. stability of competitive conditions; and
7. likelihood that technological disruption will destabilise coordination.

Thus, Impala reinforces the distinction between an economically plausible scenario and an adequately substantiated prediction.

10. Case Law: Tetra Laval's Evidentiary Principle and AI Innovation

A particularly important feature of Tetra Laval is its treatment of predicted innovation effects.

The General Court examined the Commission's evidence regarding the possibility that the merger would reduce incentives to innovate and concluded that the relevant prediction had not been established to the required standard.

This is directly relevant to AI.

An authority might argue:

Acquisition of an AI start-up by a dominant platform will reduce innovation.

But what exactly must be demonstrated?

Potential evidence includes:

• R&D expenditure;
• number of researchers;
• model-release frequency;
• technical benchmarks;
• patent activity;
• open-source contributions;
• model-training capacity;
• investment plans;
• customer switching;
• developer adoption;
• product roadmaps; and
• internal assessments of competitive threats.

The further the authority moves from observable evidence toward long-term technological prediction, the greater the need for a coherent evidentiary chain.

11. Case Law: Microsoft / Commission

Microsoft Corp v Commission, Case T-201/04

The Microsoft litigation demonstrates the importance of evidence concerning technological interoperability, product integration and exclusionary effects in rapidly evolving technology markets.

The case is useful by analogy because competition assessment had to deal with:

• technological functionality;
• interoperability;
• innovation;
• network effects;
• product integration; and
• potential foreclosure.

AI relevance

Modern AI ecosystems raise comparable questions.

For example:

Should an AI platform be required to provide interoperability with competing AI agents?

Evidence could include:

• API access;
• switching costs;
• technical compatibility;
• model portability;
• data portability;
• developer dependency;
• integration costs;
• technical standards; and
• evidence of actual customer switching.

The challenge is determining whether technical integration is merely an efficiency or whether it is likely to create durable foreclosure.

12. Case Law: United States v Microsoft Corp

United States v Microsoft Corp., 253 F.3d 34 (D.C. Cir. 2001)

The Microsoft case is particularly instructive concerning technological markets in which today's competitive position can be used to influence the development of tomorrow's market.

The case concerned Microsoft's conduct involving operating systems, browsers and distribution channels.

AI relevance

An AI platform may similarly possess control over a critical distribution layer:

Operating system → AI assistant → application distribution

or:

Cloud → compute → foundation model → API → AI applications.

Evidence must therefore address not only current market shares but also strategic control over future market development.

13. The Special Problem of AI Forecasting Models

Competition authorities may use:

• econometric models;
• simulations;
• machine-learning forecasts;
• demand estimation;
• agent-based models;
• event studies;
• counterfactual simulations;
• productivity forecasts; and
• technical benchmarking.

These can be useful but create a new evidentiary question:

Can a model-generated prediction itself constitute sufficiently reliable evidence of future competition?

The answer should generally be: only if the model's assumptions, inputs, methodology and limitations are adequately demonstrated.

A sophisticated model does not automatically produce sophisticated evidence.

14. Explainability and Reproducibility

An AI-generated prediction may be difficult for the opposing party or court to challenge if:

• the model is proprietary;
• the training data are confidential;
• the model is not reproducible;
• the algorithm is probabilistic;
• feature selection is opaque;
• hyperparameters are undisclosed;
• the model changes over time; or
• the prediction depends on assumptions that cannot independently be verified.

This creates a procedural fairness problem.

The authority should ideally be able to explain:

Input → Methodology → Assumptions → Model → Output → Economic inference.

If the chain cannot be reconstructed, the evidentiary weight of the prediction may be reduced.

15. The Problem of Endogeneity

AI market forecasts may also suffer from endogeneity.

Suppose an authority predicts:

Firm A will become dominant because developers increasingly adopt Firm A's AI ecosystem.

But developer adoption may itself depend upon:

• the merger;
• the authority's investigation;
• access to subsidised compute;
• exclusive distribution;
• API pricing;
• existing ecosystem size.

Thus, the predicted competitive outcome may partly be a consequence of the very conduct being investigated.

This creates a difficult counterfactual problem:

What would developer adoption have been without the challenged conduct?

16. Internal Documents as Evidence

Internal documents can be extremely important in AI cases.

Examples include:

• board presentations;
• investment memoranda;
• AI strategy documents;
• product roadmaps;
• acquisition presentations;
• competitor assessments;
• model-development plans;
• compute procurement documents;
• pricing strategies;
• internal risk assessments.

The CMA's guidance specifically recognises the importance of internal documents and explains that their evidentiary weight depends on context, including whether documents were created before the parties contemplated the transaction and whether they are consistent with other evidence.

Evidentiary difficulty

Internal documents can conflict.

