It’s Bananas: The Third Circuit in ROSS Views Headnotes as ‘Creative Spark’
The Third Circuit’s ROSS AI copyright ruling explains the creative spark in legal headnotes, the banana-costume analogy, and why this training use was not fair
What the Third Circuit’s ROSS AI copyright opinion reveals about the creative spark
AI copyright is reaching the federal courts of appeals, and the lines are becoming more specific. In September, the Ninth Circuit decided a narrow dispute about copyright management information in Doe v. GitHub. Thirteen days later, the Third Circuit affirmed a partial-summary-judgment ruling that original Westlaw headnotes were copyrightable and that ROSS’s use of them to train a competing legal-research platform was not fair use. These are consequential appellate decisions, but they answer different questions.
The Third Circuit also managed to explain part of its reasoning through a banana costume. That may sound like a detour. It is actually a useful way to understand what copyright protects when an AI developer says it is learning from information that belongs to everyone.
In my earlier article, When Copying Doesn’t Show Up in the Output: The Legal Doctrine Behind AI Training, I explored why the absence of copyrighted material from a product’s output does not end the legal inquiry. ROSS now brings another question into sharper focus: what, precisely, was original in the material used to build that product?
The creative spark comes before fair use
A judicial opinion is public legal material. A privately written headnote is an editor’s account of a particular point within that opinion. The difference matters before a court ever reaches fair use.
Writing a useful headnote requires choices. Which legal proposition deserves its own summary? Which facts are necessary to understand it? How much reasoning must survive the compression? How can the editor make the paragraph accurate and intelligible without requiring the reader to reconstruct the entire case?
That work can have a creative spark even though it is constrained by accuracy. Creativity in legal writing does not mean inventing the law. It can reside in selecting, framing, and expressing what the law says.
The Third Circuit found that spark in the headnotes before it. It rejected ROSS’s characterization that protecting them would give Thomson Reuters ownership of the law: the judicial opinions remained available for everyone to publish. Copyright protected the editors’ original contribution. The standard was minimal originality, not literary brilliance or the number of hours spent drafting.
The holding also has a limit worth keeping in the sentence: this appeal involved 2,243 headnotes that did not copy judicial opinions verbatim. Footnote 5 expressly leaves the copyrightability of verbatim headnotes unanswered. A broader statement that every headnote is protected would outrun the decision.
Where the banana comes in
The merger doctrine prevents copyright from swallowing an idea when that idea can be expressed in only one or a few ways. If protecting the expression would effectively monopolize the underlying idea, the two have merged for copyright purposes.
ROSS argued that this principle defeated protection for the headnotes. The Third Circuit disagreed because legal points can be expressed in different ways. To explain why, it drew on Silvertop Associates Inc. v. Kangaroo Manufacturing Inc., its decision of Aug 1, 2019 concerning a full-body banana costume. The banana facts belong to that earlier dispute; ROSS draws on its reasoning as precedent.
Silvertop held that the costume’s particular combination of non-utilitarian artistic features was copyrightable, without giving anyone ownership of the banana concept. Designers still had many ways to vary curvature, proportions, color, texture, and other features, so merger did not defeat protection. ROSS applied the same logic to headnotes: protecting an editor’s original wording leaves others free to explain the same public-domain law differently.
There is a doctrinal distinction inside the joke. Silvertop also addressed which artistic features could be separated from a useful article, the wearable costume. That is a separate copyright question. ROSS borrows the case’s merger reasoning; it does not classify headnotes as useful articles or transplant the entire costume analysis into legal publishing.
The point is not that every way of summarizing the same rule will be original, or that all similar summaries infringe. It is that accuracy does not eliminate every meaningful expressive choice. Merger cannot do all the work merely because two writers must remain faithful to the same underlying law.
The creative spark inside the data
A judicial opinion is public legal material. A privately written headnote captures a particular legal point within it. The editor decides what matters, which facts provide necessary context, and how to compress the proposition accurately enough to stand alone.
Those decisions can carry original expression. The ROSS court found the required creative spark in 2,243 nonverbatim headnotes. The threshold was minimal originality, rather than literary brilliance or hours of effort. Footnote 5 expressly leaves the copyrightability of verbatim headnotes unresolved. Protecting the editors’ contribution does not give Thomson Reuters ownership of the underlying law.
