KOHEN AVOCATS · PARIS, FRANCE
Fable 5.2 Launch: What Should Creators Check About U.S. Copyright and AI Training?

Hassan Kohen is a Paris-based lawyer admitted to the Paris Bar (Avocat au Barreau de Paris). Initial strategic analysis: €80 incl. VAT, personal reply within 24 hours.

100% confidential · Legal professional privilege · No obligation

Article generated by artificial intelligence, through a process designed and supervised by the firm

Sources: U.S. court opinions and official U.S. government publications, read for this article

Maître Hassan KOHEN, avocat au Barreau de Paris
Hassan KOHEN
Attorney at the Paris Bar

Fable 5.2 Launch: What Should Creators Check About U.S. Copyright and AI Training?

Speak to a lawyer 06 89 11 34 45

Anthropic’s help center says Fable 5.2 launched on November 9, 2026, alongside temporary usage offers for eligible subscribers. A new AI release also gives writers, artists and businesses a practical reason to ask a separate question: what does U.S. copyright law say when a work is used in building an AI system, supplied by a user, or appears in an output?

The answer depends on the particular work, the specific copy or use, the applicable permission and any defense. Federal copyright law protects original expression; it does not automatically protect every idea or give an owner control over every use. Courts assessing training disputes have separated the purpose of a copy, the evidence of market effects and the stage of the case. An allegation that survives a motion to dismiss is not a finding that infringement occurred.

The discussion that follows is general in scope: it describes checks useful in comparable situations and makes no findings about Anthropic’s or any other named party’s current situation. It does not establish which materials were used to train any particular model. A creator can make better decisions by keeping records of the original work, agreements, observed outputs and the source of each factual claim before drawing conclusions.

I. When can a creator’s work be used in AI training?

A. Identify the protected work, the owner and the specific use

Start with the work itself. Under federal 17 U.S.C. § 102, copyright protection applies to original works of authorship fixed in a tangible medium. The statutory categories include literary works, pictures, musical works and audiovisual works. The same section says protection does not extend to an idea, procedure, process, system, method of operation, concept, principle or discovery as such. A description of a subject and the author’s particular expression of that subject are therefore different things to examine.

For a writer, the useful first record is the actual text: a dated draft, the published version and the portions that appear to matter. A photographer may keep the original file and the delivered edit; a designer may retain source files and export versions. Dates and version history help distinguish a work that existed before an alleged use from a later revision. A familiar subject, style or factual claim alone does not identify which protected expression was copied.

Ownership is a separate question. Federal 17 U.S.C. § 201 says copyright initially vests in the author or authors, while making specific provision for works made for hire, contributions to collective works and transfers of ownership. A contract may divide rights among an author, publisher, employer or other party. Check the agreement and the actual role in creating the work; the person who holds a file or posted it online is not necessarily the owner of every copyright interest.

A useful rights inventory separates the work from the rights attached to it. Record who created the text or image, whether it was created for an employer, which rights a client or publisher received, and whether a license applies to a particular use. If an anthology contains several contributions, the rights in an individual contribution and those in the collective work can be different. Avoid treating a platform’s ability to host a file as proof that it may authorize every other use of that file.

Federal 17 U.S.C. § 106 gives copyright owners specified exclusive rights, including reproduction, preparation of derivative works and distribution, subject to statutory limitations. Those rights make it necessary to ask what actually happened: was a file downloaded, retained in a collection, transformed for a particular task, supplied by a customer, or reproduced in an output? Each answer points to a different factual record and potentially a different legal analysis.

Do not begin by treating all stages as one event. A training dataset may be assembled from many sources. A user may later upload a document into an application. A system may then generate a response for that user. A business may save, edit or publish that response. Those steps can involve different people, agreements and copies. The presence of one agreement does not by itself answer every question about the others.

For a creator investigating a possible use, separate what is known from what is suspected. A dataset inventory, a model card, a license, a transaction record or an actual output may provide different kinds of evidence. A model’s apparent ability to write about the same subject is not, by itself, proof that a particular manuscript was included in training. Conversely, a public description of a data source should be read closely to see whether it identifies a particular version, period or set of materials rather than the entire history of a system.

For a business building a dataset, record where each collection came from, who supplied it, what permission was obtained and what the permission covers. Keep the operative contract together with any amendments and relevant documentation of access or delivery. Distinguish a license for a defined use from ownership of the underlying copyright. If permission depends on restrictions, the useful operational step is to make those restrictions visible to the people who select, process and retain the material.

B. Fair use is a case-specific defense, not a universal training rule

Federal 17 U.S.C. § 107 provides that fair use is not copyright infringement and identifies four factors for the inquiry: the purpose and character of the use, the nature of the work, the amount and substantiality used, and the effect on the potential market for or value of the work. A commercial use is one consideration within the first factor; it does not answer the entire inquiry alone. Nor does the mere presence of a new technology decide all four factors.

