COPYRIGHT INFRINGEMENT IN AI GENERATED OUTPUTS: STYLE COPYING,SUBSTANTIAL SIMILARITY AND LIABILITY IN GENERATIVE AI : Author: Shigarfa Shamshad
The increasing use of generative artificial intelligence has raised difficult questions concerning copyright infringement in AI-generated outputs. This article examines three closely connected issues: the copying of artistic style, the test of substantial similarity, and the allocation of liability between users and AI developers. It distinguishes between the imitation of an author's or artist's general style, which is not ordinarily protected by copyright, and the reproduction of protected expression from an identifiable copyrighted work. The article further considers how substantial similarity may be assessed where an AI-generated output resembles an existing work and how factors such as prompts, training data and model memorisation may affect the analysis
ABSTRACT
The increasing use of generative artificial intelligence has raised difficult questions concerning copyright infringement in AI-generated outputs. This article examines three closely connected issues: the copying of artistic style, the test of substantial similarity, and the allocation of liability between users and AI developers. It distinguishes between the imitation of an author's or artist's general style, which is not ordinarily protected by copyright, and the reproduction of protected expression from an identifiable copyrighted work. The article further considers how substantial similarity may be assessed where an AI-generated output resembles an existing work and how factors such as prompts, training data and model memorisation may affect the analysis. It also examines whether liability should rest with the user who generates or distributes the output, the AI provider that develops the model, or both. The article argues for a fact-specific approach that separates style imitation, protected expression and actual reproduction, while applying existing copyright principles to the particular conduct of the parties involved.
Introduction
Generative AI has made copying a more complicated legal problem. A person can now ask an AI system to produce an image, story, song or other creative work that resembles something already created by another person. The difficulty is that resemblance alone does not necessarily amount to copyright infringement.
This is particularly clear in the case of artistic style. A user may ask an AI system to produce an image in the style of a particular artist. The output may contain similar colours, lighting, composition or visual techniques, but it may not reproduce any particular work of that artist. Copyright law generally does not protect style in the abstract. It protects the original expression contained in a work.
The more difficult case is where an AI-generated output reproduces identifiable and protected elements of an existing work. At that point, the question is no longer simply whether the AI has copied a style. The issue becomes whether it has reproduced a substantial part of a copyrighted work.
This distinction is important because it determines both whether infringement has occurred and who, if anyone, should be held liable.
1. Style Copying and Copyright
The demand for AI-generated content "in the style of" a particular artist has raised concerns among creators. An artist may spend years developing a recognizable recognisable visual language, only to find that an AI system can generate thousands of images that resemble it.
There is, however, a difference between copying a style and copying a work.
Copyright protects the expression of an idea, not the idea itself. The same principle limits the protection available for artistic techniques, genres and general styles. If style alone were protected, an artist could potentially prevent other artists from working in a similar aesthetic. That would give copyright a scope considerably wider than its traditional function.
For example, an AI-generated painting may use the colour palette, lighting and broad visual characteristics associated with a particular painter. If it does not reproduce the composition or other distinctive expression of a particular copyrighted painting, establishing copyright infringement may be difficult.
The position changes where stylistic imitation is accompanied by reproduction of protected elements from an identifiable work. If the AI output reproduces a distinctive character, composition, arrangement of objects or other original expressive features from a particular painting, the copyright owner has a stronger claim.
The legal problem, therefore, is not whether an AI system can imitate an artist's style. It clearly can. The question is whether the imitation has crossed the line into copying protected expression.
2. Substantial Similarity
The doctrine of substantial similarity provides a useful framework for analysing AI-generated outputs.
Copyright infringement does not require a work to be copied word-for-word or reproduced exactly. A work may infringe where it takes a substantial part of the protected expression of another work.
Indian law has long recognised the distinction between an idea and its expression. In R.G. Anand v. Deluxe Films, the Supreme Court held that there can be no copyright in an idea, subject matter, theme or plot as such. The question is whether the later work has copied the expression of the earlier work.
This principle is particularly relevant to generative AI.
Suppose two AI systems produce images of a person standing in front of a historical building. The subject matter is common and does not establish copying. But if one output reproduces the unusual composition, positioning, background elements and other distinctive features of a particular copyrighted photograph, the comparison becomes different.
