If you've ever wanted to write a story in a favorite author's style, you may have asked ChatGPT to "write like so-and-so." Lately, that request doesn't work the way it used to. In Ars Technica's own tests, ChatGPT refused to imitate living authors such as Stephen King and J.K. Rowling. When asked to imitate deceased authors such as Charles Dickens and Ernest Hemingway, it likewise avoided direct imitation and returned writing with a "similar feel." Yet another study found the opposite: ChatGPT complied with requests to imitate deceased authors. The line, in practice, is not uniform. OpenAI has never officially announced this change in behavior, and that silence reflects the company's calculation of copyright risk.
Refusals for living authors are consistent, but results split for deceased authors
In tests Ars Technica ran in July 2026, ChatGPT was asked to write in the style of three living authors: Stephen King, J.K. Rowling and Amy Tan. It refused direct imitation in every case. For Stephen King, it reportedly responded to the effect that it could write in line with traits such as atmospheric, character-driven horror and the unease lurking in small towns, but could not imitate Stephen King's exact style or distinctive voice. It then offered original writing with a "similar feel" to the specified author's style. In the same test, ChatGPT also avoided direct imitation when asked to imitate two deceased authors, Charles Dickens and Ernest Hemingway. In Ars Technica's own verification, whether the author is alive or dead did not determine the outcome.
A comparison of five major AI chatbots conducted in July 2026 by the technology analysis outlet No Latency reported that ChatGPT and Perplexity refused or paraphrased in response to requests to imitate living authors, Claude and Copilot complied with a disclaimer, and Gemini complied directly without one. Simply posing the same question to five AIs produced a range from refusal to no guardrails. However, this audit was a small-scale test using only two types of prompts each for living and deceased authors, and it does not guarantee each company's standing policy.
On how the tools treat deceased authors, there is a result that conflicts with Ars Technica's own verification. In No Latency's comparison, four companies, ChatGPT, Claude, Copilot and Gemini, complied with requests for pastiche of deceased authors, while only Perplexity applied the same standard it uses for living authors and refused. Even for the same ChatGPT, Ars Technica's verification found it avoided imitating deceased authors, whereas No Latency's found it complied, opposite results that suggest the boundary may be less a fixed rule than behavior that shifts with the prompt or the timing of the test. All five companies complied with requests to imitate a broad literary genre without naming an author, and with requests to "analyze" rather than "imitate" a living author's style. The restriction is concentrated in a narrow area: direct copying of a style by naming a specific living individual.
No official announcement and no Model Spec entry: what the silence means
An OpenAI spokesperson did not respond to Ars Technica's request for comment on this change in behavior. Checking the latest published Model Spec (the version released December 18, 2025) directly, no explicit provision prohibiting imitation of living authors' styles can be found. As far as can be confirmed from outside, no official explanation backs this change, and only a shift in the model's output behavior has been observed.
The Model Spec is a document that governs the model's behavior through instruction-based rules, but it contains no clause that names and prohibits imitating living authors' styles. By adjusting only training data or reinforcement-learning reward design, it is technically possible to change behavior without publishing any written policy. The fact that only ChatGPT's output changed while the Model Spec's text did not is consistent with the view that such adjustments were made.
OpenAI faces class-action lawsuits alleging copyright infringement, and a loud announcement that it refuses to imitate living authors could be read as admitting that past imitation was a problem. Quietly changing behavior without codifying it as a rule can be explained as a choice that eases criticism without touching the issues in dispute in the litigation. It cannot be asserted for certain, but there is room to see this kind of calculation as one reason for the silence. The silence stands out more because, for DALL·E 3 image generation, OpenAI explicitly stated in 2023 an official policy of refusing to imitate living artists' styles.
Gemini, which complied without a disclaimer in No Latency's comparison, and Claude and Copilot, which complied with one, have not put into writing, in the public documents we could check this time, how they handle requests to imitate living authors. Users see only the differences in behavior, and the reasoning behind them is, for now, not visible from any company. OpenAI's silence is one piece of a structure shared by all five. This is not a problem confined to one company.
US law's premise: style is not protected
A living author's style itself is not subject to protection under US copyright law. What is protected is concrete "expression," and abstract features called "style" fall, in principle, outside copyright. There is an exception, however.
If the imitated result is judged "substantially similar" to the original, infringement can be found. One of the few precedents is the 1987 federal district court decision in Steinberg v. Columbia Pictures Industries, in which similarity of artistic style was treated as one factor in finding infringement. Robert Brauneis, a professor at George Washington University Law School and co-director of its intellectual property program, told Bloomberg Law in May 2023 about this situation: "We have never been in a situation where an individual's authorship could be imitated this cheaply and this elaborately."
The standard Steinberg set is whether the protected expression in a work is so substantially similar that an ordinary observer would recognize it as having been appropriated from the original. What is asked is similarity of expression, and whether someone could mistake the author is not an element. Nearly 40 years after the ruling, the principle that style alone does not constitute copyright infringement remains unshaken. Whether the "similar feel" writing ChatGPT returns exceeds this standard can only be judged output by output, and there is no boundary a model provider can declare uniformly safe. Switching to a practice of refusing imitation requests outright can be interpreted as the surest way to nip risk in the bud without waiting for judicial rulings on each output.
On September 20, 2023, the Authors Guild and 17 authors including John Grisham, George R.R. Martin and Jodi Picoult sued OpenAI in the US District Court for the Southern District of New York as a class action alleging copyright infringement. On this ambiguous boundary, OpenAI is already standing in court. In its "AI Best Practices for Authors," revised on May 11, 2026, the Authors Guild recommends that members not use generative AI to deliberately imitate another author's distinctive style or voice in ways that diminish the value of that author's work or profit from the imitation. That an industry body spells this out in its own guidelines shows that stylistic imitation has become a practical flashpoint.
