News Analysis

Can AI Imitate Authors Without Breaking Copyright Law?

Traditional copyright law protects an author’s words but not their style. Generative AI, which can imitate literary style at scale, pushes this distinction to breaking point.

Can an author’s style be copyrighted?

For most of copyright history, the answer has been: generally, no. Copyright protects the particular expression of an idea, not the idea itself. It protects the words on the page, not the general 'feel' of an author's work – the short, snappy sentences, the gothic atmosphere, dry wit or psychological suspense.

Copyright Protects Expression, Not Style 

The legal starting point is the distinction between ideas and expression.

The US Copyright Office puts the point plainly: copyright protects expression, but not “ideas, procedures, methods, systems, processes, concepts, principles, or discoveries.” Its FAQ says the same thing in practical terms: copyright does not protect an idea or concept, though it may protect the way that idea is expressed in writing or images.

The UK position is similar in broad outline. UK government guidance explains that copyright protects types of works such as literary, dramatic, musical and artistic works, and that copyright is an automatic right. UK legal guidance also frames the point in familiar terms: ideas are not protected by copyright; the expression of ideas fixed in material form may well be.

That matters because “style” usually sits at a high level of abstraction. Sparse prose, gothic mood, satirical social comedy, noir narration, fragmented modernist consciousness or a clipped thriller voice are not usually treated as property in themselves. They are modes of writing. They are ways of making art.

Protected Expression: a Clearer Case

Copyright becomes clearer when the similarity moves from general style into protected expression: copied passages, distinctive scenes, developed characters, dialogue, plot structure, arrangement or other substantial expressive elements. 

In other words, the law is more comfortable asking whether one work directly copied another than whether one work merely ‘feels like’ another.

Of course, that does not mean style is valueless. On the contrary, style may be one of the most valuable things an author has. It is part of how readers recognise them. It is part of the commercial promise attached to their name. And this is precisely why the efficiency and power of large language models in their ability to reproduce style poses legal and ethical problems. 

Why AI Changes the Practical Stakes

Human imitation has always existed. Apprentices copy masters. Young writers pass through phases of sounding too much like the writers they love. Satirists and parodists deliberately mimic tone. Literary influence is not a bug in culture; it is one of the ways culture evolves.

The difference with AI is scale. A human pastiche requires labour, skill and time. It usually involves some degree of transformation, even when the imitation is obvious. 

However, a generative AI system can produce a plausible imitation instantly, repeatedly and cheaply. It can do so for thousands or millions of users. While the output may not copy a specific paragraph, but it may still compete in the same imaginative territory.

The crux of the matter is this: traditional copyright may have little to say if the imitation remains at the level of style. A large language model could theoretically become a factory production line for writing that looks remarkably like the work of a living author without clearly infringing copyright.

Living vs Deceased Authors in Copyright Terms

The issue is sharpest for living authors, but deceased authors’ estates and legacies are impacted as well. 

If a system can produce “a new story in the style of [Author X]”, the commercial harm for a living author is not necessarily that it has reproduced a protected passage. The harm may be that it offers readers, publishers or platforms a cheap approximation of the experience associated with that author. 

Such an author will have an active market, an active reputation and an ongoing relationship with readers. Their style is not just a literary fingerprint. It is part of their commercial identity. If AI-generated work is presented as being “in the style of” that author, it may create reader confusion, reputational risk or market substitution even where the copyright analysis is uncertain. For this reason, various industry bodies have been outspoken in their opposition to these practices.   

As we explored in our previous piece, OpenAI has now introduced restrictions on its ChatGPT model, preventing it from producing text in the style of living authors. However, users can generate text that resembles writing in broader genre categories or historical literary traditions.  

Although this update creates a modicum of protection for living writers, the restriction does not extend to deceased writers, a fact that continues to create ethical and legal difficulties in this space. 

The Case of Deceased Authors

In the UK, government guidance states that literary, dramatic, musical and artistic works generally last until 70 years after an author’s death. In the US, the Copyright Office says that, as a general rule, works created after January 1, 1978 are protected for the life of the author plus 70 years. 

Older US works have more complicated rules. Duke Law’s Center for the Study of the Public Domain notes, for example, that on January 1, 2026, works published in 1930 entered the US public domain, while also warning that copyright terms differ outside the US.

Just because an author is deceased, therefore, does not mean their works are in the public domain.

Even where copyright has expired, estates may still control other assets: trademarks, authorised editions, licensing programmes, adaptations, merchandise, archives or the commercial use of a literary brand. Those rights are not the same as copyright in the author’s prose style, but they can still shape what happens commercially.