The court would therefore have to identify the protected elements of the original work and determine whether those elements have been reproduced in the AI output.
This is why simply saying that an AI output "looks similar" is insufficient. Similarity must relate to material that copyright actually protects.
3. The Problem of Similarity Without Direct Copying
Generative AI creates an unusual evidentiary problem because similar outputs can arise without a straightforward act of copying.
A model may have been trained on millions of images. An output may contain features that appear in many of those images rather than features taken from one particular work. Similarity may therefore result from common artistic conventions or from the model's generalisation of patterns in its training data.
This makes it difficult to establish the connection between an original work and a particular output.
The U.S. Copyright Office has specifically identified this problem. It has asked whether the traditional substantial-similarity test is sufficient for generative-AI output cases and how copyright owners can establish copying when developers do not disclose the training material used by their models.
The problem is especially serious where the output is similar but not identical. A copyright owner may be able to demonstrate that the output resembles their work but still face difficulty proving that protected expression from that particular work was reproduced.
4. Memorisation and Regurgitation
The question becomes more serious where an AI model appears to reproduce material from its training data.
There is a difference between a model learning general patterns and reproducing a particular work. An AI system that generates a new landscape after learning from thousands of landscape paintings is not necessarily reproducing any one of them.
The position is different if the system produces a passage of text that is identical or nearly identical to a copyrighted article, or an image that closely reproduces a particular photograph.
In such cases, evidence of memorisation or "regurgitation" may become relevant.
The recent decision of the Delhi High Court in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC is particularly relevant here. The Court considered whether ChatGPT memorised and reproduced ANI's copyrighted literary works and whether its outputs amounted to substantial reproduction. The Court ultimately held, at the interim stage, that ANI had not established substantial similarity or memorisation and regurgitation in the examples placed before Indian Kanoon
The decision is significant because it shows that the mere presence of copyrighted material in an AI training process does not automatically establish infringement by every output. The copyright owner still has to connect the allegedly infringing output with protected expression from the original work.
5. Training and Output Are Separate Issues
A distinction must also be made between the legality of training an AI model and infringement caused by a particular output.
Training may involve making copies of copyrighted works. Whether those copies are permitted under copyright law is a separate legal question.
An output is a different matter.
A model could be trained on copyrighted material and still generate an output that does not substantially reproduce any particular work. Conversely, an output could reproduce a particular copyrighted work closely enough to create an infringement claim.
The Delhi High Court's 2026 ANI decision illustrates this distinction. The Court was required to consider both the training of the model and the nature of the outputs produced for users. It concluded, at the interim stage, that the storage of ANI's works for training fell within the relevant Section 52(1)(a) exception and that the outputs produced through the RAG process were not substantially similar to ANI's works.
The decision was expressly limited to the interim application and did not finally determine the entire suit.
7. Liability of the User
The user is the most obvious potential defendant when the user deliberately uses AI to reproduce copyrighted material.
Consider a person who uploads a copyrighted illustration and asks an AI system to reproduce it with minor changes, then sells the resulting images. The user's involvement in the reproduction is direct and substantial.
The situation is different when a user enters an ordinary prompt and receives an unexpectedly similar output. It would be difficult to justify a rule under which users are automatically liable for every similarity generated by an AI system.
Liability should therefore depend on the user's conduct and the applicable copyright rules. The fact that an output was generated by AI does not, by itself, determine whether the user has infringed copyright.
8. Liability of the AI Provider
The position of the AI provider is less straightforward.
An AI provider develops the model, determines aspects of its training process and controls the system through which outputs are generated. Copyright owners may argue that the provider should be responsible where its system repeatedly produces substantially similar copies or where the provider knowingly facilitates infringing uses.
The provider may argue that it supplies a general-purpose technology capable of lawful uses and that users, rather than the provider, determine what individual outputs are requested and distributed.
The legal question is therefore not simply whether the provider's technology was capable of producing an infringing output. The relevant issue is whether the provider's own conduct satisfies the requirements for direct or secondary liability under the applicable law.
The U.S. Copyright Office has identified this as an unresolved issue, including whether liability should attach to model developers, developers of systems incorporating AI models, end users or other parties when an AI output infringes copyright.
