Evidence Disclosure Standards for Proprietary Algorithms

1. Introduction

Evidence disclosure standards for proprietary algorithms concern the extent to which a party relying upon an algorithm in litigation, regulatory proceedings, competition investigations, or administrative decision-making must disclose information about that algorithm to opposing parties or adjudicators.

The central tension is between:

• procedural fairness and effective judicial review, which may require disclosure sufficient to test the algorithm's operation; and
• trade-secret, confidentiality, intellectual-property, cybersecurity, and commercial interests, which may justify limiting disclosure of source code or technical architecture.

The modern legal position is generally not that proprietary algorithms automatically receive absolute confidentiality. Rather, courts and regulators increasingly distinguish between disclosure of the existence and effects of an algorithm, disclosure of its methodology and inputs, and disclosure of the underlying source code.

2. Meaning of a Proprietary Algorithm

A proprietary algorithm may include:

1. source code;
2. machine-learning models;
3. model weights;
4. decision trees;
5. pricing algorithms;
6. ranking algorithms;
7. recommendation systems;
8. fraud-detection systems;
9. credit-scoring models;
10. automated risk-assessment tools;
11. demand-prediction systems;
12. algorithmic trading systems; and
13. proprietary datasets and feature-engineering processes.

In litigation, the relevant question is usually not whether the algorithm is commercially confidential, but:

What information must be disclosed so that the opposing party can meaningfully challenge the decision or evidence generated by the algorithm?

3. Why Algorithmic Evidence Creates a Disclosure Problem

Traditional documentary evidence is comparatively transparent. An algorithm may instead operate as a black box.

For example:

Input data → preprocessing → algorithm/model → weighting → prediction → automated decision → economic consequence

A party may produce only the final output:

"The algorithm determined that the defendant's conduct caused a 17% price increase."

That output may be insufficient if the opposing party cannot determine:

• what data were used;
• whether the data were complete;
• what assumptions were made;
• which variables were weighted;
• whether the model was trained on biased data;
• whether the algorithm changed during the relevant period;
• whether human intervention occurred;
• whether the model was validated;
• whether alternative models produce different results; and
• whether the result can be reproduced.

Consequently, algorithmic disclosure is increasingly treated as an evidentiary reliability problem rather than merely an intellectual-property problem.

4. Different Levels of Algorithmic Disclosure

A useful legal framework is to divide disclosure into five levels.

The law should not automatically require Level 5 disclosure in every case.

Instead, disclosure should normally be proportionate to the algorithm's importance to the dispute.

5. Core Legal Principles

A. Relevance

Algorithmic material must be disclosed where it is relevant to a disputed issue.

If an algorithm is merely administrative and has no bearing on liability, complete technical disclosure may be unnecessary.

Where the algorithm is the primary basis for liability, however, substantially greater disclosure may be required.

B. Necessity for Effective Challenge

The most important principle is whether the opposing party can effectively challenge the evidence without access to the underlying algorithmic information.

For example, disclosure of:

"The model calculated damages of $50 million"

may be inadequate where the opposing party cannot determine:

• the variables used;
• the counterfactual;
• the training period;
• the assumptions;
• the model specification; or
• the calculation methodology.

C. Proportionality

Courts generally balance:

need for disclosure ↔ confidentiality burden

A court may therefore order:

• limited disclosure;
• disclosure to experts;
• confidentiality undertakings;
• protective orders;
• inspection rather than copying;
• source-code escrow;
• redaction;
• secure-room examination; or
• disclosure of outputs and methodology without source code.

6. Trade Secrets Do Not Automatically Defeat Disclosure

Proprietary status is relevant, but it is not necessarily decisive.

A company cannot ordinarily transform evidence into an evidentiary privilege merely by describing its algorithm as:

"commercially sensitive."

The court must ask whether confidentiality can be protected while still preserving procedural fairness.

This is particularly important where the algorithm is being used to establish:

• damages;
• market power;
• discrimination;
• cartel effects;
• consumer harm;
• creditworthiness;
• fraud;
• risk;
• regulatory violations; or
• causation.

7. Case Law

1. State v. Loomis — Wisconsin Supreme Court

State v. Loomis, 881 N.W.2d 749 (Wis. 2016) is one of the leading cases concerning proprietary algorithmic decision-making.

The case involved the use of the COMPAS risk-assessment system during criminal sentencing.

The defendant argued that reliance upon a proprietary algorithm raised due-process concerns because the methodology was protected as a trade secret.

The Wisconsin Supreme Court permitted consideration of COMPAS but emphasized important limitations.

Significance

The case illustrates that a proprietary algorithm cannot simply be treated as an unquestionable source of truth.

