Direct Answer: AI copyright fair use landmark cases in 2025 and 2026—most notably Bartz v. Anthropic and Thomson Reuters v. Ross Intelligence—have dismantled the assumption that all machine learning ingestion is categorically protected fair use. Federal courts have established an essential bifurcated doctrine: training on lawfully acquired expressive works to extract latent statistical patterns can be highly transformative, but acquiring data through illicit repositories or using proprietary domain taxonomies to build direct commercial substitutes constitutes actionable copyright infringement.
For general counsel, technology executives, and enterprise builders deploying agentic AI or specialized domain models, intellectual property risk is no longer theoretical. The legal boundary has shifted from abstract model weights to data supply chain provenance and commercial market substitution.
The Judicial Fault Lines: Transformative Learning vs. Direct Market Substitution
Since the emergence of generative models, developers frequently cited Authors Guild v. Google, Inc. (the landmark Google Books decision) to argue that mass digital copying for non-expressive computational analysis qualifies as fair use under 17 U.S.C. § 107. However, recent federal rulings demonstrate that judges apply a rigorous, fact-intensive analysis to AI training across all four statutory fair use factors.
1. Lawful Acquisition vs. Illicit Provenance: Bartz v. Anthropic
In Bartz v. Anthropic (N.D. Cal. June 2025), the court drew an explicit legal boundary between the internal mechanics of neural network training and the real-world acquisition of source material:
- Transformative Training on Lawful Copies: The court held that training an AI system on books lawfully acquired through authorized commercial channels is "highly transformative." The model ingests text not to republish stories, but to teach an algorithm linguistic relationships and conceptual reasoning.
- Actionable Infringement from Unlawful Datasets: Crucially, the court denied fair use protection for the acquisition and retention of pirated book corpora (such as shadow library datasets). The decision establishes that an organization cannot launder copyright infringement through the subsequent training of a transformative model. The illegal acquisition of raw data remains an independent, actionable basis for liability.
2. Commercial Competition and Domain Taxonomies: Thomson Reuters v. Ross Intelligence
While Bartz examined foundational models trained on broad literary works, Thomson Reuters v. Ross Intelligence addressed specialized enterprise AI trained on proprietary domain assets. Thomson Reuters alleged that legal AI startup Ross Intelligence unlawfully ingested over 2,000 Westlaw headnotes and the proprietary West Key Number classification system to train a competing legal research platform.
In February 2025, Judge Stephanos Bibas of the U.S. District Court for the District of Delaware granted partial summary judgment to Thomson Reuters, rejecting Ross's fair use defense as a matter of law. In late 2026, the Third Circuit affirmed the core liability ruling:
- Market Substitution Trumps Computational Transformation: Under the fourth statutory factor (the effect on the potential market for the copyrighted work), the court emphasized that Ross used Westlaw's curated editorial summaries to develop a direct market substitute targeting Thomson Reuters's core commercial audience.
- Editorial Taxonomies Are Protected Expressions: Westlaw's selection, arrangement, and drafting of point-of-law headnotes contain sufficient human creativity to enjoy robust copyright protection against wholesale automated harvesting.
Comparison of Landmark AI Copyright Rulings
| Judicial Decision | Jurisdiction & Forum | Primary Data Subject | Key Fair Use Holding | Enterprise Risk Implication |
|---|---|---|---|---|
| Bartz v. Anthropic | N.D. Cal. (Federal) | Published commercial books and literary works | Transformative for lawfully acquired texts; No fair use for pirated corpora | Supply chain provenance audits are mandatory; training does not cure illicit data sources. |
| Thomson Reuters v. Ross Intelligence | D. Del. / 3rd Cir. (Federal) | Proprietary legal headnotes and Key Number taxonomy | Fair use rejected as a matter of law; commercial product created direct market substitution | Ingesting competitor editorial structures or taxonomies to train vertical tools carries severe exposure. |
| Authors Guild v. Google (Foundational) | 2nd Cir. (Federal) | In-copyright print books digitized for search snippet index | Fair use upheld; snippets created public search utility without displacing book sales | Search indices that direct users to primary sources remain protected; direct answer generators face higher scrutiny. |
Practical Blueprint: Enterprise IP Due Diligence in the Agentic Era
Organizations implementing internal AI tools, deploying autonomous workflows, or fine-tuning models on domain data should establish clear compliance controls:
- Audit Vendor Data Provenance: Require commercial foundation model vendors to provide contractual representations and indemnities confirming training datasets were assembled from licensed, public domain, or lawfully acquired sources.
- Respect Competitor Taxonomies in Vertical Agents: When developing domain-specific agents (in legal, accounting, healthcare, or insurance), ingest public primary records (such as court opinions, public dockets, or statutory codes) rather than proprietary secondary editorial digests or competitor database schemas.
- Implement RAG Output Safeguards: Ensure retrieval-augmented generation systems do not regurgitate lengthy verbatim excerpts of licensed materials. Output moderation guardrails should restrict verbatim quoting to fair quotation norms.
- Maintain Segregated Training Ledgers: Keep timestamped, auditable records of all internal corpora ingested for fine-tuning or embedding generation, verifying corporate ownership or commercial enterprise licenses.
To explore how high-reliability agent architectures maintain auditable execution and state verification in production environments, read our analysis on Closed-Loop Verification for Production AI Agents.
Frequently Asked Questions (FAQ)
Can an enterprise be held liable if an AI model recites copyrighted text?
Yes. If an AI system generates substantial verbatim excerpts of copyrighted works that compete with or displace the original expression, both the platform operator and downstream commercial users may face claims of direct or contributory copyright infringement.
Why did fair use fail in Thomson Reuters v. Ross Intelligence?
Fair use was rejected primarily because Ross used Westlaw's copyrighted editorial headnotes to build a legal research product that directly competed for Thomson Reuters's paying customers, causing demonstrable market substitution rather than creating an orthogonal, public-benefit search utility.
How does lawful data acquisition affect fair use in AI training?
Under Bartz v. Anthropic, the method of data acquisition is legally distinct from model training. Even if training is deemed transformative, acquiring and storing pirated or unauthorized copies constitutes actionable copyright infringement that fair use will not excuse.