Expert Witness Standards For Machine Learning Systems In Antitrust Cases

1. Introduction

Machine-learning systems are increasingly relevant to antitrust disputes involving algorithmic pricing, personalized pricing, recommendation systems, ranking algorithms, digital advertising, platform self-preferencing, automated bidding, demand prediction, market allocation, algorithmic collusion, data-driven exclusion, and merger effects.

The difficulty is that an expert may not merely present an economic model. The expert may rely upon a machine-learning model that itself contains multiple layers of data selection, feature engineering, training, validation, prediction, optimization, and interpretation.

Accordingly, the central question is not simply:

Is the expert qualified to discuss machine learning?

It is:

Has the expert demonstrated that the ML system, the underlying data, the methodology, and the application of the model are sufficiently reliable, valid, reproducible, and relevant to the particular antitrust question?

In the United States, the principal framework is Federal Rule of Evidence 702, as interpreted through Daubert, Joiner, and Kumho Tire. The 2023 amendment to Rule 702 expressly emphasizes that the proponent must demonstrate that it is more likely than not that the requirements are satisfied, including reliable application of principles and methods to the facts.

The DOJ's antitrust guidance similarly recognizes economic modeling, econometric analysis, data, testimony, and business documents as important evidence, while emphasizing the probative value and reliability of the analysis.

2. Why ML Expert Evidence Creates Special Problems

Traditional economic expert evidence might involve:

• regression analysis;
• market-share calculations;
• price elasticity;
• merger simulation;
• event studies;
• damages calculations;
• market definition.

An ML-based expert analysis can additionally involve:

1. massive datasets;
2. proprietary algorithms;
3. neural networks;
4. automated feature selection;
5. ensemble models;
6. black-box predictions;
7. continuously changing models;
8. feedback loops;
9. training-data problems;
10. model drift;
11. algorithmic optimization;
12. automated pricing decisions.

Therefore, an antitrust court should distinguish between technical sophistication and evidentiary reliability.

A highly sophisticated neural network is not automatically better evidence than a transparent regression model.

3. Governing Legal Standard

A. Federal Rule of Evidence 702

An expert may testify where the expert's specialized knowledge will help the fact-finder and the testimony:

• rests on sufficient facts or data;
• results from reliable principles and methods; and
• reflects reliable application of those principles and methods to the facts.

The 2023 amendment strengthened the emphasis on the proponent's burden to demonstrate admissibility and on reliable application.

This is particularly important for ML because an expert cannot simply say:

"The algorithm produced this result."

The expert must explain why the algorithm's output is sufficiently reliable for the particular legal proposition for which it is offered.

4. Daubert v. Merrell Dow Pharmaceuticals, Inc.

Principle

Daubert established the federal judicial gatekeeping function for scientific expert evidence.

Relevant considerations include:

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

These are flexible considerations rather than rigid requirements.

Application to ML antitrust evidence

An ML expert should therefore be prepared to identify:

• the model architecture;
• training methodology;
• validation methodology;
• testing procedure;
• error rate;
• confidence intervals where appropriate;
• reproducibility;
• relevant benchmarks;
• limitations;
• assumptions.

A model that cannot be meaningfully tested or whose outputs cannot be reproduced creates a substantial admissibility problem.

Antitrust significance:
If an expert claims that an algorithm caused supracompetitive prices, the court should ask whether the methodology actually distinguishes algorithmic effects from ordinary market factors.

5. General Electric Co. v. Joiner

Principle

Joiner confirmed that courts may examine the connection between the methodology and the expert's ultimate conclusion.

It is insufficient for an expert to use reliable scientific methods but make an unsupported inferential leap.

ML application

This is particularly important where an expert moves from:

"The model predicts higher prices under condition X"

to:

"The defendant's algorithm caused anticompetitive price increases."

Those are not necessarily equivalent propositions.

An ML model may identify correlation, while antitrust law may require evidence concerning:

• causation;
• exclusionary effects;
• competitive effects;
• market power;
• foreclosure;
• consumer harm.

Thus, Joiner supports scrutiny of the inferential bridge between the model's output and the antitrust conclusion.