The court was concerned about:

• proprietary methodology;
• inability of the defendant to examine the complete algorithm;
• potential statistical limitations;
• accuracy concerns; and
• the possibility that the assessment could influence sentencing.

Principle

A proprietary algorithm may be used in decision-making, but procedural safeguards may be necessary where the affected person cannot independently scrutinize its methodology.

The case is particularly important for modern AI because it demonstrates the tension between trade-secret protection and due process.

8. 2. United States v. Microsoft Corp. — Algorithmic and Proprietary Evidence Context

United States v. Microsoft Corp., 253 F.3d 34 (D.C. Cir. 2001) is not a pure source-code disclosure case, but it is highly relevant to proprietary technology evidence.

The litigation required examination of Microsoft's proprietary technological practices and software architecture in determining whether conduct constituted exclusionary behavior.

Importance

The case demonstrates that proprietary technology cannot be evaluated solely from the defendant's characterization of its technical system.

Competition analysis may require examination of:

• technological design;
• interoperability;
• technical restrictions;
• product architecture;
• strategic technological decisions.

Principle

Where proprietary technology is central to proving anticompetitive conduct, commercial confidentiality does not eliminate the need for meaningful evidentiary scrutiny.

9. 3. FTC v. Amazon — Algorithmic Evidence and Digital Competition

Digital-platform litigation has increasingly required examination of proprietary algorithmic systems.

In Federal Trade Commission v. Amazon.com, Inc., algorithmic evidence concerning marketplace practices, pricing, advertising and platform design became relevant to the assessment of alleged exclusionary conduct.

Although the case does not establish a simple universal rule requiring source-code production, it illustrates the broader evidentiary problem:

Digital platforms frequently possess information about algorithms that opposing parties cannot independently reproduce.

Significance

Where algorithmic conduct is alleged to affect:

• ranking;
• pricing;
• seller visibility;
• advertising;
• consumer choice; or
• competitive access,

disclosure may need to extend beyond the final output to technical and operational evidence explaining how the system functioned.

10. 4. FTC v. Meta Platforms

The litigation involving FTC v. Meta Platforms, Inc. demonstrates another important dimension of proprietary algorithmic evidence.

Large digital platforms often possess extensive internal information concerning:

• recommendation systems;
• user engagement;
• data collection;
• product development;
• interoperability;
• acquisition strategy; and
• internal experimentation.

Significance

The evidentiary importance of algorithmic material may arise even where the algorithm itself is not the formal subject of the dispute.

Internal documents concerning how algorithms were designed and modified can reveal:

• strategic intent;
• competitive effects;
• exclusionary mechanisms;
• product substitutability; and
• consumer behavior.

Principle

Courts may therefore distinguish between disclosure of source code and disclosure of business records describing algorithmic operation and strategic use.

The latter may be considerably easier to justify.

11. 5. Sorrell v. IMS Health Inc.

Sorrell v. IMS Health Inc., 564 U.S. 552 (2011) concerned the use of prescriber information and data analytics by pharmaceutical companies.

Although it was principally a First Amendment and data-use case rather than a source-code-disclosure dispute, it demonstrates the legal significance of commercially generated data and analytical systems.

Relevance to algorithmic evidence

Modern proprietary algorithms frequently depend upon large datasets.

Therefore, disclosure disputes may concern not only:

"What is the algorithm?"

but also:

"What data did the algorithm process?"

An algorithm can be mathematically transparent yet produce unreliable conclusions if the underlying dataset is:

• incomplete;
• selectively constructed;
• biased;
• outdated;
• improperly labelled; or
• contaminated.

Principle

Algorithmic transparency without data provenance may provide only partial evidentiary transparency.

12. 6. Carpenter v. United States

Carpenter v. United States, 585 U.S. 296 (2018) concerned government access to historical cell-site location information.

The case is important to algorithmic evidence because modern surveillance and analytical systems can transform large datasets into highly revealing conclusions about individuals.

Significance

The Court recognized that technological systems can generate information about individuals that would have been practically impossible to obtain through traditional methods.

This supports a broader principle:

The legal significance of automated analysis cannot be determined merely by examining the raw data; the method of processing and aggregation can materially affect the legal consequence.

Thus, where an algorithm transforms apparently innocuous data into legally significant conclusions, disclosure of the analytical process may become important.

13. 7. State ex rel. Cincinnati Enquirer v. Hamilton County

Public-records litigation in the United States has repeatedly confronted the question whether proprietary software and algorithmic systems can be withheld because they contain trade secrets.

Cases involving governmental use of proprietary software demonstrate the importance of distinguishing:

• source code;
• system functionality;
• underlying records;
• decision criteria; and
• vendor-confidential information.