One document may say:

"We expect significant competitive entry."

Another may state:

"No credible competitor is likely."

The authority therefore cannot necessarily select the most convenient statement.

It must assess:

• author;
• date;
• purpose;
• audience;
• level of technical knowledge;
• assumptions;
• subsequent developments; and
• consistency with objective evidence.

17. The "Black Box" Problem

AI companies may argue that their technology is inherently difficult to explain.

For example, the authority may ask:

Why is Model A likely to become a competitive constraint on Model B?

The company may respond with benchmark scores.

But benchmark superiority does not necessarily establish commercial competition.

The authority may need to connect:

Technical performance → customer preference → switching → revenue → competitive constraint.

This is a crucial evidentiary bridge.

A model can outperform another model on a benchmark but still fail commercially because of:

• high inference cost;
• poor integration;
• weak developer ecosystem;
• limited availability;
• lack of enterprise support;
• inadequate safety features; or
• insufficient distribution.

18. Open-Source AI Creates Additional Uncertainty

Open-weight and open-source AI create particularly difficult forecasting problems.

An authority might predict that proprietary models will maintain dominance.

Yet an open model could suddenly become competitive.

Conversely, an apparently powerful open model may fail because:

• hosting costs are high;
• commercial licensing is restrictive;
• developers prefer integrated services;
• safety requirements increase costs;
• enterprises demand warranties;
• compute remains concentrated.

Therefore, technical availability is not necessarily equivalent to competitive entry.

Evidence should distinguish:

Potential technological capability

from

commercially effective competitive constraint.

19. Network Effects and Data Feedback Loops

AI markets can create feedback mechanisms:

More users → more data → better model → more users

or:

More developers → more applications → greater platform value → more developers.

But these feedback loops may be difficult to prove empirically.

The authority should investigate:

• whether additional data actually improve the model;
• whether data are unique;
• whether comparable data can be obtained elsewhere;
• whether synthetic data substitute for user data;
• whether model improvement produces customer switching;
• whether network effects are direct or indirect.

Otherwise, a theoretically plausible feedback loop may be mistaken for an established economic effect.

20. Evidence from Technical Experts

AI competition cases may increasingly require expert evidence from:

• computer scientists;
• AI engineers;
• economists;
• data scientists;
• semiconductor specialists;
• cloud architects;
• cybersecurity experts.

But expert disagreement creates another evidentiary problem.

One expert may predict:

Compute efficiency will increase rapidly.

Another may predict:

Frontier-model training will continue to require enormous compute resources.

The court or authority therefore needs a framework for evaluating competing forecasts.

Important factors include:

• methodological transparency;
• historical forecasting accuracy;
• quality of underlying data;
• consistency with observable market behaviour;
• sensitivity analysis;
• alternative scenarios; and
• independence of the expert.

21. Scenario Analysis Rather Than Single-Point Prediction

For AI markets, a more defensible evidentiary methodology may involve scenario analysis.

For example:

Scenario A — Incumbent dominance

• proprietary models improve;
• compute remains scarce;
• switching costs increase.

Scenario B — Competitive fragmentation

• open models become effective substitutes;
• inference costs fall;
• multiple providers expand.

Scenario C — Ecosystem convergence

• cloud providers integrate models;
• model differentiation declines;
• distribution becomes more important.

Scenario D — Technological disruption

• a new architecture reduces training and inference costs dramatically.

The authority can then ask:

Does the alleged competition concern remain substantial across the reasonably plausible scenarios?

This may be more robust than pretending to know a single precise future.

22. Evidence Hierarchy for AI Market Evolution

A useful hierarchy is:

The hierarchy is not absolute. Context determines weight.

23. Standard of Proof and Judicial Review

The central legal principle is that uncertainty does not automatically make future effects incapable of proof.

The question is whether the evidence, considered as a whole, sufficiently supports the predicted competitive outcome.

The CMA's present merger guidance expressly states that uncertainty about future market development does not automatically prevent an SLC finding, while also recognising that uncertainty must be appropriately reflected in the assessment.

This is especially important for AI because requiring certainty would effectively prevent enforcement in rapidly developing markets.

But the opposite extreme is equally problematic:

Technological uncertainty cannot become a licence for speculative enforcement.

24. Key Evidentiary Challenges

The principal challenges can therefore be summarised as follows:

1. Temporal uncertainty

AI markets can change faster than the evidentiary record can develop.

2. Counterfactual uncertainty

It may be unclear what would have happened without the challenged transaction or conduct.

3. Technological uncertainty

The next generation of models may fundamentally alter competitive conditions.

4. Data scarcity

Nascent AI markets may lack sufficiently long historical datasets.

