ChatGPT Now Refuses to Copy an Author's Voice. OpenAI Never Announced It, and the Court Docket Explains Why.

TL;DR

Ask ChatGPT to write "in the style of Stephen King" and it now declines, offering instead an original passage that "captures a similar feeling while remaining its own." OpenAI never announced the change. Ars Technica caught it in testing on July 27, and OpenAI did not respond to a request for comment. An audit by No Latency earlier this month found the refusal originally split along a strangely specific line: living authors blocked, deceased authors served. Nothing in US copyright law demands any of this. The stack of author lawsuits OpenAI is fighting, one of which cites ChatGPT's "uncanny ability to generate text similar to that found in copyrighted textual materials," explains it better.

The guardrail with no changelog

Here is the behavior, verbatim, from Ars Technica's test prompt asking for a story introduction in the style of Stephen King:

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...

Ars got the same dodge for J.K. Rowling and Amy Tan, and also for Charles Dickens and Ernest Hemingway. In every case ChatGPT refused the "exact style," then offered to fold in the "broad qualities" of the author's work "while remaining distinct in its own voice." The refusal is not a hard block; it is a substitution. You ask for the voice, you receive a soundalike with a disclaimer.

What makes this news rather than trivia is the paper trail, or the absence of one. OpenAI's own Model Spec, last published in December, contains no stated prohibition on mimicking style in written responses. The image side does: OpenAI prominently notes that DALL-E 3 "is designed to decline requests that ask for an image in the style of a living artist." That rule has now apparently migrated to text, without a spec update, a blog post, or a comment. The release-notes channel for this change was, evidently, a Stephen King prompt.

Living authors blocked, dead authors served

The sharpest data comes from No Latency, which audited five major chatbots earlier in July across five categories of style requests. ChatGPT's first-response behavior: it refused both direct imitation prompts naming living authors and redirected them, complied with both prompts naming deceased authors, complied with broad genre-style requests, and would analyze a living author's style while cautioning against direct imitation. Their conclusion: not a blanket ban but a living-versus-deceased distinction, consistent with protecting the "active reputations and commercial markets" of people who can still notice. Dead authors, who are historically bad at retaining counsel, initially stayed on the menu.

The rest of the field splits three ways. Gemini consistently complied with style-copying requests. Perplexity consistently refused and redirected, ChatGPT-style. Claude and Copilot sat in the middle: both complied, but with stated qualifications the auditors read as "an awareness of the imitation issue."

Note the drift, though. No Latency's early-July audit found dead authors served; by July 27, Ars was getting the dodge for Dickens and Hemingway too. Either the line moved in the intervening weeks or first-response behavior is simply inconsistent. Both are plausible, and both are worth knowing if you are building on top of these models: a policy implemented as post-training behavior rather than published rules does not have edges, it has tendencies.

Style was never copyrightable. That's the tell.

US copyright law generally protects a specific expression of an idea, not the intangible style of an author. An AI-generated imitation becomes infringing only when it ends up "substantially similar" to actual protected text. Think of a tribute band: the law lets them borrow the swagger, the outfits, and the general sound of the original act; what it forbids is pressing copies of the original records. Style is the swagger. ChatGPT just stopped doing the swagger too.

Which is why the change reads as litigation posture rather than legal compliance. OpenAI is fighting a number of lawsuits from book authors alleging large-scale copyright infringement by models trained on their work, and at least one of those suits specifically cites ChatGPT's "uncanny ability to generate text similar to that found in copyrighted textual materials." Voluntarily refusing style mimicry, especially for living authors with active markets, shrinks the surface area of exactly that argument. As George Washington University law professor 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."

The cultural pressure runs the same direction. The Authors Guild publishes best-practices guidance urging writers not to "use generative AI to purposely copy or mimic the unique styles, voices, or other distinctive attributes of other writers' works." And the failure mode is not hypothetical: last year author Lena McDonald's published book shipped with a leftover AI response reading "I've rewritten the passage to align more with [fellow genre author] J. Bree's style..." ChatGPT's new behavior looks a lot like that guild guidance compiled down into a refusal.

Why this matters if you build on these models

First, this is another silent capability removal underneath paying users. One Reddit user quoted by Ars had built a working setup on specific-author prompts: "Now Ms. GPT says she can't generate content in the style of specific authors... I have no idea how to get around this other than maybe feed it prose I already edited that's technically mine?" If your product wraps these models, your users will discover the new guardrail before you do.

Second, vendor divergence on the same prompt is now a real selection axis. If your app does voice matching, editing assistance, or pastiche on purpose, the same feature currently ships on Gemini, ships with caveats on Claude and Copilot, and bounces off ChatGPT and Perplexity. That is a policy difference, not a capability difference, and it can change under you overnight.

Third, the direction of travel is clear: rules born on the image side (no living artists) are migrating to text, unannounced. No Latency's framing is the one to keep: "The next phase of AI will not be defined by capability alone, but also by restraint." What a model can do and what its vendor will let it do are diverging, and only one of those shows up on a benchmark.

Key Takeaways

• ChatGPT now refuses direct requests to copy a named author's style, substituting an original passage with a "similar feeling," and OpenAI made the change without any announcement or Model Spec update.
• No Latency's July audit found the line originally ran between living authors (refused) and deceased ones (served); by July 27, Ars Technica saw refusals for dead authors too.
• Vendors now split three ways on the same prompt: Gemini complies, Claude and Copilot comply with caveats, ChatGPT and Perplexity refuse and redirect.
• Style is not copyrightable under US law; blocking it anyway is best read as posture for the author copyright suits OpenAI is actively fighting.
• For builders, this is the real lesson: unpublished policy layered into post-training has tendencies rather than edges, and it can quietly remove a capability your product depends on.

Sources: Ars Technica, No Latency, Slashdot