OpenAI may be rewriting its legal defense one prompt at a time. ChatGPT is now refusing direct requests to mimic specific authors’ styles, offering instead to channel the ‘broad qualities’ of a writer while staying, as the model puts it, ‘distinct in its own voice.’ That’s a subtle shift, but in the context of active litigation, subtle is exactly the point.
As reported by Ars Technica, when asked for a story opening in the style of Stephen King, ChatGPT now responds that it can capture ‘atmospheric, character-driven horror and small-town dread’ but cannot write in King’s ‘exact style or closely imitate his distinctive voice.’ The model returned similar deflections for J.K. Rowling, Amy Tan, Charles Dickens, and Ernest Hemingway. An earlier analysis by No Latency found the same pattern for living authors, though ChatGPT apparently still complied with style requests for deceased ones at that point.
So why does the wording matter? In US copyright law, style itself is not protected. Only specific expression is. But if an AI-generated text becomes ‘substantially similar’ to a copyrighted work, that’s where infringement claims get real traction. OpenAI is currently fighting multiple lawsuits from book authors, at least one of which directly cites ChatGPT’s ability to produce text that mirrors copyrighted material. Switching from ‘I’ll copy this author’s voice’ to ‘I’ll evoke a similar feeling’ is a small but legally meaningful reframe.
George Washington University Law Professor Robert Brauneis put it plainly to 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 context is worth holding onto. This isn’t just an OpenAI policy tweak. It reflects a broader industry uncertainty about where imitation ends and infringement begins.
Not every major model handles this the same way. According to the No Latency study, Google’s Gemini complied fully with style-copy requests. Perplexity refused and redirected, similar to ChatGPT’s new behavior. Anthropic’s Claude and Microsoft’s Copilot both complied but added qualifications that signaled some awareness of the issue. OpenAI’s DALL-E 3 already declines image requests mimicking living artists, but the company’s published model spec from December contained no clear prohibition on style replication in text.
For some users, the change is already frustrating. One Reddit user complained that highly specific style prompts that previously worked now get blocked, with no obvious workaround. That friction is real. But for the authors who have spent years watching AI tools reproduce their voice on demand, this shift, however modest, is at minimum an acknowledgment that the problem exists.
Whether ‘capturing a feeling’ versus ‘copying a voice’ will actually hold up as a legal distinction remains to be seen. But OpenAI clearly thinks it’s a distinction worth making.