6. Kumho Tire Co. v. Carmichael

Principle

Kumho Tire extended the gatekeeping obligation beyond traditional "scientific" testimony to technical and specialized expertise.

This is extremely important for machine-learning evidence.

An ML expert cannot avoid Rule 702 scrutiny merely by characterizing the testimony as:

• computer science;
• engineering;
• data science;
• software expertise;
• technical experience.

The court may examine the reliability of the technical methodology.

ML antitrust application

For example, an expert claiming that two pricing algorithms learned to coordinate may need to establish:

• how the algorithms were modeled;
• what information each algorithm received;
• whether the algorithms actually observed competitors' prices;
• whether coordination was intentional, emergent, or merely simulated;
• whether alternative explanations were tested;
• whether the experiment can be replicated.

Kumho Tire therefore provides the doctrinal foundation for rigorous scrutiny of technical ML testimony.

7. Concord Boat Corp. v. Brunswick Corp.

Principle

Concord Boat is particularly important to antitrust expert evidence because the court scrutinized an economic model that was insufficiently connected to the actual competitive conditions of the industry.

The case illustrates an essential principle:

An economically respectable model may still be unreliable if its assumptions do not adequately correspond to the real-world market.

The DOJ has specifically discussed Concord Boat in connection with the admissibility of economic models and the requirement that expert testimony fit the factual circumstances of the case.

ML application

This becomes even more important with machine learning.

Suppose an expert trains a model using:

• national pricing data;

but the alleged conduct concerns:

• a particular geographic market;

or trains a model using:

• historical consumer behavior;

when the alleged anticompetitive conduct concerns:

• a newly introduced AI pricing system.

The model may have excellent predictive accuracy while nevertheless having poor antitrust fit.

Key lesson

Predictive accuracy ≠ legal relevance.

8. In re Google Play Consumer Antitrust Litigation

This litigation illustrates the importance of Rule 702 scrutiny in modern platform antitrust litigation.

Courts have examined expert economic evidence under Rule 702, including questions concerning:

• sufficient data;
• reliable principles and methods;
• reliable application;
• economic modeling;
• classwide methodology.

The court recognized the significance of the 2023 Rule 702 amendment and the requirement that the proponent establish the necessary elements of admissibility.

ML relevance

In platform cases, experts may use machine-learning techniques to analyze:

• app-store transactions;
• ranking;
• consumer substitution;
• developer behavior;
• pricing;
• commissions;
• user switching.

The expert must establish that the model measures the relevant competitive phenomenon rather than merely predicting user behavior.

9. In re Google Digital Advertising Antitrust Litigation

This litigation provides a particularly useful modern example because expert evidence concerning digital advertising involves sophisticated economic and technological questions.

Courts have applied Daubert and Rule 702 to experts analyzing:

• digital advertising markets;
• auction mechanisms;
• damages;
• benchmarks;
• competitive effects.

One recent ruling emphasized that the court must examine whether the expert's methodology is reliable and whether it is properly applied to the case. It also recognized benchmark analysis as an accepted methodology where appropriately constructed.

ML significance

Digital advertising systems increasingly use:

• automated bidding;
• prediction models;
• personalization;
• auction optimization;
• machine-learning targeting.

Therefore, an expert should distinguish between:

A. What the algorithm predicts

and

B. What the prediction means economically.

For example:

Algorithm predicts advertiser conversion probability → does not automatically establish → market foreclosure.

There must be an economically defensible intermediate analysis.

10. Associated Newspapers Ltd. v. Google LLC

Recent litigation concerning Google's digital advertising activities provides another illustration of Rule 702's application to sophisticated economic expert evidence.

The court emphasized that Daubert is a flexible inquiry and that the reliability of an expert's analysis must be examined at each significant stage. It also recognized that criticisms going merely to the weight of properly grounded evidence should ordinarily be distinguished from defects serious enough to make the methodology unreliable.

ML significance

This distinction is critical.

An opposing party might argue:

"The ML model could have been more accurate."

That may affect weight.

But if the expert:

• used contaminated training data;
• trained and tested on overlapping observations;
• ignored obvious confounders;
• failed to disclose material preprocessing;
• cannot reproduce the results;

the problem may go to admissibility, not merely weight.

