A few dozen pictures are enough to teach a model somebody's manner. Copyright has never protected a manner. The interesting law, and the practical protection, sits in the gap between those two facts.
Style is what is left when you remove every particular photograph, which is exactly why copyright has so little to say about it.
Yes, a model can copy a photographer's style, and no, style as such is not protected by copyright. Both halves of that answer are well established. What is new, and what this article is about, is the mismatch between them: the law's unit is the individual work, and the machine's unit is the body of work. A fine-tuned model takes nothing recognisable from any one photograph and something unmistakable from all of them. Copyright was not designed to see that. Other things, mostly contracts, are.
This article is general information, not financial or legal advice. Checked on 18 November 2026.
What does it mean for a model to copy a style?
Two different techniques go by the name.
Style transfer is the older. In August 2015 Leon Gatys, Alexander Ecker and Matthias Bethge showed that a neural network could separate the content of one image from the style of another and recombine them: your holiday photograph rendered in the brushwork of a painting. It works image by image, and it needs a specific style image to borrow from.
Fine-tuning is the one that matters now. A large text-to-image model already knows, in a general way, what photographs look like. Fine-tuning adjusts it with a small extra set of images so that it acquires something particular. The DreamBooth paper of August 2022, from Google researchers, described fine-tuning a pretrained model on "just a few images" of a subject so that a made-up identifier in the prompt would call that subject up in new scenes. The same approach, and lighter-weight relatives that train a small add-on to the model instead of the whole of it, works for a manner as well as for a subject. Feed in a consistent set of one person's pictures and the add-on learns what they share.
What they share is the interesting part. For a photographer it is lighting habits, lens choice, colour, how close they stand, what they leave out. None of that is in any single frame. It exists only across many, which is why our comparison of fine-tuning and foundation training treats a coherent single-author archive as fine-tuning material by nature.
There is a third, lazier route: typing "in the style of" followed by a name. That works only to the extent the base model saw enough captioned examples of the name during training, so it favours the famous and the heavily published.
Is style protected by copyright?
Not on its own. The principle is one of the oldest in the field: copyright protects expression, not ideas. The United States wrote it into statute. Section 102(b) of the Copyright Act says that in no case does protection extend to "any idea, procedure, process, system, method of operation, concept, principle, or discovery". UK legislation contains no equivalent sentence, but the courts apply the same distinction when they ask whether a substantial part of a work has been taken, and international agreements to which the UK is party state it in terms.
A way of lighting jewellery, a preference for hard shadows, a palette: these are methods and ideas. Anyone may adopt them, which is how photography has always developed. Assistants learn from the photographers they work for and then go and do something recognisably descended from it, and any browse through our index of photographers is a browse through manners handed on.
Two cautions. First, the line between a style and a particular image's expression is a matter of degree, and photographs have been held to infringe where a second image recreated the distinctive composition of a first without copying it mechanically. The English example is Temple Island Collections v New English Teas ([2012] EWPCC 1), where the Patents County Court held that a separately taken picture of a red bus against a monochrome Westminster reproduced a substantial part of the claimant's photograph; the judge added that many pictures of the same subject would not infringe. Whether a given output crosses that line is a legal judgement about two particular images; a side-by-side view, such as our compare two images tool gives, shows what differs in the pixels and decides nothing about the law. Second, "not protected by copyright" does not mean "unprotected". It means the protection, if any, comes from somewhere else.
What do primary sources say about AI and style?
The fullest official discussion is in the US Copyright Office's report on generative AI training, Part 3 of its study, released in a pre-publication version in May 2025. Its status needs stating carefully: the report is not binding, the cover says a final version would follow with no substantive changes expected, and the Office's website still listed only the pre-publication version when we checked.
The report accepts the orthodox position that copyright does not protect style as such, and then declines to treat that as the end of the matter. In its discussion of market harm it says that "stylistic imitation made possible by its use in training may impact the creator's market", even where outputs are not substantially similar to any specific work. It notes in a footnote that replicating a style may in some cases capture protectable elements. Its summary lists, among the ways outputs can dilute the market for the works trained on, "generating material stylistically similar to those works".
That is a notable move. It relocates style from the question "was something copied?" to the question "was the market harmed?", which in US law is one of the factors in fair use. It is an agency's analysis, not a court's holding, and US courts have not yet ruled on fair use for image models.
Where do the court cases stand?
Andersen v Stability AI (Northern District of California) is the case most directly about artists and style. It was filed in January 2023 by illustrators against Stability AI, Midjourney, DeviantArt and later Runway. Copyright claims and some trade mark related claims survived motions to dismiss in August 2024. There has been no trial. According to a docket tracker, the summary judgment hearing was moved in February 2026 to 17 February 2027, and a further extension of about three months was reported on 28 September 2026. Reports elsewhere that the case went before a jury in September 2026 are contradicted by the docket-based sources.
