Evidence Admissibility Standards For Machine-Generated Economic Inference

1. Introduction

Modern competition-law investigations increasingly rely on machine-generated economic inferences: algorithmic price-concentration models, automated event studies, market-reconstruction models, anomaly detection, econometric simulations, network analysis, machine-learning predictions, and automated assessments of market power or coordinated conduct.

The central legal question is not simply whether an algorithm produces a statistically sophisticated result. The more important question is:

When, and under what conditions, should a court or competition authority admit and rely upon an economic inference generated wholly or partly by a machine?

There is no universally established doctrine that makes machine-generated economic evidence automatically admissible or inadmissible. Existing evidentiary principles generally focus on reliability, relevance, methodological validity, transparency, reproducibility, provenance, expert competence and the possibility of meaningful challenge.

A useful distinction is therefore:

Machine-generated output → admissibility → evidential reliability → probative weight → substantive inference

An output can be admissible but ultimately receive little weight if the underlying model is poorly specified, trained on biased data, impossible to reproduce, or incapable of explaining its conclusion.

2. Meaning of Machine-Generated Economic Inference

A machine-generated economic inference occurs when software, an algorithm, an econometric model or an AI system transforms raw information into a conclusion relevant to an economic or competition-law issue.

Examples include:

• estimating a relevant market;
• predicting counterfactual prices;
• estimating cartel overcharges;
• detecting coordinated pricing;
• identifying parallel conduct;
• estimating market power;
• calculating diversion ratios;
• identifying exclusionary effects;
• reconstructing prices absent an alleged infringement;
• detecting suspicious bidding patterns;
• determining whether conduct caused foreclosure;
• predicting consumer switching;
• estimating damages;
• identifying structural breaks in market data.

For example:

Raw transaction data → algorithm → predicted competitive price → estimated overcharge

or:

Millions of bids → machine-learning model → probability of bid coordination

The legal difficulty is that the final inference may be mathematically generated but legally contested.

3. Core Evidentiary Principle

The strongest approach is to treat machine-generated economic evidence according to the following principle:

The machine does not become the witness merely because it generated the inference. The party relying on the output must establish the reliability of the data, methodology, processing system and inferential chain through admissible evidence.

This produces four separate questions.

A. Authenticity

Can the party establish that the data and output are what they purport to be?

B. Reliability

Was the system functioning properly and was the methodology scientifically/economically sound?

C. Reproducibility

Can another competent expert examine the methodology and obtain substantially similar results?

D. Probative value

Even if technically reliable, does the inference actually prove the proposition for which it is being offered?

This distinction is particularly important in competition proceedings because correlation is not necessarily causation, and a statistical prediction is not automatically proof of unlawful conduct.

4. Relevance and Materiality

Machine-generated evidence must first be relevant to a legally material issue.

For example, an AI model showing that two competitors changed prices within several minutes of each other may be relevant to:

• coordination;
• algorithmic monitoring;
• conscious parallelism;
• information exchange;
• market transparency.

But the same model does not necessarily establish:

• an agreement;
• concerted practice;
• dominance;
• abuse;
• causation of consumer harm.

Thus:

Algorithmic correlation ≠ legal infringement

The evidentiary chain must connect the machine output to the legal elements of the offence.

5. Data Provenance

Data provenance is particularly important where economic conclusions depend upon large datasets.

A party should ordinarily be able to establish:

1. where the data originated;
2. when it was collected;
3. how it was stored;
4. whether it was altered;
5. what variables were included;
6. what variables were excluded;
7. how missing observations were treated;
8. whether duplicates were removed;
9. how outliers were treated;
10. how the dataset was transferred into the analytical system.

An apparently sophisticated economic model can become unreliable if its underlying dataset is defective.

Example

Suppose an authority's algorithm concludes that a dominant platform increased prices by 14%.

If:

• 8% of transactions are missing,
• cancelled transactions are excluded,
• promotional prices are incorrectly coded,
• inflation adjustments are inconsistent,

the output may have substantial mathematical sophistication but limited evidentiary value.

6. Black-Box Algorithms

The most difficult issue is the black-box problem.

A machine-learning system may produce:

“Probability of coordinated pricing = 93%.”