11. In re High Fructose Corn Syrup Antitrust Litigation

This antitrust litigation is important for the broader proposition that economic evidence must be capable of demonstrating the alleged competitive mechanism rather than merely describing parallel market outcomes.

ML relevance

In algorithmic-collusion litigation, an expert may observe:

• parallel prices;
• synchronized price increases;
• similar algorithmic responses.

But parallel outcomes alone do not necessarily establish unlawful coordination.

The expert should investigate competing explanations such as:

• common cost shocks;
• common demand shocks;
• industry-wide software;
• common input prices;
• automated reactions to public information;
• rational unilateral optimization.

Key standard

The ML expert must test plausible alternative explanations.

12. In re Air Cargo Shipping Services Antitrust Litigation

Air-cargo antitrust litigation demonstrates the importance of econometric and statistical evidence in determining whether observed prices and conduct are consistent with coordinated conduct.

ML relevance

Modern versions of the same inquiry might use:

• anomaly detection;
• clustering;
• network analysis;
• causal forests;
• reinforcement-learning simulations;
• graph neural networks;
• temporal models.

But technological sophistication does not change the fundamental evidentiary question:

Does the methodology reliably establish the competitive proposition being asserted?

An ML model detecting anomalous pricing is evidence of an anomaly—not automatically evidence of an antitrust violation.

13. Six Core Case Laws at a Glance

The first three are the foundational expert-evidence authorities; the latter cases demonstrate their particular significance in antitrust/economic disputes.

14. Proposed Expert-Witness Standard for ML Antitrust Cases

A useful framework is a seven-stage ML reliability test.

Stage 1 — Expert Qualification

The expert should demonstrate appropriate expertise in the specific field involved.

A data scientist specializing in image recognition should not automatically qualify as an expert in:

• antitrust economics;
• market definition;
• merger simulation;
• algorithmic pricing;
• causal inference.

Conversely, an economist should not automatically qualify as an expert on:

• neural-network architecture;
• software engineering;
• model training;
• cybersecurity.

Preferred approach

Complex ML antitrust disputes may require multidisciplinary expert teams, such as:

Economist + ML/data scientist + industry/technical expert.

15. Stage 2 — Data Integrity

The expert should disclose:

• data sources;
• collection period;
• sampling methodology;
• missing-data treatment;
• data cleaning;
• feature construction;
• exclusions;
• transformations;
• labeling methodology.

Particular attention should be paid to:

Data leakage

Information unavailable at the relevant historical point must not inadvertently enter the training data.

Selection bias

The dataset must represent the relevant competitive environment.

Survivorship bias

The model must not ignore firms or transactions that exited because of the alleged conduct.

Measurement error

Prices, quantities, users, transactions and market shares must be accurately defined.

16. Stage 3 — Model Reliability

The expert should disclose:

• model type;
• architecture;
• parameters;
• hyperparameters;
• training procedure;
• validation procedure;
• test set;
• performance metrics;
• error rates.

Depending on the model, appropriate metrics might include:

• precision;
• recall;
• F1 score;
• AUC;
• RMSE;
• MAE;
• calibration;
• out-of-sample prediction accuracy.

But the appropriate metric depends upon the antitrust question.

High predictive accuracy alone does not establish causation.

17. Stage 4 — Reproducibility

A particularly important requirement is whether another competent expert can reproduce the analysis.

The expert should, where appropriate, preserve:

• source data;
• code;
• model version;
• software environment;
• random seeds;
• preprocessing pipeline;
• feature definitions;
• model parameters;
• training configuration.

This is especially important where the defendant's algorithm is continually changing.

Reproducibility problem

If an expert says:

"The model currently produces this result."

the court should ask:

"Did the same model produce the same result during the relevant antitrust period?"

18. Stage 5 — Causal Validity

This is arguably the most important distinction.

Prediction

"When X occurs, Y is likely to occur."

Causation

"X caused Y."

Antitrust litigation frequently requires the second proposition.

An ML model trained for prediction may therefore be inappropriate for causal inference without additional methodology.

