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.
| Element | Question the court asks | Typical outcome if missing |
|---|---|---|
| Borrowed expression | Does the new work copy protected expression, not only ideas or style? | No derivative work; at most inspiration |
| Substantial similarity | Would an ordinary observer recognize the original in it? | Use is too remote to infringe |
| New authorship | Did someone add original, creative choices? | Result is a mere copy, not a new work |
| Fixation or form | Is the result a concrete, perceptible work? | No separate work exists to protect |
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.
| Category | Relation to the source | Rights consequence |
|---|---|---|
| Exact reproduction | Copies the work without new authorship | Reproduction right applies directly |
| Derivative work | Transforms protected expression with new authorship | Needs permission; new layer may be protected |
| Compilation | Selects and arranges existing works | Protection only for selection and arrangement |
| Independent work | Shares ideas, genre or style only | No permission needed; ideas are free |
| Parody or pastiche | Evokes the source to comment or imitate | Often covered by an exception or fair use |
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 stage | Derivative work argument | Prevailing view and uncertainty |
|---|---|---|
| Training copies | Datasets reproduce works before learning | Treated as reproduction; TDM exception or fair use decides |
| Model weights | Parameters encode the training works | Usually not derivative unless works can be extracted |
| Generated output | Output resembles a specific source work | Derivative only with substantial similarity |
| Fine-tuned model | Narrow tuning on one artist or catalog | Higher risk because memorization is more likely |
| Retrieval output | System inserts retrieved text into answers | Can reproduce or adapt sources directly |
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.
| Situation | Recommended response | Reason |
|---|---|---|
| Prompt names a specific copyrighted character | Obtain a license or redesign | Characters are protectable expression |
| Output closely matches a known image or text | Discard or rework the output | Substantial similarity indicates copying |
| Fine-tuning on one author’s catalog | Secure AI training licensing first | Memorization and market harm are likely |
| Generic prompt in a broad genre | Use with routine similarity checks | Genre and ideas are not protected |
| Source licensed under CC BY-SA | Credit the author and keep the license | ShareAlike binds adaptations |
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.