AINEWS
ChatGPT now refuses to mimic named authors, but the 'similar feeling' output it offers is what pending book author copyright suits against OpenAI actually target.
Ask ChatGPT to write like Stephen King and the model will hand you a refusal. "I can definitely write with the hallmarks of atmospheric, character-driven horror and small-town dread, but I can't write in Stephen King's exact style or closely imitate his distinctive voice. Here's an original opening that captures a similar feeling while remaining its own…" That second sentence is the one the courts will be reading.
The behavior change, first reported by Ars Technica last week, is not a safety feature. It is OpenAI redrawing the line where copyright applies, in real time, while a stack of book-author copyright suits is moving toward trial. And the "similar feeling" output the model still produces on request is the part those suits are built to attack.
Robert Brauneis told Bloomberg Law: "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." That is the legal frame the refusal does not address. The model is no longer offering to imitate King's style. It is offering to imitate King's feeling, at the same cost, and the doctrine the lawsuits will lean on is the one that asks whether the result is substantially similar to a protected work — not whether the user said the author's name out loud.
The dodge is consistent across authors. Ars Technica's in-product test found ChatGPT refusing author-style mimicry for living writers (J.K. Rowling, Amy Tan) and for deceased writers (Charles Dickens, Ernest Hemingway). That is a change from a No Latency analysis earlier in the year, which found ChatGPT still complying with style-copying for deceased authors. OpenAI declined to comment. The closest public company anchor is the OpenAI Model Spec updated December 18, 2025, which is where style and imitation rules live when the company bothers to publish them.
The legal stakes are concrete. One pending suit specifically targets ChatGPT's "uncanny ability to generate text similar to that found in copyrighted textual materials." That language is the lawsuit's allegation, not a generic complaint, and it lands on the model's "feeling" output as easily as on a direct copy. US copyright doctrine protects specific expression, not intangible style, so a clean feeling-versus-style line might look like a legal win. The counterargument is that an AI stylistic imitation can still be infringing if the output is "substantially similar" to a protected work, and "atmospheric, character-driven horror and small-town dread" is doing more than a passing impression of King's voice. "Similar feeling" is a product description. "Substantially similar" is a legal test. Those are not the same line, and OpenAI is the one drawing them.
The workaround is also the test case. A direct style-mimicry request is now a visible refusal; the "original opening that captures a similar feeling" is the output the user can actually use. If that output is the new product surface, it is also the new infringement surface. Courts will not be persuaded that the boundary between "King's exact style" and "the feeling of atmospheric, character-driven horror" is a meaningful one when the user is asking for King by name and getting prose in the same register, on the same topic, in the same paragraph shape.
What to watch: the active book-author suits move toward discovery, where OpenAI will have to defend the Model Spec's rules in deposition, and where the "similar feeling" output becomes a piece of evidence. The refusal is a litigation position dressed as a product update, and the lawsuits are the audience it was written for.