Technical Summary: Barriers to Evidence in AI-Related Cases and the Privatization of Proof

Problem Statement

Evidence is the cornerstone of legal litigation, yet in disputes involving artificial intelligence (AI) and automated decision-making systems, obtaining admissible and reliable evidence has become increasingly difficult. The authors identify a structural barrier where decisive facts are often hidden within proprietary models, platform logs, and protected databases controlled by private actors. This creates a "Catch-22" for claimants: they cannot substantiate a request for access to evidence without first possessing supporting information, yet the producing party (the AI developer or deployer) refuses to grant access without further substantiation.

This dynamic leads to the "privatization of proof," a phenomenon where control over the evidence necessary to prove or defend a claim falls into the hands of private entities. These entities can demand justification for access while simultaneously ensuring that the justification remains out of reach due to information asymmetries. Consequently, many AI-related claims are settled or dismissed at the pre-trial stage, undermining private litigation as a viable mechanism for AI accountability.

Methodology

The authors conducted a systematic analysis of case law involving litigation against AI and automated decision-making systems. Their methodology included:

1. Case Selection: Surveying cases spanning various legal theories, including constitutional claims (due process, equal protection), federal statutory claims (disparate impact under Title VII, Fair Housing Act, copyright infringement), and state common law claims (contract, tort).
2. Scope: The research covered multiple procedural stages where evidentiary access disputes arise, such as pre-trial admissibility hearings, motions to dismiss, discovery, and summary judgment.
3. Data Sources: The authors utilized existing legal knowledge, the Law and Algorithms casebook, news reporting on ongoing litigation, and AI-assisted legal research tools (Westlaw AI-Assisted Research and PrecedentPilot) to identify interlocutory court orders and evidentiary disputes that often do not appear in final published opinions.
4. Analysis Framework: The authors coded cases to identify recurring arguments used by producing parties to resist evidence requests and categorized the sources of asymmetry between litigants.

Key Contributions

1. Identification of the "Privatization of Proof"

The paper defines the privatization of proof as a state where private actors control the evidence necessary for litigation and can demand justification for access while ensuring that justification is impossible to provide without that same access. This is exacerbated by the concentration of capital, resources, infrastructure, and expertise within private AI companies.

2. Seven Sources of Asymmetry

The authors identify seven specific dimensions of asymmetry that contribute to the privatization of proof in AI cases:

1. Model: Access to weights, activations, outputs, and API access.
2. Data: Access to pre-training, fine-tuning, human feedback, and benchmark data.
3. Documentation: Information revealing design decisions and organizational knowledge.
4. Post-deployment Information: Logs and operational traces of system behavior.
5. Expertise: Specialized knowledge required to formulate discovery requests and evaluate technical claims.
6. Compute: Resources required for red-teaming, reproduction, and systematic testing.
7. Infrastructure: Tooling that assists in development and analysis.

3. Fungibility of Access

A central theoretical contribution is the argument that access is multidimensional and that different forms of access are functionally exchangeable (fungible). Denying one type of access (e.g., full source code) does not necessarily foreclose the gathering of evidence if alternative forms of access (e.g., query access, structured evaluation sets, or logs) can provide a functionally equivalent amount of information. This observation allows for negotiated access solutions that protect sensitive artifacts while satisfying evidentiary needs.

4. A Three-Part Test for Access Disputes

Drawing on existing legal doctrines of relevance, proportionality, and necessity, the authors propose a structured three-part test for courts to evaluate disputes over AI-related evidence:

• Step 1: Degree of Asymmetry. The court evaluates whether the requesting party lacks access to relevant models, data, documentation, compute, infrastructure, expertise, or downstream information compared to the producing party. The test terminates if the requesting party can prove their claim with current access or cannot prove it even with the producing party's access.
• Step 2: Justified Access (Proportionality). The court assesses whether "privatization of proof" has occurred by comparing the benefit-to-risk ratio of the requested access against the benefit-to-risk ratio of the minimal access needed to reproduce the elements of the cause of action.
  • Benefit: The value to the requesting party.
  • Risk: The risk to the producing party (e.g., trade secret leakage, privacy violation).
  • If the requested access offers a higher benefit-to-risk ratio than the minimal access required for the cause of action, the request is favored.
• Step 3: Alternatives and Safeguards. If the request passes Step 2, the producing party may propose alternative access plans (e.g., sandboxes, phased discovery, adverse inference) that are functionally equivalent to the requested access and do not reduce the benefit-to-risk ratio established in Step 2.

Results and Case Analysis

Through the analysis of cases such as People v. Superior Ct. (Chubbs), State v. Loomis, The New York Times Co. v. Microsoft Corp., and Kadrey v. Meta, the authors observe:

• Recurring Arguments: Producing parties frequently resist access by claiming that high-level methodological descriptions are sufficient (Argument 1), that developer-generated validation studies are adequate (Argument 2), or that the AI plays only a secondary role to human discretion (Argument 3). Conversely, they invoke trade secrecy, privacy, technical infeasibility, and cost burdens to block production.
• Judicial Outcomes: Courts often deny access to source code or training data, relying on developer-provided documentation or validation studies. In many instances, this leads to settlements or dismissals before the merits of the AI system's behavior are fully examined.
• The "Catch-22" in Practice: In cases like TrueAllele and COMPAS, courts have accepted developer-generated evidence as sufficient for reliability, even when independent evaluation was impossible due to lack of access. In The New York Times v. Microsoft, technical barriers and data retention issues (e.g., deletion of search results in a sandbox) further complicated evidence gathering.

Significance

The paper argues that without addressing these barriers, AI-related decisions and outcomes may become uncontestable without special access, support, or knowledge. The significance of this work lies in:

1. Systematizing the Problem: Moving beyond ad hoc judicial responses to identify a consistent pattern of "proof privatization" driven by specific asymmetries.
2. Proposing a Framework: Offering a practical, three-part test that judges can use to resolve access disputes consistently, drawing on established legal principles rather than creating new procedural rules.
3. Enabling Accountability: By recognizing the fungibility of access, the framework suggests that courts can order evidence production that satisfies the requesting party's evidentiary needs without necessarily forcing the disclosure of the most sensitive proprietary assets, thereby balancing accountability with legitimate business and privacy interests.

The authors conclude that private litigation remains a critical route for curbing AI risks, but its viability depends on resolving the asymmetries that currently prevent the production of necessary evidence.