The expert may need:

• causal inference;
• natural experiments;
• difference-in-differences;
• instrumental variables;
• regression discontinuity;
• controlled experiments;
• counterfactual analysis.

19. Stage 6 — Counterfactual Reliability

Antitrust analysis frequently depends upon a counterfactual:

What would the market have looked like absent the challenged conduct?

ML systems can construct counterfactual predictions, but the expert must explain:

• what the counterfactual is;
• what assumptions generate it;
• whether those assumptions are testable;
• whether the model has historical support;
• whether the counterfactual extrapolates outside the training distribution.

For example:

Actual world:
Platform algorithm suppresses rival products.

Counterfactual:
What ranking would have occurred without the suppression?

The expert must avoid simply instructing the ML model to "remove anticompetitive behavior" and treating its output as objectively established.

20. Stage 7 — Antitrust Fit

The ultimate question is whether the model addresses the legal-economic issue.

The expert should identify the exact proposition supported by the model.

For example:

This distinction is essential to prevent algorithmic outputs from becoming substitutes for legal reasoning.

21. Standards for Algorithmic-Collusion Experts

Algorithmic collusion presents a particularly difficult problem.

An expert should examine:

1. Communication

Did algorithms receive information about rivals?

2. Observability

Could one algorithm observe another's price or output?

3. Optimization objective

What was the algorithm actually instructed to optimize?

4. Learning mechanism

Did it use:

• reinforcement learning;
• supervised learning;
• unsupervised learning;
• dynamic optimization?

5. Constraints

Were there:

• pricing floors;
• inventory constraints;
• margin requirements;
• regulatory limits?

6. Alternative explanations

Could the observed behavior result from independent optimization?

7. Counterfactual experimentation

What happens when:

• one algorithm is replaced;
• information is restricted;
• optimization objectives change;
• competitors are introduced?

22. Standards for Algorithmic Pricing Experts

In pricing cases, an expert should examine:

• input variables;
• competitor-price data;
• demand forecasts;
• inventory information;
• customer segmentation;
• elasticity estimates;
• optimization objectives;
• price-update frequency;
• experimentation;
• feedback mechanisms.

A particularly important issue is feedback loops.

For example:

Algorithm raises price → competitors observe price → competitor algorithm raises price → first algorithm interprets increase as market demand → price rises again.

An expert should determine whether this constitutes:

• independent adaptation;
• conscious parallelism;
• algorithmically facilitated coordination;
• or some other economically significant mechanism.

23. Standards for Platform Ranking Algorithms

Where an ML ranking system allegedly excludes rivals, the expert should examine:

1. ranking inputs;
2. ranking weights;
3. training labels;
4. treatment of sponsored content;
5. personalization;
6. experimentation;
7. changes to the ranking algorithm;
8. competitor visibility;
9. consumer substitution;
10. downstream effects.

The expert should distinguish:

technical ranking change

from

competitive foreclosure.

24. Standards for ML-Based Market Definition

Machine learning may be used to identify:

• consumer substitution;
• geographic patterns;
• product similarity;
• switching behavior;
• demand clusters.

However, clustering does not automatically constitute legal market definition.

An algorithm may classify products as similar because of:

• language;
• technical characteristics;
• consumer reviews;
• purchasing behavior.

But antitrust market definition may depend upon:

• substitutability;
• competitive constraints;
• demand response;
• supply substitution;
• geographic conditions.

Thus:

Machine-learning classification should inform—not replace—the legal-economic market-definition inquiry.

25. Black-Box Models

A major problem is the use of opaque models.

The argument:

"The neural network is too complex to explain."

should not automatically be accepted as an excuse for inadequate expert disclosure.

The expert should ordinarily provide, to the extent technically feasible:

• model architecture;
• feature importance;
• sensitivity analysis;
• robustness analysis;
• counterfactual explanations;
• validation results;
• error analysis.

Interpretability methods such as SHAP or LIME may assist, but they themselves must be used carefully because an interpretability technique is not necessarily a complete explanation of causation.

26. Model Drift

ML systems may change over time.

This creates a special antitrust problem.

A model that operated in:

2022

may have been retrained in:

2023

and materially modified in:

2024.