Principle

A government agency cannot necessarily avoid transparency obligations merely because the government procured the relevant technological system from a private vendor.

This becomes particularly significant where automated systems are used for:

• public benefits;
• policing;
• risk assessments;
• procurement;
• taxation; or
• regulatory enforcement.

14. 8. R (Bridges) v Chief Constable of South Wales Police

R (Bridges) v Chief Constable of South Wales Police [2020] EWCA Civ 1058 is particularly important from a UK perspective.

The case concerned the police use of automated facial recognition technology.

The Court of Appeal considered the legality of automated facial recognition and the adequacy of the legal framework governing its deployment.

Importance for proprietary algorithms

The case demonstrates that automated decision systems used by public authorities must be assessed against:

• legal authority;
• proportionality;
• safeguards;
• human discretion;
• equality considerations; and
• accountability.

The case therefore supports a broader proposition:

Where an algorithm materially affects legal rights or public power, opacity cannot automatically defeat legal scrutiny.

15. 9. R (Miller) v College of Policing

UK judicial review litigation concerning automated or data-driven policing also illustrates the importance of understanding the operational criteria embedded in technological systems.

Where algorithmic tools influence policing decisions, relevant disclosure may concern:

• decision criteria;
• thresholds;
• datasets;
• risk classifications;
• human intervention;
• error rates; and
• governance procedures.

The legal issue is therefore often functional transparency rather than unrestricted source-code disclosure.

16. 10. Google LLC v Competition and Markets Authority / Digital Markets Context

UK and EU digital-competition proceedings increasingly demonstrate the evidentiary significance of proprietary algorithms.

Search ranking, advertising, recommendation and platform algorithms can determine:

• visibility;
• market access;
• traffic allocation;
• advertising prices;
• consumer choice; and
• competitive conditions.

The emerging approach is generally to seek enough information to evaluate competitive effects without necessarily exposing every element of commercially sensitive source code.

17. Algorithmic Evidence in Competition Law

This issue is particularly important in competition-law litigation.

A dominant digital undertaking may argue:

"Our ranking algorithm is proprietary."

But the competition authority may need to establish:

1. how rankings are generated;
2. whether rivals are disadvantaged;
3. whether self-preferencing exists;
4. whether access is selectively restricted;
5. whether algorithmic changes alter market conditions;
6. whether prices are algorithmically coordinated; and
7. whether data advantages reinforce market power.

Accordingly, algorithmic secrecy cannot become a substitute for evidentiary accountability.

18. Algorithmic Pricing and Cartel Cases

Algorithmic pricing creates an especially difficult disclosure problem.

Suppose competitors use automated pricing systems.

The authority may need to establish whether:

A. each algorithm independently responds to market conditions,

or

B. the algorithms effectively implement coordinated pricing.

Relevant evidence may include:

• source-code architecture;
• pricing rules;
• historical outputs;
• model-training data;
• communication between firms;
• parameter changes;
• API interactions;
• logs;
• timestamps;
• override records; and
• algorithm-update histories.

However, complete source-code disclosure may not always be necessary.

Logs + inputs + outputs + methodology + expert examination may sometimes permit adequate testing.

19. AI and Machine-Learning Models

Machine-learning systems complicate traditional disclosure because the "algorithm" may not be a fixed sequence of human-written rules.

The relevant evidence may instead include:

• training dataset;
• model architecture;
• model weights;
• hyperparameters;
• feature engineering;
• training methodology;
• validation dataset;
• test results;
• model version;
• model drift;
• post-deployment updates; and
• human interventions.

Consequently, the phrase "disclose the algorithm" may be legally inadequate.

A proper disclosure order should specify which technical artefacts are required.

20. Source Code Versus Explainability

A crucial distinction is:

Source-code disclosure ≠ meaningful transparency.

A lawyer may receive 500,000 lines of source code and still be unable to determine why a decision occurred.

Therefore, courts should consider whether disclosure permits:

• independent testing;
• reproducibility;
• causal analysis;
• error identification;
• bias testing;
• statistical validation; and
• meaningful cross-examination.

In some cases, an expert-access model is superior to unrestricted source-code production.

21. Protective Orders

Courts can reconcile confidentiality with disclosure through protective mechanisms.

Common safeguards

1. Attorney's-eyes-only disclosure
2. Expert-only access
3. Confidentiality undertakings
4. Secure data rooms
5. Source-code inspection without copying
6. Redaction
7. Limited-purpose disclosure
8. Prohibition on commercial use
9. Sealed filings
10. Court-appointed technical experts

These mechanisms support a least-restrictive disclosure principle.

22. What Should Be Disclosed?

A proportional disclosure order may require:

A. Algorithm identity

• name;
• version;
• developer;
• deployment date.