There have also been cases where the flashpoint actually ignited. In 2025, when author Lena McDonald published the novel "Darkhollow Academy: Year 2," she left in the main text a response from an AI chatbot saying it had "rewritten the passage to be closer to J. Bree's style," drawing criticism from readers (the AI product used was not identified). The incident shows that stylistic imitation is a risk that is exposed to readers the moment it is published.
The same OpenAI stated officially when announcing DALL·E 3 in 2023 that it was "designed to decline requests that ask for an image in the style of a living artist." About three years passed before ChatGPT's refusal behavior was observed in text generation in July 2026. Although images and text are different media, the sequence, in which the same company set out an official policy on the same issue for images but only a change in behavior was observed for text three years later, suggests it may have approached legal risk differently depending on the medium.
Japan's Agency for Cultural Affairs goes further than the US in protecting creators
Article 30-4 of Japan's Copyright Act permits use of copyrighted works for purposes such as information analysis, including AI training, without the rights holder's permission. This looks similar in structure to the US principle that style is outside copyright protection, but the Agency for Cultural Affairs added conditions in its "General Understanding on AI and Copyright in Japan," published in March 2024. It indicated that when additional training (fine-tuning) is done using only the works of a specific creator, for the purpose of having the AI output that creator's creative expression itself, a purpose of "enjoying" the thoughts and emotions can be assessed as coexisting with the training purpose, in which case the use falls outside Article 30-4's permission-free scope. That a mere similarity of style is not treated as infringement is common to the US, but Japan's Article 30-4 is distinctive in having a provision that can call the training method itself into question. In the US, too, the legality of using copyrighted works for AI training is itself at issue in fair-use litigation, and the Authors Guild's suit questions the legality of this training stage.
"Shosetsuka ni Naro" (Narou), one of Japan's largest novel-posting sites, made it mandatory from June 9, 2026 for newly posted works to set their AI usage status, in one of four categories: direct AI use, indirect AI use, auxiliary AI use and no AI use. From September 1, 2026, attempts to update existing works (posting episodes or editing work information) without a set AI usage status will be blocked. Neglected works are not all forced to set a status uniformly, but with each update, declaring a category becomes unavoidable.
Of the four categories, "auxiliary AI use" and "no AI use" are designed not to be disclosed to third parties; the two disclosed categories are "direct AI use" and "indirect AI use." Even so, which category to choose when imitating a living author's style may become a focal point as operation progresses. The difference between the Agency for Cultural Affairs' approach and US regulation has already come down from the overseas debate to Japanese practice in the concrete form of this disclosure obligation for posters.
When Japanese publishers negotiate content licenses with overseas AI companies, the Agency for Cultural Affairs' position that the way training data is collected can itself be a problem may give them leverage in negotiations. It broadens the points that can be raised at the table compared with the US case-law doctrine, where the similarity of outputs is the main issue. For overseas AI companies that handle Japanese-language content as training data, negotiating while ignoring the Agency's position could rebound later as legal risk, and for the publishing industry this difference is already an element that cannot be ignored.
Some observers say that, as of May 2026, no copyright infringement lawsuit against a generative AI company had yet been filed in Japan. Others point to the possibility of a suit within the year, so the situation remains fluid. The next focus is whether the Agency's creator-protective view will be tested in an actual lawsuit.
Authors gain; vendors that comply without disclaimers lose out
The biggest beneficiaries are living authors like Stephen King, whose style itself has commercial value, and organizations that represent authors, such as the Authors Guild. Their styles are treated as something not to be copied without permission, giving them more bargaining chips. Moves by publishers and media companies to sign license agreements with AI companies and negotiate terms of use for their content may also find it easier to presuppose the existence of this boundary. With the line of refusing to imitate living authors drawn, winners and losers are thus becoming clearly separated.
The losers are individual users and self-publishing operators who want to use stylistic imitation as a shortcut to creation. As the Lena McDonald case shows, the usage of naming a living author and asking to "write like that person" is already blocked on ChatGPT and Perplexity. Vendors like Gemini, which complied directly without a disclaimer in No Latency's comparison audit, may not appear to have crossed a line, but they are placed in a position that makes them easy targets for criticism and tighter regulation. Compliance with a disclaimer, as with Claude and Copilot, also accepts imitation itself as long as a disclaimer is present, carrying the risk of being swept up in any future tightening of the line.
When creating internal proposals or marketing copy with generative AI, cases can arise in which the person in charge unintentionally specifies a particular author's or creator's style. If a company uses a vendor that complies without a disclaimer, the company itself bears the risk of the imitation being pointed out after the deliverable is published. When Japanese companies compare AI vendors, occasions are expected to increase in which they evaluate not only price and processing performance but also the vendor's stance on copyright. The more a use case involves mass content generation, the more differences in vendors' compliance policies translate directly into differences in the rate of post-publication trouble.
Next question: how Gemini and Claude respond
Whether it was the technical team's judgment or a request from the legal department that changed ChatGPT's behavior is an internal process that cannot be verified from outside. The only fact visible externally is that OpenAI has never codified a policy of "refusing to imitate living authors' styles." If an official line like the one DALL·E 3 set out for images in 2023 were also documented for text generation, the differences in handling that vary across the five companies might narrow.
The course of the Authors Guild lawsuit filed in 2023 and developments in Japan's system offer clues to what comes next. The Agency for Cultural Affairs' position that the way training data is collected can itself be problematic will be tested by how it is actually applied after Narou's AI usage disclosure extends to updates of existing works from September 1, 2026. How far the line protecting living authors' styles gets codified will become visible through two concrete indicators: the ruling in this lawsuit and Narou's operating record.