This resource is educational and does not provide legal advice. Authorities and technical standards should be verified for the governing jurisdiction and current matter.
AI copyright disputes are increasingly technical evidence disputes. A useful record connects the source material, the acquisition path, the corpus, training or fine-tuning process, system behavior, alleged outputs, relevant markets, and claimed damages. Broad statements about whether AI training is transformative rarely resolve those questions on their own.
Two 2025 opinions, different records
In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., the District of Delaware granted significant partial summary judgment to Thomson Reuters on direct infringement and fair use issues involving Westlaw headnotes used to train a competing legal-research product. The official opinion addresses the route through which the training material was obtained, comparisons between the materials, and expert evidence bearing on copying.1
In Kadrey v. Meta Platforms, Inc., the Northern District of California granted Meta partial summary judgment on the record before it. The court emphasized that the result did not establish a general rule that use of copyrighted material to train language models is lawful. The order discussed the plaintiffs’ market-harm proof and the limits of the evidence offered in that case.2
Provenance is a technical question
Training-data provenance asks how material entered a system’s development path. The inquiry may involve licenses, public datasets, web collection, repositories, vendors, internal data, preprocessing, deduplication, filtering, or later fine-tuning. The relevant answer is often distributed across technical, contractual, and operational records.
An expert may help reconstruct that path, identify gaps, compare claimed source material with a corpus or dataset, and explain the significance of technical artifacts. The expert’s role does not determine ownership or fair use as a legal matter. It supplies the factual analysis on which those issues may depend.
Evidence categories often worth preserving
Data-acquisition records, corpus manifests, preprocessing documentation, model and version history, fine-tuning materials, output testing, licensing records, market evidence, and damages assumptions can each matter depending on the claim.
Output evidence requires disciplined testing
Output allegations can raise different questions from training-data allegations. A claimed similarity may require testing across prompts, model versions, temperature or sampling settings, repetition, retrieval components, context supplied to the system, and the degree to which a result is reproducible. Screenshots alone rarely answer all of those questions.
A sound testing plan states what was run, when it was run, under what conditions, and how the resulting output was compared. It also identifies alternative explanations, including user-supplied context, retrieval sources, memorization claims, and ordinary similarity in subject matter or style.
Market effects and damages depend on the factual record
The 2025 opinions reinforce that market analysis must be connected to evidence. That may include licensing practice, substitutability, the nature of the relevant work, output capability, actual market behavior, and the economic assumptions used to measure claimed harm. A technical expert and an economic expert may address different parts of the inquiry.
Attorneys should therefore decide early whether the matter requires expertise in data systems, model behavior, digital forensics, copyright comparison, licensing practice, or damages. The right scope depends on the pleadings, the available record, and the factual theory being tested.
AI Expert Witness Services helps attorneys identify experts for training-data provenance, model-output analysis, digital evidence, technical discovery, and AI intellectual-property disputes. The retained expert provides independent professional analysis.
Start matter triage