For the first factor, identify the actual function of the contested copy and compare it with the function of the original. For the second, distinguish, for example, a highly creative novel from a factual compilation without assuming either is unprotected. For the third, record both quantity and the significance of the portion used. For the fourth, look for evidence concerning the relevant market and the actual theory of harm. A person assessing a potential dispute needs evidence on the use at issue, not only a general description of the industry.

In Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), the United States District Court Northern District of California on June 23, 2025 granted summary judgment on fair use for the training use and for a particular print-to-digital conversion, while treating separately the retained library copies obtained from unauthorized sources. In Cognella, Inc. v. Meta Platforms, Inc., No. 26-cv-04053-VC (N.D. Cal. September 10, 2026), the United States District Court Northern District of California cited Bartz when denying a discovery request about licensing evidence in the disputes before it. The discovery order applied the earlier reasoning in that setting; it did not decide the merits of every AI training claim. For a creator or developer, the practical lesson is to identify each copy and its purpose rather than assume that a ruling about one use resolves all uses of a work.

In Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025), the United States District Court Northern District of California on June 25, 2025 granted summary judgment on fair use for copying the named plaintiffs’ books as training data on the record before it. In Strike 3 Holdings, LLC v. Meta Platforms, Inc., No. 25-cv-06213-EKL (N.D. Cal. June 11, 2026), the United States District Court Northern District of California cited Kadrey when examining allegations of copying and denied a motion to dismiss. That later decision addressed whether allegations were sufficient at the pleading stage, not whether the alleged infringement had been proved. For the reader, a claim that survives dismissal still requires the factual issues in that proceeding to be resolved.

These are federal district court decisions from the Northern District of California. Their reasoning is useful to understand the questions a court may ask, but a decision by one district judge is not a nationwide ruling that all uses of training material are fair or unfair. The outcome of a later dispute may turn on the parties, evidence, legal arguments and governing appellate law. A summary judgment order and a decision on a motion to dismiss also do different work: one evaluates a developed record under the applicable standard; the other tests a pleading at an earlier stage.

When an output resembles a protected work, assess that output on its own evidence. Save the complete prompt, the response, the system version if shown, and the circumstances in which the response was generated. Then compare the potentially relevant expression with the source work. A shared topic, short factual statement or general style is not the same inquiry as copied expression. Likewise, an output that differs from a work does not alone establish how any training dataset was assembled.

A person considering a license should also ask a narrower commercial question: what practical permission is needed for the intended use, and who can grant it? A license may define the works, territory, duration, methods of use, payment, recordkeeping or restrictions. Those are terms to check in the actual document. Negotiating permission is a business choice and can reduce uncertainty, but an article about fair use cannot supply missing facts or decide the effect of an unsigned or unseen agreement.

II. What should users and creators check before using or challenging an output?

A. Follow the evidence from source material to a specific act

Build a short chronology that can be checked. For each work, record the title or identifying description, creation and publication history, relevant ownership documents and any permission already granted. For each observed use, record who performed it if known, the system and version identified by the available evidence, the date observed and where the material appeared. Mark unknown fields as unknown. This prevents an inference about one product, account or file from becoming an unsupported claim about another.

In Nazemian v. NVIDIA Corp., No. 24-cv-01454-JST (N.D. Cal. May 5, 2026), the United States District Court Northern District of California allowed specified direct and contributory infringement allegations to proceed at the pleading stage and dismissed a vicarious infringement claim with leave to amend. The court addressed the allegations before it rather than finding that the alleged acts had been proved. For a comparable concern, identify the work, the alleged dataset or act, the people involved and the evidence connecting them before choosing a legal theory.

That distinction matters for both sides of a business relationship. A person who believes a protected file was copied may have evidence of ownership but lack evidence of the particular dataset. A developer may have a broad description of its data pipeline but need the actual supplier agreement or processing records to determine whether a permission covered a given collection. Neither gap should be filled with a guess. Put the evidence beside the proposition it supports and label the remaining question clearly.

Copies can have different functions even within one workflow. A file acquired for a collection, a copy created during preparation, a copy used for training and material displayed to an end user are not automatically interchangeable descriptions. Record retention practices as well as acquisition. If a supplier delivered several versions, identify the version actually used and any exclusions or restrictions. These practical distinctions follow the courts’ focus on the particular conduct and record, without assuming the same result in a new dispute.

Consider an author who has licensed a publisher to distribute an ebook while retaining other rights. The useful questions are which text the agreement covers, which rights were granted, whether the publisher may authorize a further use, and whether a later amendment changes the answer. A separate delivery receipt may show what file was provided, but it does not by itself define the permission. The author can keep the agreement, any amendments and the delivered version together so that a future question can be matched to the actual language rather than a recollection of the deal.

