Derivative work

On this page 8

A derivative work is a new work that recasts, transforms or adapts protected expression taken from one or more preexisting works, such as a translation, a film adaptation, a musical arrangement or an edited remix. In the field of Copyright & AI, the concept matters because it decides whether an AI model, its outputs or a fine-tuned variant carry over rights that belong to the authors of the training material.

The exclusive right to prepare derivative works belongs to the copyright owner of the original. A derivative work can earn its own copyright for the new contribution, but that protection never extends to the borrowed material, and an unauthorized derivative work is an infringement unless an exception or a defense such as fair use applies.

How the law defines a derivative work

A derivative work exists when a new work incorporates protectable expression from an earlier work and adds enough original authorship to be more than a copy. United States law defines it in 17 U.S.C. section 101 and grants the owner control over it in section 106(2). The Berne Convention protects the same interest through the rights of translation, adaptation and arrangement.

European Union law has no single harmonized definition. Instead, the adaptation right is left largely to national statutes, while the reproduction right in the InfoSoc Directive 2001/29/EC captures many derivative uses because any new work that reproduces original elements also reproduces part of the source. The Court of Justice tests reproduction by asking whether the author’s own intellectual creation survives in the new material.

The table below lists the elements courts usually examine before they treat a new creation as derivative.

ElementQuestion the court asksTypical outcome if missing
Borrowed expressionDoes the new work copy protected expression, not only ideas or style?No derivative work; at most inspiration
Substantial similarityWould an ordinary observer recognize the original in it?Use is too remote to infringe
New authorshipDid someone add original, creative choices?Result is a mere copy, not a new work
Fixation or formIs the result a concrete, perceptible work?No separate work exists to protect
Core elements of a derivative work

Derivative works compared with neighboring concepts

The derivative work concept sits between pure copying and independent creation, and it is often confused with related categories. The difference matters because each category triggers different rights and different defenses.

CategoryRelation to the sourceRights consequence
Exact reproductionCopies the work without new authorshipReproduction right applies directly
Derivative workTransforms protected expression with new authorshipNeeds permission; new layer may be protected
CompilationSelects and arranges existing worksProtection only for selection and arrangement
Independent workShares ideas, genre or style onlyNo permission needed; ideas are free
Parody or pasticheEvokes the source to comment or imitateOften covered by an exception or fair use
How derivative works differ from related categories

A key point is that style itself is not protected. A painting in the manner of an established artist is not derivative unless it reproduces specific expressive elements of identifiable works. This distinction is central when users prompt an image generator for a picture in the style of a named illustrator.

Where AI systems raise derivative work questions

Generative AI creates three separate derivative work questions: whether the training process creates derivatives, whether the trained model itself is a derivative of its training data, and whether individual outputs are derivatives of specific works. Each question has a different legal answer and different evidence.

AI stageDerivative work argumentPrevailing view and uncertainty
Training copiesDatasets reproduce works before learningTreated as reproduction; TDM exception or fair use decides
Model weightsParameters encode the training worksUsually not derivative unless works can be extracted
Generated outputOutput resembles a specific source workDerivative only with substantial similarity
Fine-tuned modelNarrow tuning on one artist or catalogHigher risk because memorization is more likely
Retrieval outputSystem inserts retrieved text into answersCan reproduce or adapt sources directly
Derivative work questions across the AI lifecycle

Courts have so far been skeptical of the claim that model weights are derivative works as a whole, because most models do not contain recognizable copies of individual works. The picture changes when a model memorizes and regurgitates training items, since verbatim or near-verbatim output is strong evidence of copying. In the European Union, the text and data mining exception of the DSM Copyright Directive and the TDM opt-out govern the training copies, but neither exception authorizes outputs that reproduce protected expression.

Ownership of the new layer

A lawful derivative work can attract a separate copyright in the added material, owned by its author. That protection is thin and layered: it covers only the new creative contributions and never extends the term of the underlying work or revives material that is already in the public domain.

For AI-assisted derivatives, two hurdles apply at once. The user needs permission or an exception for the borrowed expression, and the new layer is protected only where a human made creative choices, because most jurisdictions deny protection to purely machine-generated material. A Creative Commons license may allow adaptations, but the NoDerivatives (ND) condition forbids sharing adapted material, and the ShareAlike (SA) condition requires releasing the derivative under the same license.

Assessing derivative work risk in practice

Organizations that build or use generative tools reduce derivative work exposure by controlling inputs, prompts and outputs. The decision table below summarizes common situations and the response experienced counsel usually recommends.

SituationRecommended responseReason
Prompt names a specific copyrighted characterObtain a license or redesignCharacters are protectable expression
Output closely matches a known image or textDiscard or rework the outputSubstantial similarity indicates copying
Fine-tuning on one author’s catalogSecure AI training licensing firstMemorization and market harm are likely
Generic prompt in a broad genreUse with routine similarity checksGenre and ideas are not protected
Source licensed under CC BY-SACredit the author and keep the licenseShareAlike binds adaptations
Decision guide for AI-assisted derivative uses

Common misconceptions about derivative works

  • Changing a fixed percentage is not a safe harbor. No rule says that altering 10 or 30 percent of a work avoids infringement; the test is qualitative similarity, not arithmetic.
  • Crediting the author is not permission. Attribution satisfies some licenses and moral rights, but it does not replace consent to adapt.
  • Transformation alone does not end the analysis. In the United States, the Supreme Court decision in Andy Warhol Foundation v. Goldsmith (2023) narrowed how far a new purpose can excuse a commercial adaptation.
  • AI involvement does not launder a work. An output that reproduces protected expression infringes regardless of which tool generated it.

The simple explanation

A derivative work is something new built on top of someone else’s creative work, like a movie based on a novel or a cover song with new lyrics. The original creator decides who may make such versions, so making one usually needs permission.

With AI, the question is whether a generated picture or text copies recognizable parts of a real work. Imitating a general style is usually fine. Reproducing a specific drawing, character or passage is not, even if a machine produced it. The closer the result looks to one particular original, the higher the legal risk.

How it is used in practice

A mid-sized publisher plans to release an illustrated children’s series and asks its design team to use an image generator for concept art. During review, the legal team notices that one output reproduces the distinctive costume and pose of a well-known cartoon character almost exactly, although the prompt mentioned only a fox detective in a raincoat.

The team treats that image as a likely unauthorized derivative work and discards it. They add a similarity screening step that compares outputs against reference images of protected characters, ban prompts that name living artists or franchises, and require an illustrator to redraw every approved concept by hand. The final artwork is therefore human-authored and protectable, and the publisher documents the process in case a rightsholder ever challenges the series.