But the opposing party may reasonably ask:

• Which variables produced the result?
• What training data were used?
• What assumptions were incorporated?
• What model architecture was used?
• What alternative models were tested?
• What was the error rate?
• Was the model overfitted?
• How sensitive is the conclusion to particular variables?
• Could legitimate parallel conduct generate the same result?

If these questions cannot be answered, the evidence may still technically be admitted in some systems, but its weight and reliability may be seriously undermined.

7. Expert Evidence and Machine-Generated Inference

Where a machine-generated economic inference is presented through an economist or technical expert, the expert should normally be able to explain:

Input → Method → Processing → Output → Economic interpretation

An expert cannot simply say:

“The computer produced this result.”

The expert must understand the methodology sufficiently to defend its application.

This is particularly important where the model contains:

• proprietary algorithms;
• neural networks;
• automated variable selection;
• automated feature engineering;
• reinforcement learning;
• probabilistic inference;
• synthetic data;
• automated counterfactual modelling.

8. Six Important Case Laws

Case 1 — Daubert v Merrell Dow Pharmaceuticals, Inc. (1993)

Jurisdiction: United States
Court: U.S. Supreme Court

Principle

Daubert is one of the foundational authorities concerning the admissibility of scientific and technical expert evidence.

The Court emphasized judicial scrutiny of the reliability of expert methodology, including considerations such as:

• whether the methodology can be tested;
• whether it has been subjected to peer review;
• its known or potential error rate;
• applicable standards;
• general acceptance.

Relevance to machine-generated economic inference

A machine-learning or econometric model should not receive automatic evidentiary legitimacy simply because it is technically sophisticated.

A court may examine:

• whether the model is testable;
• its error rate;
• validation methodology;
• methodological acceptance;
• whether it is reliably applied to the facts.

Competition-law significance

If an economic expert uses an AI model to infer cartel overcharges or market power, Daubert-type scrutiny supports examination of the model itself rather than merely accepting the expert's conclusion.

9. Case 2 — General Electric Co. v Joiner (1997)

Jurisdiction: United States
Court: U.S. Supreme Court

Principle

Joiner confirmed that courts may examine whether there is an adequate analytical connection between the underlying evidence and the expert's ultimate conclusion.

The Court rejected the notion that an expert could simply bridge substantial analytical gaps through assertion.

Relevance

This is extremely important for AI-generated economic inference.

Suppose:

Data → AI model → statistical correlation → conclusion of anticompetitive coordination

The court may ask whether the intermediate analytical steps actually support the ultimate conclusion.

Key lesson

A sophisticated model cannot cure an analytical gap.

Thus:

More computation does not necessarily mean stronger proof.

10. Case 3 — Kumho Tire Co. v Carmichael (1999)

Jurisdiction: United States
Court: U.S. Supreme Court

Principle

Kumho extended the reliability-oriented approach of Daubert beyond traditional scientific testimony to technical and specialized expert evidence.

The Court recognized that courts must exercise a gatekeeping role with respect to expert evidence.

Relevance to AI-generated economic evidence

Economic evidence generated through:

• software;
• statistical systems;
• technical algorithms;
• data analytics;

cannot escape reliability scrutiny merely because it is described as “technical” rather than “scientific.”

Importance

The principle is especially relevant to algorithmic competition cases because economic models increasingly combine:

• economics;
• statistics;
• computer science;
• machine learning;
• data engineering.

A court may therefore scrutinize the technical foundation as well as the economic interpretation.

11. Case 4 — Tyson Foods, Inc. v Bouaphakeo (2016)

Jurisdiction: United States
Court: U.S. Supreme Court

Principle

Tyson Foods concerned statistical and representative evidence used to establish compensation-related claims.

The Court recognized that statistical evidence can be legitimate evidence where it is capable of establishing relevant propositions, while emphasizing that its appropriateness depends upon the circumstances.

Relevance to machine-generated economic inference

The case illustrates an important proposition:

Statistical evidence is not inherently inadmissible simply because it relies upon aggregation or modelling.

For competition cases, this supports the potential use of:

• statistical sampling;
• regression models;
• representative transaction datasets;
• automated classification;
• predictive economic models.

But the model must still be appropriately connected to the proposition being proved.

Competition-law application

A machine-generated estimate of cartel overcharge or damages may be useful where transaction-level evidence is extraordinarily large.

