In late July 2026 OpenAI quietly tightened ChatGPT so it stops honoring a request creators had leaned on for years: "write this in the exact style of [named author]." Where the model had previously refused only for living writers, it now declines the request for named, copyrighted authors whether they are alive or dead — telling a user, for example, that "Agatha Christie's works are still under copyright, so I can't provide text that closely imitates her distinctive style," then offering the "hallmarks" of the genre "while remaining distinct in its own voice." There was no announcement; the tech press surfaced the change through its own testing and the story spread from there. The obvious read is that a useful feature got taken away. The more useful read, and the one this guide takes, is that the restriction only touches one specific thing — borrowing a third party's copyrighted voice — and leaves untouched the thing a real content operation actually needs, which is generating in a brand voice you own. This guide separates those two ideas cleanly, explains exactly what changed and why the copyright lawsuits pushed OpenAI to draw the line here, makes the case that "write in the style of a famous author" was always a fragile strategy for anyone publishing commercially, shows why the inconsistency between AI providers is the deeper lesson, and lays out how to define and apply a voice that is genuinely yours — so a vendor's next policy swing cannot reach it.
Sometime around July 27–28, 2026, without any announcement, OpenAI changed how ChatGPT handles one of its most-used writing tricks. Ask it to "write in the exact style of" a named author and it now declines when that author's work is under copyright — and, crucially, it declines for dead authors too, where it used to comply. It does not slam the door; it redirects. Asked to imitate Agatha Christie it will say her works are still under copyright and it cannot closely imitate her distinctive style, then offer to write with the general "hallmarks" of the genre "while remaining distinct in its own voice." Stephen King gets the same treatment: the hallmarks of atmospheric, character-driven horror, but not King's exact style. The tech press surfaced the shift through its own testing rather than any OpenAI disclosure, and the change appears to be a quiet, policy-level tightening rather than a model update.
The easy story is "OpenAI removed a feature." The more useful story, and the one this guide is built around, is that the restriction lands on exactly one thing — borrowing a specific third party's copyrighted voice — and leaves the thing a real content operation depends on completely intact: generating in a voice you define and own. Those two ideas get blurred together under the label "voice-specific content," and separating them is what turns this from a loss into a prompt for a better strategy. For the news-desk account of the change itself, see ChatGPT now refuses direct requests to copy an author's style; what follows is the operator's version — what it means for anyone using AI to produce branded content, and what to do about it.
The precise shift is narrow but real. Earlier in 2026, testing showed ChatGPT would refuse to mimic living authors on copyright grounds but still happily produce text "in the style of" long-dead ones like Hemingway or Dickens. The late-July change extended the refusal to named authors whose work remains under copyright regardless of whether they are alive — so the line moved from "is the author living?" to "is the work copyrighted?" That is a meaningfully different and more defensible test, because copyright, not mortality, is what actually creates the legal exposure. Public-domain figures are less likely to trigger it; a modern, in-copyright author reliably does.
The form of the refusal matters as much as the fact of it. This is a disclaimer-and-redirect, not a hard block: the model states it cannot imitate the named author's exact style, then immediately offers to capture the broad qualities of the relevant genre in its own voice. So the capability to write atmospheric horror or a cozy whodunit did not disappear — what disappeared is the model's willingness to badge that output as a specific person's signature style. There was no press release, which is itself telling: OpenAI already applied comparable limits to visual style imitation in its image tools, so extending the same posture to text reads as a consistent legal stance being tightened quietly rather than a headline feature launch. The ChatGPT product page reflects the same late-July behavior shift.
The timing is not a coincidence. The change arrived in the middle of a wave of copyright litigation against OpenAI from authors and publishers, some of it explicitly citing the model's ability to generate text closely resembling copyrighted work. In that context, a chatbot that will reproduce a living, in-copyright author's "exact style" on request is a standing liability — it is a demonstration, on demand, of the exact capability the plaintiffs are pointing at. A chatbot that declines the named-style request and offers the general feeling of a genre instead is standing on the safer side of the argument. Read that way, the restriction is risk management, not a moral or technical judgment about style itself.
This framing is important because it tells you how durable the restriction is and how to relate to it. It is not that imitating a voice suddenly became impossible; it is that OpenAI decided the legal downside of offering it as a named feature outweighed the upside. That kind of decision can tighten further, loosen, or vary by author and jurisdiction, and it is made by one company about one product. A strategy that depends on a specific vendor's current appetite for legal risk is a strategy resting on someone else's liability calculus — which is a shaky foundation to publish a business on.