A commercial user should read the agreement that governed the transaction. Check whether it addresses inputs, outputs, data retention, disclosure to third parties, user permissions and changes in terms. If a workplace or customer project is involved, compare that agreement with any separate confidentiality or client obligations. The legal effect of a clause depends on the actual wording and circumstances; a generic label such as enterprise, private or commercial is not a substitute for reading it.

Before placing another person’s work into a system, identify what the intended task requires. For an internal summary, the relevant questions may include whether the user is entitled to share the document, whether the contract permits the intended processing and whether sensitive information should be removed. For a public marketing asset, the questions also include whether the final expression incorporates material belonging to someone else. The answer may differ for a public-domain work, a licensed work and a confidential unpublished manuscript.

If an output appears unusually close to an existing work, preserve the complete response rather than only a striking excerpt. Keep the original work as it existed at the relevant time, the prompts and any subsequent human changes. A side-by-side comparison is more useful when it distinguishes original expressive choices from ideas, facts and standard language. Do not edit away context that could explain the similarity, and do not represent a single response as evidence of every response the system can produce.

If the question is whether a work entered a training collection, preserve the evidence that actually speaks to that question: a dataset listing, a relevant admission, a disclosed license, a supplier record or another traceable source. A screenshot of a social-media claim can document that a claim was made; it does not automatically establish the claim. A court filing may describe allegations, while an opinion may decide only a defined issue. Read the document and note which category of evidence it represents.

B. Separate rights in an input from rights in a human-edited output

The rights in an existing work placed into an application and the possible protection for a new output are separate inquiries. Section 102’s protection for original fixed expression remains the starting point. The federal Copyright Office’s March 16, 2023 registration guidance explains how the agency examines works containing AI-generated material and applies its human-authorship requirement. It is agency guidance about registration practices, not a judicial decision that resolves every training or infringement dispute.

The guidance asks how the technology was used and whether the traditional elements of authorship were supplied by a person or produced by a machine. It explains that a work can contain both human-authored and AI-generated material, and that the human selection, arrangement or modification may matter to the examination. The practical step is to retain the drafts and describe the actual human contribution accurately. Do not assume that typing any prompt automatically makes every resulting element a protected human-authored work.

Where registration is contemplated, read the agency’s instructions about identifying and excluding material that should not be claimed. Keep a record of which portions were generated, which were written or substantially revised by a person, and how the final arrangement was assembled. A record of creative decisions is useful only if it corresponds to what actually happened. Creating a polished final file does not reconstruct a missing history of authorship.

A creator may also need to consider material contributed by other people. A collaborator’s text, a commissioned illustration and an employer’s preexisting template can each raise a separate ownership or permission question. Review the agreement for each contribution and document what may be used in the final project. Under federal copyright law, a model’s terms cannot grant a permission that neither the user nor the provider holds or is authorized to give. Equally, a creator’s interest in their own contribution does not automatically settle ownership of every outside element.

For a team using a tool, assign responsibility for checking inputs and reviewing outputs before release. One person can identify the source of material uploaded; another can examine the final text or image and compare it with any reference material. Keep the version that was reviewed and the version that was actually published. If the output is revised, note what changed. These measures help the team answer a concrete question later without claiming that a checklist guarantees a particular legal outcome.

For a user preparing a report, a practical review can begin with a list of the inputs and the source of each one. Next, compare the generated draft with the documents supplied and identify passages that require permission, correction or independent support. Finally, document the human revisions in the working file. This process addresses three different questions: authority to use the input, accuracy and originality of the material selected for publication, and proof of the contribution made to the finished report. It does not turn an unsupported assertion into a verified source.

When a person receives a complaint about a particular work, preserve the communication, the cited material and the relevant agreements before making factual assertions. Identify exactly which use is challenged: training, an upload, an output, a public display or something else. Then check the evidence for that use and the owner’s rights. An allegation is not a judgment, and a refusal or settlement in one matter does not automatically resolve a different transaction.

When a creator suspects unauthorized use, the same disciplined approach works in the other direction. Save a stable copy of the work and ownership records, record what was observed and distinguish an original document from commentary about it. If a source is a court opinion, identify the court, date, procedural stage and actual issue decided. District court decisions can provide a detailed example, but their reach and the quality of the evidence remain part of the analysis.

A useful final review asks four questions. First, what expression is involved? Second, what act concerning that expression is actually supported by evidence? Third, what permission, restriction or statutory defense is relevant to that act? Fourth, what remains unknown? Those questions are more productive than treating a product launch as proof of a training source or assuming that a familiar output establishes a legal violation.

Conclusion

U.S. copyright analysis starts with a protected work, an identifiable act and the rights or permissions relevant to that act. Fair use requires attention to the statutory factors and the evidence in the particular dispute. Keep ownership records, contracts, inputs, outputs and version history separate, and distinguish a reported allegation from a court’s actual decision. That record gives creators and users a sound basis for deciding what to check next.

📄 Official circular