9. Who Should Bear the Risk?
Generative AI makes it possible for several parties to contribute to the production and distribution of one allegedly infringing work.
The model developer may have created the technology. The user may have written the prompt. Another company may have incorporated the model into its platform. A third party may have distributed or sold the resulting work.
It would be difficult to impose identical responsibility on all of them.
A more useful approach is to examine each actor's conduct separately. The relevant questions include who supplied the copyrighted material, who requested the reproduction, who controlled the output, who distributed it and who knew or had reason to know about the alleged infringement.
This approach is consistent with the basic structure of copyright law, where liability depends on the particular exclusive right involved and the conduct of the person alleged to have violated it.
10. The Limits of the Style Argument
The debate over AI-generated art sometimes treats an artist's style as if it were equivalent to a copyrighted work. That is not necessarily correct.
A style can be commercially valuable without being protected by copyright as such.
An artist may have a recognisable style, but allowing copyright protection over that style would raise difficult questions. How similar must another artist's work be before it becomes unlawful? Can a creator prevent others from using a particular colour scheme or painting technique? Can a writer prevent another writer from using a similar sentence structure?
These questions show why the law has generally drawn a line between style and expression.
The stronger copyright claim arises when the AI output can be traced to a particular work and reproduces protected expression from that work. A claim based only on the general appearance of an artist's style is considerably more difficult.
11. A Better Framework for AI Output Cases
AI copyright disputes can be analysed through a relatively simple sequence of questions.
First, identify the original work and determine whether it is protected by copyright.
Second, identify the specific elements of that work that are allegedly reproduced.
Third, compare those elements with the AI-generated output and determine whether the similarities concern protected expression.
Fourth, examine how the output was produced, including the user's prompt, any uploaded reference material and, where relevant, evidence concerning the model's training or memorisation.
Finally, determine the role of each party involved in generating and distributing the output.
This approach avoids treating every AI-generated similarity as infringement while still protecting creators when an AI system produces a substantial reproduction of their work.
Conclusion
The copyright problem created by generative AI is not simply that machines can imitate artists. The more difficult problem is determining when imitation becomes reproduction of protected expression.
Style, by itself, should not be confused with copyrightable expression. An AI-generated work may resemble an artist's general style without copying a particular copyrighted work. The position changes when the output reproduces distinctive elements of an identifiable work.
Substantial similarity therefore remains central. The ANI Media v. OpenAI decision demonstrates the importance of identifying actual instances of reproduction rather than relying on the general proposition that an AI model was trained on copyrighted material.
Liability presents a separate question. A user who deliberately asks an AI system to reproduce a copyrighted work is in a different position from a user who receives an unexpected similarity. The position of the AI provider must likewise depend on its own conduct rather than merely on the existence of an infringing output.
The most appropriate legal approach is therefore a factual one: identify the protected expression, compare it with the AI output, establish the connection between the two, and then determine the responsibility of the parties involved. Generative AI may require courts to apply these principles to new technological facts, but it does not remove the underlying distinction between an idea, a style and protected expression.
References
Cases
ANI Media Pvt. Ltd. v OpenAI OpCo LLC, CS(COMM) 1028/2024, Delhi High Court, judgment
Andersen v Stability AI Ltd, 744 F Supp 3d 956 (ND Cal 2024).
Eastern Book Company v D.B. Modak, (2008) 1 SCC 1.
R.G. Anand v Deluxe Films, (1978) 4 SCC 118; AIR 1978 SC 1613
Feist Publications, Inc. v Rural Telephone Service Co., 499 US 340 (1991).
Google LLC v Oracle America, Inc., 593 US 1 (2021).
Authors Guild v Google, Inc., 804 F 3d 202 (2d Cir 2015).
Nichols v Universal Pictures Corp., 45 F 2d 119 (2d Cir 1930).
Computer Associates International, Inc. v Altai, Inc., 982 F 2d 693 (2d Cir 1992).
Statutes
The Copyright Act, 1957 (India), ss 13, 14, 51 and 52.
Copyright, Designs and Patents Act 1988 (UK).
Copyright Act 1976 (United States), 17 USC.