The ethical position is also different. A deceased author cannot suffer personal reputational distress in the way a living author can. But readers, estates, publishers and cultural institutions may still care deeply about how a literary legacy is used. 

It remains to be seen whether restrictions will eventually be demanded by literary estates, publishers or readers who object to AI-generated pastiche of deceased authors.

Training Data: a Related but Separate Dispute

The style debate also connects to the separate fight over AI training data.

AI models are trained on large datasets. Authors and publishers have alleged in multiple lawsuits that those datasets included copyrighted books, journalism and other protected works without permission. The Authors Guild describes the author cases as copyright lawsuits based on the alleged unauthorised copying of authors’ works to train generative AI models. OpenAI and other AI companies have generally argued that training can be transformative and may fall within US fair use principles; that argument remains heavily contested and fact-dependent.

Recent US decisions have begun to draw some lines, but they have not ended the debate. In Bartz v. Anthropic, a federal court found that using lawfully acquired books to train AI models could be fair use, while treating the use of pirated books to build a central library very differently. In Kadrey v. Meta, the court also found for Meta on the record before it, while commentators stressed that the outcome was tied to the evidence and arguments presented. Reuters summarised the emerging 2025 picture as courts showing greater openness to fair use where training data is lawfully sourced, while remaining alert to piracy, market harm and output-related claims.

The Anthropic litigation also shows how quickly the legal and commercial tracks can diverge. Anthropic later agreed to a $1.5 billion settlement over claims relating to pirated books, with Reuters reporting in April 2026 that nearly 120,000 authors and other copyright holders were seeking a share of the settlement.

For the style question, the key distinction is this:

• Input question: was the model trained lawfully?
• Output question: should the product generate this kind of imitation?

Those questions overlap, but they are not the same. A court might find certain training uses lawful while a platform still decides that outputs imitating living authors are too risky. Equally, a model trained on lawfully licensed material could still produce an infringing output if it reproduces protected expression too closely.

Licensing May Become the Practical Compromise

One possible future is clearer law. Courts may define when training is fair use, when outputs infringe, and how market harm should be measured. But litigation is slow, expensive and jurisdiction-specific.

The faster route may be licensing.

Publishers and AI companies are already making deals. Reuters has reported on AI content licensing arrangements involving major media organisations, including the New York Times’ agreement with Amazon and Reuters’ own licensing agreement with Meta. The Reuters Institute has also described a broader pattern of licensing announcements involving publishers such as Associated Press, Axel Springer, Prisa Media and Le Monde, alongside continuing lawsuits by news organisations.

Licensing does not solve every style problem. A licence to train on a corpus is not necessarily permission to impersonate a living author. A publisher may have rights in books but not unlimited rights in an author’s name, likeness or personal reputation. Estates may control some assets but not others. Different jurisdictions will draw different boundaries.

Still, licensing creates a practical route through the uncertainty. It allows AI companies to secure higher-quality data, reduce litigation risk and build products around consent rather than scraping first and arguing later.

The Future: Courts, Contracts And Guardrails

The likely future is not one clean answer. It is a mixture of law, licensing and product design.

Courts will continue to decide training-data and output cases. Publishers and AI companies will keep negotiating licences. Authors may demand opt-outs, attribution, compensation or controls over style imitation. AI companies may build more behavioural guardrails around living creators, copyrighted characters, long excerpts and outputs that look like market substitutes.

The most immediate changes may not come from copyright reform. They may come from AI companies deciding which capabilities are too risky to offer at scale.

That is an important shift. In the pre-AI world, style imitation was mainly governed by norms: embarrassment, reputation, editorial judgement, parody conventions and the labour required to pull it off. In the AI world, imitation is governed by interface design. A product either allows the prompt or refuses it. A model either offers a direct pastiche or redirects towards broader traits.

The boundary between influence and imitation is becoming a product setting.

Why The Style Problem Is Really A Governance Problem

Copyright is comfortable asking whether one text copies another. It is less comfortable asking whether a system can mass-produce a substitute for a writer’s creative identity.

That does not mean courts should simply convert style into property. There are good reasons to resist that. If style were too easily protected, literature would become dangerously cramped. Writers need room to learn from one another. Genres need shared conventions. Parody, criticism and homage need breathing space.

But generative AI makes that old compromise more fragile. When imitation required human labour, the law could tolerate a wide zone of influence, homage and stylistic borrowing. When imitation can be automated, scaled and commercialised, the same boundary starts to look less stable.

The law may continue to say that style is not copyrightable in the abstract. But generative AI is forcing a harder question: if style can be replicated instantly, endlessly and commercially, can the old boundary between influence and imitation still hold?

The answer may not come from copyright law alone. It may come from a new mix of legal doctrine, licensing markets, platform rules and creator consent.