"Voice-specific content" hides two completely different activities. The first is imitation: prompting a model to write like a famous, identifiable author — Stephen King, J.K. Rowling, a named columnist — so the output carries that person's recognizable signature. The second is authorship in a defined voice: instructing a model to write in a tone and style you specify, built from your own material and rules, so the output sounds like you or your brand. ChatGPT's restriction hits the first and does not touch the second. You can still tell any assistant to write in short, punchy sentences, to be warm and plainspoken, to avoid jargon, to open with a concrete scene — because none of that is claiming a specific copyrighted person's identity.
Once you see the split, the restriction stops looking like a loss for serious creators, because almost no durable content operation should have been in the imitation bucket in the first place. Borrowing a famous author's voice for commercial output was always the weaker play: it is not yours, it does not build a recognizable identity for your own brand, and — as of late July — it is the exact thing a major provider will now refuse. Building your own voice is the play that compounds. That is the argument made at length in AI personality as a competitive advantage: a distinct, owned voice is the moat now that raw generation is a commodity, and this policy change just made the alternative more expensive to keep leaning on.
Set the policy aside and the imitation prompt has three problems that predate it. First, ownership: a piece written to sound like someone else's copyrighted work is legally and ethically shaky ground to publish commercially, and it exposes you to the same authorship and infringement questions the AI companies are now litigating. Second, identity: if your content sounds like a famous author, it sounds like them, not like you — you are spending effort making your audience remember someone else's voice instead of building recognition for your own. Third, brittleness: the whole approach depends on a model being willing to do it, which is precisely the dependency that just broke. A tactic that can be switched off by a silent policy update on a vendor's side was never infrastructure; it was a convenience you were renting.
The tell is that the creators most upset by the change are the ones who were using AI as a style Xerox rather than a drafting tool with a governed voice. If your workflow collapses because a model stopped impersonating a named author, the workflow had a single point of failure sitting inside another company's legal department. The fix is not to find a model that still allows it and hope it lasts; it is to stop needing permission to sound like someone you are not.
Different AI assistants handle the "write like a named author" request differently, and as of mid-2026 there is no shared standard — some decline, some comply with caveats, some comply freely. It is tempting to treat that as a shopping problem: just use whichever tool still does what you want. That is the wrong lesson, because every one of those postures can change without notice, exactly as ChatGPT's just did. A capability that exists today on one provider and not another, governed by shifting legal calculus rather than a stable spec, is not something to build a repeatable content process on top of.
The resilient posture is to depend on nothing model-specific for the part that defines your brand. Your voice should live in a written specification you control, applied through whatever model happens to be best or available, so a policy swing at any single provider is an inconvenience rather than an outage. This is the same principle behind not betting your reach on one platform's algorithm, applied to generation: own the asset (here, the voice spec), rent the commodity (the model), and never let the rented layer hold the asset hostage. And keep it clearly on the right side of impersonation — the discipline of instructing a model to write as a defined brand without claiming to be a specific real person is the non-impersonation prompt pattern.
Replacing "write in the style of X" with a real voice spec is straightforward and pays off beyond this one policy change. Start from your own corpus: gather a handful of pieces that best represent how you want to sound — your strongest posts, emails, or scripts. Read them for patterns rather than vibes. What is the typical sentence length and rhythm? Which words and constructions recur, and which do you never use? How do openings work; how blunt or hedged is the point of view; how much jargon is allowed? Distill those observations into a written spec: a short description of the voice, a do-and-don't list, a banned-word list for the tells and clichés you want stripped, and two or three gold-standard excerpts the model can pattern against.
That spec is categorically different from imitation because it is derived from material you own and rules you set, not from a named third party's copyrighted body of work — which is exactly why no provider needs to refuse it. Feed it as the governing instruction on every generation, then keep a human in the loop to add the specific number, example, or first-hand detail a model would never invent; that is what pushes the output past generic-but-on-brand into unmistakably yours, and it is the same discipline that keeps AI content from carrying the obvious AI tells. If you want a step-by-step version of building the voice itself, how to train AI to think like you walks the extraction process in detail.
For most people producing on-brand content, the practical effect of the restriction is small and healthy: the "in the style of a famous author" habit dies, and generation moves to a persona-governed model where the voice is defined once and applied everywhere. The creators who feel real pain are the narrow set who built output around impersonating specific writers, and their fix is the migration above. Everyone else should treat this as a nudge to formalize something they should already have — a documented voice — and to make sure it lives somewhere model-agnostic. The broader authenticity strategy this sits inside, keeping your work defensibly and specifically yours as AI scales, is laid out in AI content authenticity in social media.
The strategic takeaway is simple. A voice you borrow can be revoked by a vendor; a voice you own cannot. This policy change is a preview of a pattern that will keep repeating as the legal ground shifts under AI content — capabilities that touch someone else's intellectual property will keep getting fenced off, and the operations that thrive are the ones whose brand identity was never on the other side of that fence.