The issue becomes whether the methodology produces a sufficiently reliable inference rather than whether every individual transaction is separately analysed.

12. Case 5 — Airtours plc v Commission (2002)

Jurisdiction: European Union
Court: General Court of the European Union

Principle

Airtours is a major EU competition-law authority concerning collective dominance and the evidentiary requirements for establishing coordinated market behaviour.

The Court required sufficiently convincing evidence to establish the conditions necessary for tacit coordination.

Relevance to machine-generated inference

This is highly relevant to algorithmic competition enforcement.

Suppose an authority's AI system identifies:

• parallel price movements;
• repeated price matching;
• stable margins;
• reduced deviations;
• predictable reactions.

These patterns may be economically suggestive.

But Airtours demonstrates that structural or statistical evidence must satisfy the substantive legal test.

Key lesson

A predictive model identifying a “high probability of coordination” does not itself establish the legally required conditions for coordinated conduct.

13. Case 6 — Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba (2016)

Jurisdiction: European Union
Court: Court of Justice of the European Union

Principle

Eturas involved an electronic travel-booking system through which a technical restriction concerning discounts was introduced.

The case is particularly significant for the evidentiary implications of electronic systems and communications in competition law.

The Court examined whether knowledge of the electronically transmitted message and continued participation could support an inference of concerted practice.

Relevance to machine-generated economic inference

Eturas demonstrates that electronic-system evidence can form part of the evidentiary chain establishing competition-law conduct.

But the crucial issue remains:

What does the electronic evidence actually establish?

A system-generated event may demonstrate:

• that a technical instruction existed;
• that information was transmitted;
• that participants had access to it;
• that subsequent behaviour occurred.

It does not automatically establish every element of the infringement.

Importance for AI systems

The case is particularly useful where a competition authority uses:

• platform logs;
• algorithmic records;
• automated pricing instructions;
• system messages;
• API communications;
• machine-generated audit trails.

14. Case 7 — Tetra Laval BV v Commission (2002)

Jurisdiction: European Union
Court: Court of Justice of the European Union

Principle

Tetra Laval is an important authority concerning the evidentiary standard for prospective economic predictions in merger control.

The Court emphasized that prospective theories must be supported by sufficiently convincing evidence.

Relevance

Machine-generated economic models are often predictive rather than historical.

For example:

“The acquisition will increase prices by 12%.”

or:

“The merger will make algorithmic coordination more sustainable.”

Such predictions cannot be treated as established facts merely because they emerge from a quantitative model.

Key principle

Forecasting requires evidentiary substantiation.

An AI-generated prediction should therefore disclose its:

• assumptions;
• scenarios;
• model limitations;
• sensitivity analysis;
• counterfactual;
• uncertainty range.

15. Case 8 — Commission v Tetra Laval (C-12/03 P)

The subsequent appellate consideration of Tetra Laval reinforced the importance of examining whether a prospective theory of harm is supported by sufficiently persuasive evidence.

This is significant for machine-generated models because competition authorities increasingly use predictive systems to estimate:

• merger effects;
• foreclosure;
• coordinated effects;
• innovation effects;
• entry barriers.

A model may strengthen an inference, but it cannot replace the underlying evidentiary burden.

16. The Evidentiary Chain for AI-Generated Economic Evidence

A useful legal framework is:

Stage 1 — Data authenticity

Where did the data come from?

↓

Stage 2 — Data integrity

Was the data altered, corrupted or selectively filtered?

↓

Stage 3 — Algorithmic reliability

Does the model operate as represented?

↓

Stage 4 — Methodological validity

Is the economic methodology appropriate?

↓

Stage 5 — Reproducibility

Can an opposing expert reproduce or meaningfully test the analysis?

↓

Stage 6 — Error assessment

What is the model's error or uncertainty rate?

↓

Stage 7 — Causal connection

Does the output establish causation or merely correlation?

↓

Stage 8 — Legal relevance

Does the inference prove an element of the competition-law claim?

↓

Stage 9 — Weight

How persuasive should the tribunal regard the evidence?

17. Admissibility Versus Weight

This distinction is crucial.

Admissibility

The question is:

“Can the tribunal consider this evidence?”

Weight

The question is:

“How much reliance should the tribunal place upon it?”