B. Functional description

• purpose;
• inputs;
• outputs;
• decision thresholds.

C. Data

• relevant datasets;
• provenance;
• collection period;
• sampling methodology.

D. Technical information

• architecture;
• variables;
• parameters;
• model version;
• relevant code.

E. Validation

• accuracy;
• error rates;
• testing methodology;
• validation results.

F. Operational evidence

• logs;
• timestamps;
• audit trails;
• model changes;
• human overrides.

G. Reproducibility material

• data sufficient to reproduce the relevant result;
• calculation scripts;
• model version;
• relevant configuration files.

23. What May Properly Remain Confidential?

Courts may legitimately protect:

• unrelated source code;
• cybersecurity credentials;
• encryption keys;
• commercially irrelevant architecture;
• unrelated training data;
• third-party confidential information;
• security vulnerabilities;
• sensitive customer information;
• trade-secret material unrelated to the dispute.

The important point is that confidentiality should ordinarily be tailored to the evidentiary need.

24. Burden of Proof

A useful approach is:

Step 1 — Party relying on algorithm

The party should establish:

• what the algorithm does;
• why it is relevant;
• its basic methodology;
• its reliability;
• the relevant version.

Step 2 — Opposing party

The opponent identifies specific grounds for challenging:

• reliability;
• causation;
• methodology;
• bias;
• data quality;
• reproducibility.

Step 3 — Court

The court determines the minimum disclosure necessary for meaningful scrutiny.

Step 4 — Confidentiality

The court determines what protective mechanisms are appropriate.

25. Consequences of Non-Disclosure

Where an algorithm is material and disclosure is unjustifiably withheld, possible consequences include:

• exclusion of algorithmic evidence;
• adverse inference;
• inability to rely on particular calculations;
• sanctions;
• limitation of expert testimony;
• reopening of evidence;
• adverse procedural orders;
• judicial review of the decision;
• damages methodology being rejected.

The precise remedy depends upon the jurisdiction and procedural context.

26. Proposed Legal Test

A useful six-part proportionality test for proprietary algorithm disclosure is:

1. Materiality

Is the algorithm material to liability, causation, damages or jurisdiction?

2. Necessity

Can the opposing party meaningfully challenge the evidence without disclosure?

3. Reliability

What evidence establishes that the algorithm produces reliable results?

4. Confidentiality

What genuine trade-secret or security interests are threatened?

5. Alternatives

Can equivalent scrutiny be achieved through less intrusive disclosure?

6. Protection

Can a protective order adequately protect the proprietary interests?

This can be expressed as:

Materiality → Necessity → Reliability → Confidentiality → Alternatives → Protective Order

27. Special Rule for Government Algorithms

Where the algorithm is used by a public authority, the disclosure threshold may be higher.

The public body should generally be able to explain:

• statutory authority;
• purpose;
• decision criteria;
• data sources;
• safeguards;
• human oversight;
• error mechanisms;
• review procedures.

A private vendor's trade-secret claim should not automatically prevent a citizen from challenging a governmental decision.

This is particularly important in:

• welfare allocation;
• immigration;
• policing;
• taxation;
• public procurement;
• education;
• criminal justice; and
• regulatory enforcement.

28. Special Rule for Competition Authorities

Competition authorities should have enhanced powers to obtain:

• algorithmic audit trails;
• historical model versions;
• pricing logs;
• ranking changes;
• internal testing;
• model-development documents;
• communications concerning algorithmic objectives;
• datasets used to make competitive decisions.

However, authorities should also maintain strict confidentiality procedures because algorithmic disclosure can expose extremely valuable intellectual property.

29. Key Case-Law Principles — Summary

30. Conclusion

Evidence disclosure standards for proprietary algorithms should not be based upon an absolute choice between secrecy and complete transparency.

The preferable approach is proportionate algorithmic disclosure.

The central legal question should be:

What information is necessary to permit meaningful scrutiny of an algorithmically generated conclusion while protecting legitimate proprietary interests?

Where the algorithm is peripheral, limited disclosure may be sufficient.

Where it is central to liability, damages, market power, discrimination, or governmental decision-making, disclosure should ordinarily extend to the methodology, relevant inputs, outputs, validation material, version history, audit logs and other information necessary for independent testing.

Source code should be ordered where it is genuinely necessary, but courts should recognize that source-code access is not synonymous with explainability.

The emerging principle can therefore be stated as:

Proprietary status protects confidentiality; it does not create evidentiary immunity.

For AI and digital competition disputes, the strongest legal model is consequently risk-based, purpose-specific and proportional disclosure supported by confidentiality protections and independent expert scrutiny.