The honest framing first. Kompozy did not gain a feature because ChatGPT lost one, and it cannot and would not help you impersonate a copyrighted author — that is the thing the whole industry is now fencing off, for good reason. What Kompozy does is make the durable alternative operational: it treats your voice as an owned asset, defined once and enforced on every piece of content, through whatever model is doing the writing. If your strategy was already "sound like us," nothing here breaks; the restriction simply validated the approach the engine is built around.
Concretely, voice in Kompozy lives in a governing Persona Brief plus banned-word filters — a written specification of how your brand writes, built from your material and rules, not a prompt that names a famous author. Every generation runs against that brief, so the output is consistently on-brand by construction rather than by a per-prompt instruction someone has to remember, and it is your voice, which means there is nothing for a provider to refuse. Because the brief governs the model rather than being baked into any one of them, and generation can run on Claude or OpenAI with the option to bring your own key, a silent policy change or a differing stance at a single provider is an implementation detail, not a business risk — the asset that defines you stays put while the commodity underneath it swaps freely.
And the brief is not text-only. Kompozy is a generation and multi-platform publishing engine, so the same owned voice drives net-new formats a chat assistant cannot make — Persona Shorts and avatar video, carousels and infographics rendered to exact brand styling, Blog Articles, and Email Newsletters — every one of them governed by the same brief and routed through a per-post human review gate before it ships across the eight social platforms plus blog and email on Autopilot. That is the whole point of owning your voice instead of renting a famous one: it applies uniformly across everything you publish, on every surface, without asking any model for permission to sound like a person you are not. Explore Kompozy to see the persona-governed generate-review-publish flow.
ChatGPT's late-July 2026 tightening looks like a lost feature and is really a clarifying one. It restricts a single thing — writing in a named, copyrighted author's exact style, now for dead authors as well as living ones — as a hedge against the copyright suits piling up around AI training and output. It does not restrict generating in a voice you define and own, which is what any serious branded-content operation actually needs. "Write in the style of a famous author" was always the fragile play: not yours, not distinctive to you, and dependent on a vendor's shifting appetite for legal risk. The response is not to hunt for a model that still allows it, but to codify a voice built from your own material and rules, keep it model-agnostic, and apply it consistently with a human in the loop — so the next policy swing, whenever it comes, cannot touch the thing that makes your content recognizably yours.
In late July 2026 OpenAI quietly updated ChatGPT so it declines direct requests to write in the exact style of a named, copyrighted author — and, unlike before, the refusal now applies whether the author is living or dead. Rather than refusing outright, it redirects: asked to imitate Agatha Christie or Stephen King, it says it cannot copy their distinctive style because the work is under copyright, then offers to write with the broad "hallmarks" of the genre while staying "distinct in its own voice." There was no blog post; the tech press surfaced the behavior through its own testing.
The change lands in the middle of mounting copyright litigation against OpenAI from authors and publishers, some of it citing the model's ability to generate text closely resembling copyrighted work. A system that reproduces a living or in-copyright author's "exact style" on demand is a legal liability; one that offers the general feeling of a genre while staying in its own voice sits on the safer side of that line. The restriction is best read as risk management, not a technical limit — the model can still write well in a described tone, it just won't badge the output as a specific person's style.
No — and that distinction is the whole point. The restriction targets borrowing a third party's copyrighted voice ("write like this famous author"), not generating in a voice you define and own. You can still instruct any assistant to write plainly, warmly, bluntly, or in whatever tone you specify, and you can build that instruction from your own published material and rules. What is discouraged is renting a named author's identity. A brand voice derived from your own corpus and style guide is unaffected and, unlike a borrowed style, is actually yours to use commercially.
Not identically, and that inconsistency is itself the lesson. As of mid-2026 the assistants differ in how strictly they handle "write in the style of a named author," with some declining, some complying with caveats, and some complying freely — and any of them can change without notice, exactly as ChatGPT did. That means a workflow built on one provider's current willingness to imitate a specific voice is one silent policy update away from breaking. Building on a voice you own, applied through whatever model you choose, removes that dependency.
Codify your own voice and generate against it. Pull representative samples of your best writing, distill the recurring patterns — sentence rhythm, vocabulary, what you never say — into a written spec with a banned-word list and a few gold-standard examples, and feed that as the governing instruction instead of a famous name. The output is on-brand, defensible commercially, and portable across models. Then keep a human in the loop to add the specific details a model cannot invent, so the voice reads as a real person rather than a generic imitation.
In late July 2026 OpenAI quietly tightened ChatGPT so it declines direct requests to write in a named, copyrighted author's exact style — living or dead — offering the "hallmarks" of a genre instead, amid mounting author copyright suits. The restriction touches borrowing a third party's voice, not building your own. For creators who generate branded, voice-specific content, the durable response is to stop renting a famous style and instead codify a brand voice you own and can apply consistently across models and platforms.
Get started → · ← All guides · Compare Kompozy vs other tools