For example, an authority may submit an algorithmic model that is:

• properly authenticated;
• technically functional;
• relevant;
• supported by an expert.

The tribunal might admit it.

But after hearing the opposing expert, it might conclude:

“The model is admissible but unreliable for proving causation.”

Thus:

Admissible ≠ persuasive

This distinction is likely to become increasingly important in AI-related competition litigation.

18. Explainability as an Evidentiary Requirement

Explainability should not necessarily mean that every mathematical operation must be disclosed.

Rather, the opposing party should ordinarily receive sufficient information to understand:

1. the relevant data;
2. the variables;
3. the methodology;
4. the principal assumptions;
5. the model's limitations;
6. the validation procedure;
7. the output;
8. the relationship between the output and the legal proposition.

This can be called functional explainability.

19. Proprietary Algorithms and Confidentiality

A major practical problem arises where the algorithm belongs to a private company.

The party may argue:

“The model is proprietary and therefore cannot be disclosed.”

Competition proceedings, however, require a balance between:

• trade-secret protection;
• confidentiality;
• due process;
• procedural equality;
• effective cross-examination.

Possible safeguards include:

• confidentiality rings;
• expert-only access;
• source-code inspection under restrictions;
• disclosure of model documentation;
• disclosure of variable definitions;
• independent validation;
• controlled reproducibility.

A completely unreviewable algorithm creates a serious procedural problem where its output is relied upon as decisive evidence.

20. Reproducibility

Reproducibility is especially important for machine-generated evidence.

An opposing expert should ideally be able to reproduce the result using:

• the same data;
• the same code/version;
• the same model;
• the same parameters;
• the same assumptions.

Perfect numerical identity may not always be necessary.

But there should be sufficient transparency to determine whether the result is robust rather than accidental.

21. Model Drift

AI systems create a problem that traditional economic evidence does not always present: model drift.

An algorithm that produced a reliable output in 2024 may behave differently in 2026 because:

• market conditions changed;
• consumer behaviour changed;
• data distributions changed;
• the model was retrained;
• software was updated;
• variables changed;
• an API changed;
• training data changed.

Therefore, the tribunal may need to know:

Which version of the model produced the challenged inference?

This creates a form of algorithmic chain of custody.

22. Training Data as Evidence

Where machine learning is involved, training data can be just as important as the final output.

For example, a model trained primarily on:

• cartelised markets;
• highly concentrated industries;
• historical enforcement cases;

might systematically over-predict coordination.

Conversely, a model trained primarily on competitive markets might under-detect coordination.

Therefore:

Training-data composition can affect the evidentiary reliability of the final inference.

23. False Positives and False Negatives

Competition authorities should distinguish:

False positive

The algorithm identifies unlawful conduct where none exists.

False negative

The algorithm fails to identify actual unlawful conduct.

Both matter.

For enforcement systems, an unusually high false-positive rate can generate:

• unnecessary investigations;
• reputational damage;
• burdensome disclosure;
• incorrect merger intervention;
• mistaken cartel allegations.

Therefore, error rates should be disclosed wherever material.

24. Algorithmic Pricing and Tacit Coordination

Machine-generated evidence becomes particularly difficult in algorithmic pricing cases.

Suppose four competitors use dynamic pricing systems and prices become highly synchronized.

An algorithm detects:

97% price convergence.

This may establish economic coordination, but the legal conclusion remains separate.

Possible explanations include:

1. common cost shocks;
2. common software;
3. common market information;
4. rational price matching;
5. conscious parallelism;
6. explicit information exchange;
7. algorithmic coordination;
8. an actual agreement.

The machine cannot automatically choose the legally correct explanation.

25. Event Studies and Machine-Generated Inference

Event-study methodology is another important application.

A model may calculate:

Abnormal return = Actual return − Expected return

A large abnormal return following an announcement can be useful evidence.

But the tribunal should consider:

• alternative market events;
• market-wide shocks;
• industry-specific events;
• leakage;
• confounding variables;
• choice of event window;
• model specification.

The same principle applies:

Statistical significance does not automatically equal legal causation.

26. AI-Generated Market Definition

Machine learning may classify consumers according to:

• purchasing behaviour;
• switching patterns;
• search behaviour;
• transaction histories;
• geographic information.

It could produce an inferred market boundary.