The evidence mattered. ROSS’s expert compared training-memo questions with headnotes and underlying opinions. A subset closely resembled the headnotes while differing significantly from the opinions. After independent comparisons, the district judge found copying of the protected editorial contribution. The appellate account does not turn an expert’s similarity score into an admission of every infringement element.
This is the human contribution that the word “data” can obscure. A dataset may contain measurements and public facts alongside someone’s selection, explanation, organization, and phrasing. Treating all of it as one interchangeable input makes the rights analysis less accurate and the product’s dependencies harder to see.
AI-generated conceptual illustration of different expressions of one idea.
Why ROSS’s own expert evidence mattered
The evidence makes this distinction unusually concrete. ROSS’s expert, Barbara Frederiksen-Cross, compared the training-memo questions with Westlaw headnotes and compared the headnotes with the underlying judicial opinions.
That comparison had a sensible defense function. If a memo question and a headnote resembled each other because both drew on the same public-domain opinion, resemblance alone would not show appropriation of Westlaw’s original expression. But a damaging subset pointed the other way: questions closely resembled the headnotes while the headnotes differed materially from the opinions. That pattern supported an inference that the questions had taken the editorial contribution.
This was not an expert confessing that her client had infringed copyright. In 2023, Judge Bibas rejected treating similarity scores alone as an admission of legal substantial similarity. In 2025, he treated the adverse comparisons as evidence of actual copying, then made his own side-by-side comparisons and separately found actual copying and substantial similarity for 2,243 headnotes. The distinction between an evidentiary finding and a legal conclusion is part of the story.
Fair use asks a different question
Recognizing copyrightable expression does not decide whether someone else may use it fairly. Fair use is a separate inquiry into the use, evaluated through the statutory factors. The Third Circuit weighed the first, third, and fourth factors against fair use, while the second factor slightly favored it.
For ROSS, purpose and market effects were central. Thomson Reuters used headnotes to help people find responsive legal material and to train its own search products. ROSS used the copied material to train a commercial legal-search platform aimed at the same task. The Third Circuit regarded that purpose as, at best, minimally transformative. A change in the technological process did not establish a sufficiently different use.
Nor did the absence of headnotes in the user-facing output resolve market harm. Headnotes can have commercial value as part of a research subscription even without a standalone market in which customers buy individual headnotes. The court considered their value in attracting Westlaw users, competition in legal research, and evidence of a developing market for licensing headnotes as AI training data.
That reasoning should not become a circular rule that every unlicensed use is harmful because a license could have been sold. Here, the court identified a highly similar commercial purpose, evidence of direct substitution, and a developing licensing market. Those features explain the result and constrain the analogy to other systems.
Retrieval is not the same as generative RAG
ROSS answered legal questions by returning passages from existing judicial opinions. The Third Circuit expressly described it as non-generative. Calling this a RAG case in today’s retrieval-augmented-generation sense would blur the architecture: retrieving source passages is not itself generating an answer grounded in retrieved material.
Footnote 7 makes the distinction explicit when discussing generative-AI litigation. The court contrasted models capable of generating original responses with ROSS’s platform and emphasized ROSS’s intended commercial substitution. That limits what this decision settles. It does not give generative systems an automatic fair-use exemption, and it does not decide every use of copyrighted material in a retrieval pipeline.
What leaders should ask before the next training run
For companies developing or procuring AI, the practical work begins with the source and the use. Identify which material is public and which layers reflect someone else’s original contribution. Examine what vendors actually supplied, including questions, labels, summaries, and annotations. Trace how each layer enters training, retrieval, and outputs rather than assuming the final interface tells the whole story.
Then test the business explanation. What different purpose does the use serve? What evidence supports that account? Is the product helping users do something distinct, or using a competitor’s editorial expression to perform substantially the same commercial function? Those are legal and architectural questions that should be answered together.
A banana costume is an unexpectedly good reminder that familiar subject matter can leave room for original expression. The legal rules remain available to everyone. The human decisions that make those rules easier to find and understand can still matter to copyright. An AI strategy should know which of those things it is taking.
Sources and further reading
ROSS appellate opinion, p. 12. Merger doctrine and the banana-costume analogy.
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.