An expert must therefore establish which model version corresponds to the alleged conduct.

Otherwise, the expert may be analyzing the wrong algorithm.

27. Adversarial and Strategic Manipulation

Experts should also consider whether the data or model could have been manipulated.

Potential problems include:

• strategic input manipulation;
• bots;
• fake transactions;
• synthetic data;
• manipulated reviews;
• adversarial examples;
• strategic competitor behavior;
• gaming of ranking systems.

A model that performs well on ordinary test data may perform badly when market participants strategically adapt to it.

28. Confidential and Proprietary Algorithms

Antitrust defendants may argue that disclosure of:

• source code;
• model weights;
• training data;
• proprietary features;

would expose trade secrets.

Courts may therefore need mechanisms such as:

• protective orders;
• confidentiality designations;
• expert-only access;
• source-code review;
• confidentiality rings.

But confidentiality should not eliminate the opposing party's meaningful opportunity to test the expert's methodology.

29. Weight Versus Admissibility

Courts should carefully distinguish between:

Minor methodological criticism

Examples:

• alternative parameter choice;
• reasonable disagreement about model specification;
• different but defensible training methodology.

These may affect weight.

Fundamental reliability defect

Examples:

• training and test data contamination;
• undisclosed data manipulation;
• no meaningful validation;
• impossible reproduction;
• unsupported causal inference;
• model unrelated to the relevant market.

These may justify exclusion.

This distinction is consistent with modern Rule 702 jurisprudence emphasizing rigorous examination of the expert's factual basis, methodology, and application.

30. A Proposed "ML Antitrust Expert Reliability Matrix"

31. Special Problems in Merger Cases

ML experts may be used to predict:

• diversion ratios;
• customer switching;
• innovation effects;
• product substitution;
• future competitive entry;
• coordinated effects;
• unilateral effects.

The expert should disclose the assumptions underlying the prediction.

Merger simulation already requires disciplined model calibration. DOJ materials specifically emphasize that improper calibration can make merger-simulation predictions meaningless and that expert economic evidence must employ sound methods and reliably apply them to case facts.

An ML-enhanced merger model should therefore not be treated as inherently superior to a conventional structural model.

32. Special Problems in Damages Cases

ML models may estimate:

• overcharges;
• lost sales;
• lost customers;
• lost advertising revenue;
• lost transactions;
• reduced traffic.

The principal question is:

Does the model reliably reconstruct the but-for world?

An expert should conduct:

• sensitivity testing;
• alternative specifications;
• out-of-sample validation;
• benchmark comparison;
• placebo testing;
• robustness testing.

A sophisticated neural network producing a precise-looking damages number can still be unreliable if the counterfactual is poorly constructed.

33. Special Problems in Digital Advertising

Digital advertising is particularly suited to ML analysis because advertising platforms already use algorithms for:

• auctions;
• targeting;
• conversion prediction;
• bid optimization;
• ad placement.

Modern Google antitrust proceedings demonstrate the scale at which economic and technical expert evidence is being deployed in digital-platform cases, including testimony and demonstrative evidence concerning advertising markets.

The expert should therefore separate:

algorithmic efficiency

from

algorithmic exclusion.

An algorithm that improves advertising efficiency is not necessarily anticompetitive.

34. Expert Disclosure Requirements

For a serious ML antitrust case, an expert report should ideally contain:

A. Expert qualifications

• education;
• publications;
• ML experience;
• economic expertise;
• prior testimony.

B. Data description

• sources;
• dates;
• observations;
• variables;
• preprocessing.

C. Model specification

• architecture;
• algorithms;
• parameters;
• hyperparameters.

D. Training and validation

• training set;
• validation set;
• test set;
• cross-validation;
• error rates.

E. Reproducibility materials

• code;
• model version;
• computational environment;
• randomization procedures.

F. Economic methodology

• market definition;
• causal model;
• counterfactual;
• competitive mechanism.

G. Sensitivity analysis

• alternative models;
• alternative assumptions;
• alternative datasets.

H. Limitations

• uncertainty;
• missing data;
• model drift;
• extrapolation;
• potential bias.

