ChatGPT no longer writes in the style of your favorite authors.

For three years, imitating the style of a famous novelist was one of ChatGPT’s favorite tricks, and one of the most commonly used (writing forums are full of ready-made instructions for imitating this or that bestseller). The exercise has just come to an end, without announcement or explanation, as if the chatbot’s publisher hoped no one would notice anything.

The chatbot declines, but keeps a plan B up its sleeve

The tests conducted over the past few days all point in the same direction.Ask for an opening scene in the style of Stephen King, and the machine responds that it cannot imitate either his “exact style” or his “distinctive voice”. Instead, it offers an original text built around small-town dread and psychological tension. The refusal applies to the living (J.K. Rowling, Amy Tan) as well as the dead, including Agatha Christie, Charles Dickens, or Ernest Hemingway. For the British novelist, the machine even cites works “still protected by copyright”.

Read also: Scanning books to train your AI: It’s legal in the United States, according to this unprecedented court ruling

However, an independent audit published a few weeks earlier showed that the chatbot was still compliant with requests targeting deceased authors: the lock was therefore put in place in recent weeks, silently. OpenAI, contacted by Ars Technica (which uncovered the issue), did not comment. Paying subscribers who had been using the feature to feed their own novels have already begun grumbling on the forums (and one admits to smiling at the thought of that romance novelist who had left her instruction—asking for a passage to be rewritten in the style of a fellow author of the genre—embedded in a book published last year).

The restriction is also less sweeping than it appears. Describe the recipe without saying the name (oral narration, psychological tension, a small town hiding a secret), and the machine complies without hesitation.One has to believe that only cybersecurity deserves genuine safeguards

Four artificial intelligences, four different sets of rules

American law, however, offers no clear answer: it protects the expression of a work, not an author’s style as such. An imitation constitutes infringement only if the result proves substantially similar to the original. This ambiguity has not prevented lawsuits from piling up against OpenAI, from The New York Times to leading novelists, the former having filed suit as early as 2023.

Why such precautions on OpenAI’s part? The reason lies with its biggest competitors. Anthropic has agreed to pay $1.5 billion to American authors, in a settlement definitively approved by the court on July 20. Yet the nuance that many summaries gloss over is worth considering. The judge had ruled that training on books was covered by American fair use: it was the pirated downloading of millions of copies that cost the company a fortune. In other words, the legal danger does not concern what the models learn, but what they were made to read and what they regurgitate. Hence the fresh attention now being paid to their outputs.

Competition, meanwhile, clearly did not receive the same memo.The same audit compared the house policies: Gemini complies without hesitation, Claude and Copilot comply by adding a warning, and Perplexity turns everyone down, living or dead. Four different policies for the same legal question, which above all says one thing: nobody knows where the line lies, and everyone sets the slider according to their exposure to the courts. On the French side, style is not any more protected as such, but moral rights and parasitism give authors angles of attack that fair use ignores.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Ars Technica