Getty Images v Stability AI (England and Wales) was decided by the High Court on 4 November 2025. Getty abandoned its claim about training during the trial, so the court did not rule on whether training infringes UK copyright; it held that the model itself was not an "infringing copy". That point went to the Court of Appeal, which heard argument on 11 and 12 November 2026. Judgment is pending. The case is about a library's photographs, not about style, and we follow it in our Getty v Stability AI timeline.
No decided case in either country, so far as we could confirm, settles whether imitating a photographer's style with a model is lawful.
If copyright does not protect style, what does?
| Route | What it addresses | Limit |
|---|---|---|
| Copyright in individual works | Outputs that reproduce a substantial part of a specific photograph; unlicensed copying of the works for training, where the law treats that as infringement | Does not reach a general manner |
| Contract | What a licensee may train, and whether a style-specific model or use of the name is allowed | Binds only those who sign |
| Name and reputation | Marketing images under a photographer's name: false attribution, passing off, false endorsement | Depends on country and on the name being used |
| Platform rules | Marketplaces that refuse artist names in prompts | Private policy, changeable |
| Technical measures | Cloaking tools that disrupt mimicry | An arms race |
On platform rules, Adobe Stock's guidelines for generative AI submissions, updated on 11 June 2026, tell contributors not to include artist names, real people or fictional characters in prompts, titles or keywords. On technical measures, the Glaze project at the University of Chicago published a tool in February 2023 that makes small changes to images before they are posted; its authors surveyed more than 1,000 artists and reported that the cloak disrupted mimicry in over 92% of cases under normal conditions and over 85% against adaptive countermeasures.
The contract row is the one with teeth. Fine-tuning needs a supply of consistent, good-quality images. Where those images are available only under licence, the licence can say what may be built.
What should a licence say about style?
For anyone licensing a single photographer's work for machine learning, in either direction, the questions are concrete:
- May the images be used to train a general model only, or also a model or add-on specific to this photographer's work?
- May the photographer's name be used in prompts, product descriptions or marketing?
- May outputs be presented as being by, or in the manner of, the photographer?
- What happens to the trained weights when the licence ends?
- Who is told, and how, that the material was used?
These are commercial terms, not legal doctrines, and they do the work that copyright's idea and expression distinction leaves undone.
Where this leaves a single-photographer archive
The Robert Harper Photograph Library holds 95,487 unique photographs by its inventory count, the work of one photographer across about thirty-five years, much of it studio work in consistent formats for magazines and commercial clients. That is precisely the kind of corpus from which a manner can be learned. It is also held by a single rights holder, 825 Ventures Ltd, which means the question of what may be built from it has one answer and one place to ask. The ownership position is set out under rights and stewardship.
We make no claim that any use is pre-approved. Licences for AI training and data use are negotiated directly with the rights holder, are subject to rights review, and are where terms about style-specific models and the use of the photographer's name belong. If that is the conversation you need to have, that page is where to begin.
Frequently asked questions
Is a photographer's style protected by copyright?
Not as such. Copyright protects particular works and the expression in them, not ideas, methods or a general manner of working. US statute says so expressly, and UK law applies the same distinction through case law. A specific photograph is protected, and copying a substantial part of its expression can infringe, but a look shared across a body of work is not owned.
Can AI copy a photographer's style?
Technically, yes. Style transfer has been possible since 2015, and fine-tuning methods published from 2022 onwards let a general image model be adapted to a subject or a manner using a small set of example pictures. The output can resemble a photographer's work without reproducing any one photograph, which is why the legal position is contested.
Is it legal to prompt an image generator in the style of a named photographer?
There is no settled general answer, and it depends on the country, the output and how it is used. Style alone is not protected by copyright, but an output may still copy protected expression, and using a living person's name to market images can raise separate issues such as false attribution, passing off or false endorsement. Take advice on specific facts.
How can photographers protect their style from AI?
By controlling the inputs and the terms. Options include licensing only under contracts that prohibit style-specific models and the use of the photographer's name, reserving rights in machine-readable form where the law recognises it, limiting high-resolution publication, and tools such as Glaze that alter images before posting. None is complete on its own.
Has any court ruled on AI style imitation?
Not definitively, as far as could be confirmed on 18 November 2026. The US artists' class action Andersen v Stability AI has not reached trial, and a further extension of its schedule was reported on 28 September 2026. In England, Getty Images v Stability AI did not decide whether training infringes; the appeal was heard on 11 and 12 November 2026 and judgment is pending.
Sources
- 17 U.S. Code 102: Subject matter of copyright, subsection (b) (Legal Information Institute)
- US Copyright Office: Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version, May 2025)
- Gatys, Ecker and Bethge, A Neural Algorithm of Artistic Style (arXiv 1508.06576)
- Ruiz et al., DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation (arXiv 2208.12242)
- Shan et al., Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models (arXiv 2302.04222)
- ChatGPT Is Eating the World: Andersen v Stability AI docket tracker (entries to 28 September 2026)
- Adobe Stock: Generative AI content submission guidelines (updated 11 June 2026)
- Getty Images v Stability AI 2025 EWHC 2863 (Ch), The National Archives
Published by Archive.London. Corrections are welcome under our editorial policy.

