ChatGPT Stops Cloning the Style of Famous Writers

OpenAI has changed how ChatGPT behaves: now, when asked to write a text “in the style” of a specific famous writer, the chatbot declines to imitate them directly. Instead, it offers to draw on the author's “general traits” while staying “in its own voice.”

When Ars Technica put it to the test with a request to write the opening of a story in the style of Stephen King, ChatGPT replied that it was happy to reproduce the hallmarks of atmospheric horror—a small-town feel and a focus on characters—but couldn't imitate King's own “exact style” and “distinctive voice.” The bot gave similarly evasive answers for other authors, both living (J.K. Rowling, Amy Tan) and dead (Charles Dickens, Ernest Hemingway). Curiously, an earlier No Latency study recorded refusals only for living writers.

Why this matters

At first glance, the distinction between “exact style” and “general feeling” seems contrived. But it's precisely this subtlety that may prove legally significant. OpenAI is facing a whole string of lawsuits from book authors who accuse the company of mass copyright infringement in training its models. One of the suits explicitly cites ChatGPT's “uncanny ability” to generate text resembling copyrighted material.

In the United States, copyright protects the specific expression of an idea, not an author's abstract style. A neural network's pastiche, however, can amount to infringement if it turns out to be “substantially similar” to the original. As George Washington University law professor Robert Brauneis noted, for the first time in history the personal style of an individual author can be imitated so well and so cheaply.

In its guidelines, the Authors Guild urges writers not to use generative AI to deliberately copy someone else's voice—beyond the ethical questions, it invites unfair-competition suits. There have already been cases of this kind: last year, writer Lena McDonald was criticized after a clearly machine-generated fragment ended up in her book, reading “I've rewritten the passage to align it more with J. Bree's style.”

Who does what

Different models behave differently. According to No Latency, Google Gemini dutifully carries out requests to copy a style, whereas Perplexity AI, by contrast, refuses and redirects them—as ChatGPT now does too. Anthropic's Claude and Microsoft's Copilot comply, but hedge their answers with caveats that betray an awareness of the problem. For image generation, OpenAI has long stated that DALL·E 3 rejects requests in the style of living artists, but no such clear-cut prohibition had existed for text until now.

Some regular users are already unhappy. On Reddit they complain that carefully honed prompts have stopped working, and that it's now unclear how to achieve the old results—short of feeding the bot their own, already-edited prose.

Source: Ars Technica