35. The "Human Expert + Machine" Problem

The court should also distinguish between:

Expert testimony generated by an expert using an ML tool

and

an expert simply reporting what an ML system produced.

The first may satisfy Rule 702 if the expert independently understands and validates the methodology.

The second presents a greater risk of:

"machine-generated expert opinion."

An expert cannot simply outsource the analytical judgment required by Rule 702 to an opaque software system.

36. AI-Assisted Expert Testimony

If generative AI is used during expert analysis, additional questions arise:

• What model was used?
• Was the output independently verified?
• Was confidential case information submitted?
• Was hallucination testing conducted?
• Were AI-generated citations checked?
• Did the AI generate substantive conclusions?
• Can the expert explain the underlying reasoning independently?

The expert should remain responsible for the final opinion.

37. Antitrust-Specific Burden of Proof

ML evidence should also be mapped onto the relevant elements of the antitrust claim.

For example:

Section 1

ML evidence may help establish:

• communication;
• coordination;
• parallel behavior;
• economic effects.

But it does not automatically establish an agreement.

Section 2

ML evidence may help demonstrate:

• exclusionary effects;
• foreclosure;
• strategic conduct;
• pricing effects.

But it does not independently establish monopoly power or unlawful conduct.

Section 7

ML may predict:

• diversion;
• substitution;
• innovation effects;
• future competition.

But the model must be connected to the statutory substantial-lessening-of-competition inquiry.

38. A Court's Recommended Gatekeeping Sequence

A court could evaluate an ML antitrust expert in the following order:

1. Qualification
↓
2. Identify exact opinion
↓
3. Examine data
↓
4. Examine model methodology
↓
5. Examine validation/error rate
↓
6. Examine reproducibility
↓
7. Examine causal methodology
↓
8. Test counterfactual
↓
9. Test alternative explanations
↓
10. Determine antitrust fit
↓
11. Separate admissibility from weight
↓
12. Admit, limit, or exclude testimony

39. Core Legal Principles Emerging From the Case Law

Principle 1 — Technical sophistication is not reliability

A neural network is not admissible merely because it is technologically advanced.

Principle 2 — Predictive accuracy is not causation

A highly accurate predictive model may still fail to prove an antitrust causal proposition.

Principle 3 — Economic fit matters

The model must correspond to the actual market and conduct, as illustrated by Concord Boat.

Principle 4 — The expert must own the methodology

The expert cannot merely become a spokesperson for an opaque algorithm.

Principle 5 — Reproducibility is increasingly important

Another qualified expert should ordinarily be able to understand and test the analytical pathway.

Principle 6 — Alternative explanations matter

Especially in algorithmic-collusion cases, parallel predictions or prices are not necessarily proof of unlawful coordination.

Principle 7 — Black-box evidence requires additional safeguards

The less transparent the model, the greater the need for validation, robustness testing, documentation and independent verification.

Principle 8 — The legal question cannot be delegated to the model

An ML system may supply evidence; it does not decide whether conduct constitutes an antitrust violation.

40. Conclusion

Expert witness standards for machine-learning systems in antitrust cases should combine traditional Rule 702/Daubert gatekeeping with heightened scrutiny of data provenance, model validation, reproducibility, causal inference, model drift, counterfactual construction and antitrust relevance.

The foundational cases—Daubert, Joiner, Kumho Tire, and Concord Boat—establish that the court must examine not merely whether an expert uses an accepted methodology, but whether that methodology is reliably connected to the facts and conclusions of the particular dispute.

Modern digital-platform cases such as Google Play and Google Digital Advertising demonstrate the increasing importance of sophisticated economic and technical expert evidence.

The most defensible standard can therefore be expressed as:

Qualified expert + sufficient data + reliable ML methodology + validated model + reproducible analysis + reliable application + causal/economic fit + defensible counterfactual + antitrust relevance.

Ultimately, machine learning should be treated as an evidentiary methodology, not as an evidentiary shortcut. The more complex and opaque the algorithm, the more important it becomes for the expert to demonstrate what the model actually establishes, what it does not establish, and why its conclusions remain reliable under reasonable alternative assumptions.