However, traditional competition-law concepts such as:

• substitutability;
• demand response;
• supply substitution;
• competitive constraints;

remain legal-economic concepts.

The algorithm may assist the analysis but should not silently replace the legal test.

27. AI-Generated Damages Estimates

Machine-generated damages models can be particularly valuable where millions of transactions are involved.

For example:

Observed price − counterfactual price = estimated overcharge

The tribunal should nevertheless examine:

• counterfactual construction;
• comparator markets;
• control variables;
• model specification;
• pass-on;
• temporal effects;
• uncertainty;
• confidence intervals.

A single machine-generated number should not be treated as an inherently objective fact.

28. Procedural Fairness

Machine-generated evidence creates a potentially serious procedural issue where one party has:

• the complete dataset;
• the algorithm;
• the source code;
• the model architecture;
• the technical personnel,

while the opposing party receives only:

“The algorithm determined X.”

Such a system may undermine meaningful challenge.

Accordingly, procedural fairness supports disclosure sufficient to allow the opposing party to test the inferential chain, subject to legitimate confidentiality protections.

29. Proposed Admissibility Test

A useful modern test can be formulated as the MACHINE Test:

M — Materiality

Is the output relevant to an issue that legally matters?

A — Authenticity

Can the data and computational output be authenticated?

C — Computational reliability

Did the system operate properly?

H — Human/expert validation

Can a qualified expert explain and defend the methodology?

I — Independence and integrity

Can manipulation, selective inputs or undisclosed assumptions be excluded?

N — Non-speculative inference

Does the output support the conclusion without an unjustified inferential leap?

E — Explainability and reproducibility

Can the opposing party meaningfully test the result?

This provides a useful framework for courts and competition authorities dealing with AI-generated economic evidence.

30. Comparative Position

31. Key Legal Lessons From the Cases

The cases collectively support several propositions.

1. Sophistication is not reliability

Daubert and Kumho Tire demonstrate that technical complexity does not remove judicial gatekeeping.

2. Analytical gaps remain fatal

Joiner demonstrates that the conclusion must actually follow from the underlying evidence.

3. Statistical evidence can be legitimate

Tyson Foods illustrates that appropriately designed statistical evidence can prove matters that cannot practically be demonstrated transaction by transaction.

4. Electronic systems can constitute evidence

Eturas demonstrates the significance of electronic-system evidence in competition proceedings.

5. Predictive theories require convincing evidence

Tetra Laval is particularly relevant to AI-generated prospective economic predictions.

6. Economic inference must satisfy the substantive legal test

Airtours demonstrates that statistical or structural evidence must ultimately establish the legal conditions for coordinated conduct.

32. Practical Checklist for Courts and Competition Authorities

Before relying substantially on machine-generated economic inference, the decision-maker should ask:

1. Who generated the data?
2. Who controlled the model?
3. What version of the algorithm was used?
4. What data were included?
5. What data were excluded?
6. Were variables transformed?
7. Were missing values imputed?
8. What assumptions were made?
9. Was the model validated?
10. What is its error rate?
11. Has it been independently tested?
12. Can the analysis be reproduced?
13. Can an opposing expert challenge it?
14. Does it establish correlation or causation?
15. Does it prove the relevant legal element?
16. Are alternative explanations addressed?
17. Has model drift occurred?
18. Are proprietary restrictions preventing meaningful scrutiny?
19. Are confidence intervals or uncertainty ranges disclosed?
20. What weight should ultimately be assigned to the output?

33. Conclusion

Machine-generated economic inference should not be treated as inherently unreliable merely because a computer produced it, nor should it receive presumptive authority merely because it is algorithmic.

The appropriate approach is to examine the entire evidentiary chain:

Data provenance → data integrity → algorithmic methodology → expert validation → reproducibility → error/uncertainty → causal inference → legal relevance → evidential weight.

The most important lesson from Daubert, Joiner, Kumho Tire, Tyson Foods, Airtours, Eturas and Tetra Laval is that technical sophistication cannot substitute for evidentiary reliability or legal proof.

In future competition litigation, the decisive question is therefore unlikely to be:

“Did an AI system reach this conclusion?”

It will increasingly be:

“Can the tribunal independently understand, test and evaluate why the system reached that conclusion, and does that conclusion actually establish the legally relevant proposition?”