5. Model opacity

AI forecasting models may be difficult to interrogate.

6. Endogeneity

Predicted market developments may themselves depend on the conduct being challenged.

7. Measurement problems

Technical benchmarks may not correspond to actual competitive constraints.

8. Multi-sided markets

Competition may occur simultaneously among models, clouds, applications, developers and users.

9. Network effects

Data and developer feedback loops can be difficult to quantify.

10. Expert disagreement

Reasonable experts may produce radically different technological forecasts.

11. Dynamic entry

A firm that appears non-threatening today may become a major competitor rapidly.

12. Open-source disruption

Open models can suddenly alter the competitive landscape.

25. Application to AI Merger Control

Consider a hypothetical:

A dominant cloud provider proposes to acquire a leading AI start-up.

The authority alleges that the start-up would otherwise become a major independent competitor.

The evidentiary record should ideally establish:

Step 1 — Capability

Does the start-up possess technically viable technology?

↓

Step 2 — Resources

Does it have access to sufficient capital, compute and talent?

↓

Step 3 — Commercialisation

Does it have customers and distribution?

↓

Step 4 — Expansion

Does objective evidence show credible expansion plans?

↓

Step 5 — Competitive constraint

Would that expansion materially constrain the incumbent?

↓

Step 6 — Counterfactual

Would those developments probably occur absent the acquisition?

↓

Step 7 — Competitive effect

Would the acquisition eliminate or substantially weaken that future constraint?

This is much stronger than simply asserting:

"The start-up is innovative and could become a competitor."

26. Application to Algorithmic Conduct

The same evidentiary problem arises where an authority alleges that AI pricing algorithms will facilitate coordination.

Evidence may include:

• algorithmic architecture;
• pricing frequency;
• access to competitor information;
• common data sources;
• algorithmic objectives;
• monitoring capabilities;
• observed pricing patterns;
• communications between firms;
• deviations from competitive benchmarks;
• economic simulations.

However, parallel prices alone may not establish unlawful coordination.

The authority must distinguish:

independent algorithmic adaptation

from

coordination or concerted conduct.

This is particularly important where algorithms respond automatically to market signals.

27. Application to AI Innovation Theory of Harm

An authority might claim:

Acquisition of an AI competitor will reduce future innovation.

The evidentiary burden should not rest merely on:

• current R&D expenditure;
• patent counts;
• number of researchers; or
• benchmark performance.

The authority should examine:

• future product pipelines;
• development milestones;
• customer demand;
• investment commitments;
• technical scalability;
• competing technologies;
• probability of successful commercialisation;
• expected substitution;
• likely future entry.

This is where Tetra Laval becomes especially instructive.

28. A Balanced Evidentiary Framework

A sound framework for AI competition enforcement should combine:

A. Observable evidence

Actual:

• customers;
• contracts;
• investment;
• deployment;
• switching;
• technical performance.

B. Documentary evidence

• internal documents;
• board papers;
• strategic plans;
• technical roadmaps.

C. Economic evidence

• demand estimation;
• entry analysis;
• switching costs;
• network effects;
• counterfactual modelling.

D. Technical evidence

• model architecture;
• compute requirements;
• interoperability;
• benchmarks;
• scalability.

E. Scenario testing

Alternative plausible technological futures should be examined.

F. Sensitivity analysis

The authority should determine whether the conclusion survives reasonable changes to assumptions.

29. Six Core Lessons from the Case Law

The common thread is that competition law permits predictive analysis but does not eliminate the need for evidentiary discipline.

30. Conclusion

Evidentiary challenges in predicting AI market evolution represent one of the central methodological problems of modern competition law.

AI markets are characterised by:

• rapid technological change;
• uncertain innovation;
• limited historical data;
• network effects;
• open-source disruption;
• vertically integrated ecosystems;
• rapidly changing business models; and
• significant uncertainty concerning future entry.

The appropriate legal response is not to demand impossible certainty, but neither should competition authorities rely upon technological speculation.

The strongest approach is a probabilistic, evidence-based and counterfactual methodology in which:

technical evidence + commercial evidence + internal documents + economic analysis + expert evidence + scenario analysis

are considered together.

The jurisprudence of Airtours, Tetra Laval, Ryanair/Aer Lingus, Bertelsmann/Sony v Impala, Microsoft and United States v Microsoft demonstrates an enduring principle: future competitive effects may be established through inference, but the inference must be anchored in reliable evidence and a coherent causal chain.

For AI markets, therefore, the critical evidentiary question is not:

"Can the authority predict the future perfectly?"

It is:

"Has the authority demonstrated, on the totality of sufficiently reliable evidence, that the predicted evolution of the AI market is a sufficiently grounded and legally defensible basis for intervention?"