ChatGPT now refuses direct requests to mimic specific authors' writing style, a shift with potential legal implications in ongoing copyright lawsuits

ChatGPT Won't Imitate Authors Anymore. Gemini Still Will.

Ask ChatGPT to write a horror opening in the style of Stephen King, and it now refuses. Same for J.K. Rowling, Amy Tan, Charles Dickens, Ernest Hemingway. Instead of cloning a named author's voice, ChatGPT offers to capture their "broad qualities" while producing something it calls its own 1.

Google's Gemini still complies with the same request without hesitation. Perplexity refuses outright. Anthropic's Claude and Microsoft's Copilot comply but add qualifications, what No Latency's researchers described as suggesting "an awareness of the imitation issue" 1. Three responses across four models, each reflecting a different bet on where copyright liability begins.

That divergence is not a product preference. It is a legal posture. OpenAI is defending multiple copyright lawsuits from book authors who allege large-scale infringement through model training. One of those suits specifically cites ChatGPT's "uncanny ability to generate text similar to that found in copyrighted textual materials" as evidence 1. If ChatGPT no longer produces that imitation, the plaintiffs lose a demonstrable capability they built their case around.

OpenAI did not announce the change. An OpenAI spokesperson did not respond to Ars Technica's request for comment 1. The behavior was documented through two separate efforts. Ars Technica's own testing found ChatGPT refusing style-copying requests for both living and dead authors 1. A separate analysis by No Latency, published earlier in July, found the same refusal for living authors but found ChatGPT still complied with style-copying requests for deceased authors 1.

The copyright stakes are already producing billion-dollar outcomes. In July 2026, a court granted final approval of a $1.5 billion settlement in the Bartz v. Anthropic copyright case, a figure tracked closely by the Authors Guild, which has been litigating AI copyright issues on behalf of writers 2. For OpenAI, every demonstration of ChatGPT mimicking a plaintiff author's prose is a piece of evidence that could surface in a courtroom.

Under US copyright law, an author's style is not itself protected. What is protected is the specific expression of an idea. An AI-generated imitation becomes potentially infringing only if it is "substantially similar" to the author's actual copyrighted work 1. ChatGPT's new response, which says it can capture a similar feeling without copying a distinctive voice, reads like a deliberate attempt to stay on the safe side of that line.

George Washington University Law School Professor Robert Brauneis told Bloomberg Law that the cost and fidelity of AI imitation is what makes the current moment unlike anything before it: "We've never had a situation in which this personal style of individual creators could be imitated as well and as inexpensively as we now have with AI" 1.

There is a telling gap in OpenAI's own policy documentation. The company's DALL-E 3 image model already declines requests for images "in the style of a living artist," a restriction OpenAI advertises prominently 1. But OpenAI's December model spec, the document meant to govern model behavior, does not include any explicit prohibition on replicating written style 1.

For users, the shift creates friction. Writers who built workflows around style-imitation prompts report being locked out of capabilities that previously worked, with one Reddit user describing carefully crafted prompts that no longer function 1. The Authors Guild, in its own best practices guidance, urges writers not to use generative AI to mimic other writers' unique styles or voices, warning that doing so could expose them to claims of unfair competition or copyright infringement 1.

OpenAI's shift means the capability plaintiffs cite in their complaints may no longer exist by the time those complaints reach trial. If that strategy holds, every AI provider has a built-in escape hatch: change the product, let the old behavior become a historical footnote. Courts will have to decide whether a model's past capabilities or its current ones define what counts as infringement. Until they do, companies choosing between GPT, Gemini, Claude, and Copilot are picking a legal risk profile as much as a tool. Comply, hedge, or refuse. OpenAI has chosen. Google has not.