The Wild West of AI Ownership Is Still Taking Shape

What the latest copyright decisions mean for the human work inside AI

A company can buy an AI service in an afternoon. Explaining the rights behind its training material, retrieved sources, and generated answers can take considerably longer. That gap between adoption and accountability remains one of the most important decisions facing enterprise leaders.

This article is a sequel to The Wild West of AI Ownership and continues the Wild West of AI series. When I introduced the series in August 2025, the frontier was already crowded with questions about training data, fair use, and who gets paid. The series has followed the cases and their aftermath. Now, appellate decisions are drawing more specific boundaries, including a consequential ruling involving Westlaw headnotes and a surprisingly useful banana costume.

The law has moved. It has not reached one universal answer. For businesses, the practical task is to identify the human contribution, the particular copying, the product’s function, and the market at stake before deciding what a ruling means for their own AI architecture.

The map has changed since 2025

The early training decisions already resisted a single headline. In Bartz v. Anthropic, the court treated training as transformative but separately rejected fair use for building and retaining a general-purpose library of pirated books. In Kadrey v. Meta, Meta prevailed on the named plaintiffs’ training claim because their market-harm evidence was insufficient. That record-specific result did not establish that every training use is lawful.

The consequences became more concrete in 2026. The Bartz settlement received final approval on Jul 20, 2026: a $1.5 billion fund covering roughly 482,000 works. The release addresses specified past conduct and preserves output and future-conduct claims. A negotiated settlement measures litigation risk and resolves claims; it does not supply a general training license or an appellate fair-use rule.

September brought two different appellate developments. Doe v. GitHub concerned the pleaded removal of copyright management information from Copilot outputs. The Ninth Circuit affirmed dismissal of that theory. It did not decide training fair use, and the plaintiffs had forfeited their separate input-stage theory. Nor did it establish literal identicality as an additional statutory requirement.

Then, on Sep 29, 2026, the Third Circuit affirmed partial summary judgment in Thomson Reuters v. ROSS. That appeal addressed originality and fair use in a competing, non-generative legal-research platform. The distinction between these two appeals is where useful analysis begins.

Fair use follows the use and the market

Copyrightability is only the first question. Fair use asks whether the challenged use of protected expression is permitted under the statutory factors.

ROSS used headnote-derived training memoranda to improve a legal-search platform competing with Westlaw. The Third Circuit’s analysis looked through that intermediate training step to the highly similar commercial purpose. It also considered extensive unnecessary copying, direct platform competition, and a developing market for licensing headnotes as training material. Factors one, three, and four weighed against fair use; the factual character of the works made factor two slightly favorable to ROSS.

This extends the question explored in When Copying Doesn’t Show Up in the Output. The absence of headnotes from the final screen did not erase the copying used to build the competing service.

The market analysis also deserves its limits. ROSS’s evidence of substitution and training-data value does not establish that every proposed licensing market must be protected. Kadrey reached a different conclusion on the training-license theory presented there. Leaders should ask which market, which use, and which evidence support the claimed harm.

AI-generated conceptual illustration of examining AI uses, sources and value.

Architecture still changes the question

ROSS returned existing judicial-opinion passages in response to legal questions. Its platform did not generate original expression. Footnote 7 expressly distinguishes generative-model litigation, including the OpenAI cases and the government’s position in them.

That distinction leaves substantial work for other courts. It does not establish an automatic exemption for generative AI. A service may train on one collection, retrieve licensed or unlicensed material from another, and compose an answer that substitutes for reading the source. Acquisition, training, retrieval, generation, and distribution can raise different questions within the same product.

The Justice Department’s September 2026 statement of interest argues for fair use of copyrighted text in LLM training while distinguishing other stages. That filing is government advocacy, not a court’s resolution of the pending generative-AI disputes. Policy preferences and legal holdings have different jobs.

AI copyright law is moving from categorical arguments about “training” toward a stage-by-stage inquiry into source, purpose, output, market, and proof.

The argument map below sets out the competing positions and turns each issue into a practical action for general counsel. The positions are arguments; the actions are an analytical checklist rather than judicial holdings.

Whose AI is it anyway?

These distinctions also clarify the ownership question. A business can own a model or contract for a service while relying on rights and contributions held by others.

What enterprise leaders should do now

The original series asked where boundaries would come from. The emerging answer is a combination of adjudicated facts, negotiated agreements, product design, and evidence. Enterprise governance needs to work at that same level of detail.

• Map the material. Record sources, acquisition methods, licenses, restrictions, and the human editorial contributions embedded in datasets and knowledge collections
• Separate the operations. Document what is copied for training, retained for retrieval, shown to users, or redistributed. A provider’s general “AI” description is insufficient
• Examine substitution. Test whether the product replaces subscriptions, source visits, research services, or other uses that customers otherwise purchase. Preserve representative evidence
• Contract around the actual use. Check training and retrieval permissions, attribution requirements, indemnity scope, exclusions, and who can respond when a rights holder objects
• Keep review current. Assign an owner for legal developments and model or dataset changes. A fair-use assessment tied to one architecture and record should not silently travel to another

The frontier is becoming more legible. Its boundaries still depend on what people created, what developers copied, and what the resulting product does. Recognizing that human work early gives leaders a stronger basis for decisions than waiting for a universal verdict on “AI.”

About the author

Lili Kazemi is General Counsel and AI Policy Leader at Anant Corporation. She writes about AI governance, copyright, and the role of human judgment in emerging technology, and is the creator of The Human Edge of AI.

About Anant

Anant builds AI capability for businesses through platform architecture, data engineering, managed operations, governance, and workforce enablement.

For more analysis of AI, law, and human judgment, read The Human Edge of AI.

📬 Subscribe to The Human Edge of AI on Substack

📬 Subscribe to The Human Edge of AI on LinkedIn

This article is for general informational purposes and does not constitute legal advice. It reflects the author’s independent analysis and is not an endorsement by any organization cited. Consult qualified counsel about specific facts and uses.