News Analysis

Why ChatGPT Won’t Write Like a Living Author

ChatGPT may refuse to imitate living authors, but the policy raises a bigger question: how should AI firms balance capability, consent and restraint?

Ask ChatGPT to write in the style of Charles Dickens and you may get a Dickensian pastiche: fog, clerks, moral discomfort and all the rest of it. Ask it to write in the style of a living novelist and the answer is likely to be different. The system may refuse, redirect the request, or offer to discuss the author’s techniques rather than imitate them directly.

That distinction has become one of the more revealing boundaries in generative AI. On the surface, it is a modest product rule. In practice, it sits at the junction of copyright law, creator rights, reputational risk, market substitution and public trust.

The dividing line is not just whether an author is alive or dead, but whether imitation risks harming an active reputation, market or creative identity.

It also raises a more uncomfortable question. If AI companies are willing to draw a line around the style of living creators, what does that say about the creative work that helped make such imitation possible in the first place?

Why Living Authors are Different

The distinction between living and deceased authors is not simply a matter of copyright. Copyright generally protects specific expression, not style in the abstract. You can copyright a novel, but not the mere idea of clipped sentences, gothic atmosphere or forensic psychological suspense.

That old distinction was easier to live with when imitation required skill, time and conscious effort. Writers have always learned from other writers; influence is certainly not theft, and pastiche is part of literary culture. A human author who writes “in the manner of” Jane Austen or Raymond Chandler is participating in a long tradition of homage, parody, apprenticeship and experiment.

For a deeper dive into the issue of whether an author's style can be protected by copyright law, and whether AIs can imitate this style without breaking copyright law, read our article on this topic.

In our piece The Automation of Human Creativity, we explore the implications of the AI era on the creative industries. For a primer on how generative AI systems work, see our guide to Demystifying AI.

A Question of Scale

Generative AI changes the scale of the problem. A model can produce fluent imitation instantly, cheaply and repeatedly. That does not automatically make every output infringing. But it does change the practical stakes.

For living authors, the risks are immediate. Their books are still on the shelves, and their names still carry commercial value. Admittedly this could be said for many historical authors. However, living authors still have a reputation; their readers may encounter low-quality imitations, misleading attributions or derivative works that feel close enough to their voice to create confusion and potentially cause brand damage. 

Therefore, even if copyright law does not clearly protect “style”, the commercial and ethical question remains: should a consumer AI product offer imitation of living creators as a standard feature?

One Rule for Images, Another for Prose

OpenAI’s position on the issue is more explicit in some areas than others.

For image generation, the company has clearly stated that DALL·E 3 is designed to decline requests for images in the style of living artists. That is a straightforward living-creator boundary. It says, in effect: you can ask for a visual style, but not the style of a living named artist.

For writing, the public documentation is more opaque. OpenAI’s Model Spec sets out principles for how its models should behave, including commitments around intellectual freedom, transparency and guardrails against real harm. In February 2025, OpenAI published a major update to that Model Spec, making model behaviour a more public governance question rather than a purely internal engineering matter.

And while OpenAI doesn’t specifically state that it prohibits its models from producing text in the style of a living author, many requests along these lines may be met with refusal or redirection. Ask for a broader mode – “write a gothic scene”, “make it spare and noir-like”, “use the conventions of nineteenth-century social comedy” – and the system is more likely to comply. 

The Training Issue

While OpenAI can decide that ChatGPT should not produce text closely imitating a living author, that does not resolve the underlying argument about how foundation models were trained.

AI models such as the GPT series, Claude and many others were trained on massive datasets. Rights holders allege that those datasets included copyrighted books, articles and other creative works, many of them by living authors. Those claims sit at the centre of ongoing disputes over whether AI training is lawful use, industrial-scale appropriation, or something courts have not yet fully defined.

Consequently, authors, publishers, newspapers and other rights holders have brought legal claims against AI companies over the use of copyrighted material in training datasets. OpenAI and other developers have argued that training on large bodies of text can fall within fair use, while claimants argue that their work has been used without permission or compensation.

Towards a System of Permission

Crawler controls such as GPTBot give website owners a way to indicate that their content should not be used to train OpenAI’s generative AI foundation models. Licensing agreements can also bring some content into the AI economy on negotiated terms. These are important mechanisms. They show that the industry is moving, however unevenly, towards more formal systems of permission, exclusion and compensation.

Yet many of these tools are forward-looking. They help govern what happens next. They do not fully answer what should happen if a writer objects after a model has already been trained on material that may have included their work.

This is the unresolved consent problem at the heart of generative AI. Opt-outs are useful, but they are not time machines. Licensing deals may help shape future datasets, but they do not automatically settle disputes about historical training. Behavioural guardrails may limit what users can generate today, but they do not end the argument over how creative labour entered the system in the first place.

Reputation Issues 

That does not make OpenAI uniquely culpable. The same issue runs across the AI industry. But it does mean that the living-authors question is more than a quirky prompt-engineering restriction. It is a small window onto a much larger governance problem.

After all there is a gap between what may be legally defensible and what may be commercially unwise, ethically dubious or reputationally toxic. A model may be capable of imitating a living writer. A company may believe it has legal arguments for how the model was trained. But that does not mean it wants to sell “instant imitation of living authors” as a consumer-facing feature.

Why this Matters Beyond Authors

Authors are only one part of the story. Similar questions are emerging around visual artists, musicians, voice actors, journalists, performers and other creative workers whose styles, voices or bodies can be simulated by AI systems.

In each case, the same pattern appears. First comes the technical capability. Then comes the public unease. Then come the lawsuits, opt-outs, licensing deals and behavioural restrictions.

This is a new kind of product governance. AI companies are not merely deciding how powerful their models should be. They are deciding which powers should be made available, to whom, in what form, and under what conditions.

That distinction will become more important as generative AI becomes more capable. The competitive question will not only be who has the best model. It will also be who can persuade users, creators, regulators and commercial partners that its model is governed responsibly.

The Line in the Sand?

Perhaps the most interesting thing about OpenAI’s approach is not that it restricts what ChatGPT can produce today. It is that it acknowledges a new reality: governing AI is no longer just about building more capable models. It is increasingly about deciding where those capabilities should stop.

The living-authors rule, however obliquely expressed, could be one of those stopping points. It suggests the next phase of AI will not be defined by capability alone, but also by restraint.