If you’ve been searching for an AI writing tool that preserves your voice, you’ve probably noticed most current tools don’t. The voice is gone by chapter three. The prose feels polished but doesn’t sound like you. Two arxiv papers from the last three weeks just explained why, and showed that the same underlying technology can also do the exact opposite.
A new paper from researcher Robert Dilworth, dated April 22, showed that you can hide your writing voice from AI by adding invisible characters to about a third of your words. The text reads completely normally to humans. AI can no longer tell who wrote it. The technique works.
That paper landed two weeks after a different one, by Berkeley researcher Tom van Nuenen, that proved AI tools flatten your writing voice when they revise your work, even when you tell them to preserve it. Same underlying tech (AI can measure how a writer writes). Opposite goal: one paper hides voice, the other showed how easily voice gets lost.
The AI writing tool space has split in two. Both halves use the same underlying tech. One half wants to make your writing voice more recognisably yours, more durable across a long project, more resistant to AI flattening. The other half wants to make your writing voice impossible to detect, impossible to identify, invisible to AI. bookmoth (which I built) is in the first half.
The split matters because it tells you which AI writing tool category fits your use case, and why most of the current frustration with AI writing tools comes from being in the wrong category for your needs.
What’s the difference between AI writing tools that preserve voice and ones that hide it?
The difference is in how each tool actually works. Tools that preserve voice (like bookmoth) take samples of your writing and turn them into a rule the AI has to follow every time it generates new prose. Tools that hide voice (like StegoStylo) add invisible characters to your writing that confuse AI detection systems. Both rely on the same way of measuring how a writer writes. They just use that measurement for opposite purposes.
Both directions use the same trick. AI can break a piece of writing down into measurable patterns: which words you reach for, how long your sentences run, whether you use contractions, how often “I” appears, the rhythm of your paragraphs, your punctuation habits. With enough of your writing fed in, those patterns form a kind of fingerprint. The fingerprint is detailed enough to identify you as a writer (which is what Kelsey Piper proved with Claude Opus 4.7 last month) and detailed enough to track what happens when AI revises your work (which is what van Nuenen measured directly two weeks ago).
Once you have that fingerprint, there are two things you can do with it. You can make the AI follow it as a rule on every generation, which keeps your voice intact across a long project. Or you can deliberately disrupt the fingerprint by inserting invisible characters, which breaks the AI’s ability to identify the writer without changing how the text reads to humans.
The split isn’t ideological. The same underlying research could go either way. bookmoth chose preservation because that’s what novelists need. StegoStylo chose hiding because that’s what privacy-conscious writers need. Both are right for their use cases.
Can AI hide your writing voice from automated detection?
Yes. The 2026 StegoStylo paper by Robert Dilworth showed that adding invisible characters (zero-width Unicode characters) to 33% or more of your words is enough to defeat the AI systems that try to identify writers. The text still reads completely normally to humans. The fingerprint just becomes statistical noise to the AI.
The setup: take a sample of writing where you know the author. Add invisible characters into a chunk of the words (the characters don’t show on screen, but the AI reads them as part of the text). Run the result through systems that try to identify the author from writing style. Measure how often the systems get it right.
The finding: at 33% coverage, the systems can’t identify the author any better than random guessing. The author’s fingerprint is gone. The paper presents this as part of a tool called TraceTarnish. The technique is simple to implement and the results hold up across tests.
What this means: if you have reasons to publish writing without anyone being able to link it back to you (whistleblowing, source protection, journalism in dangerous places, ghostwriting under another name), the technology now exists. It works, it’s been tested, and the method is publicly documented.
[Source: Robert Dilworth, “StegoStylo: Squelching Stylometric Scrutiny through Steganographic Stitching,” arxiv 2601.09056, latest revision April 22, 2026.]
Why would a writer want to hide their voice from AI?
Hiding your writing voice is a real privacy need. It serves writers protecting sources, working in dangerous places, ghostwriting under different names, or who don’t want their existing writing used to identify them in future.
These aren’t edge cases. They’re growing categories as AI’s ability to identify writers improves.
A whistleblower’s writing style can be matched against their public writing to identify them, even when their leak is anonymous. Hiding tools defeat that match. Dissidents and political writers in countries where being identified is dangerous. Hiding as a safety tool. Writers who use different names for different audiences (commercial vs literary, romance vs thriller) and don’t want their work linked across genres. Plaintiffs or defendants whose writing might be used as evidence have a real reason to understand how identification works. Writers who don’t want their existing prose to be used to fingerprint future text written about them or in their voice.
These needs are legitimate, and they’re growing. As AI’s ability to identify writers from their writing gets more accurate (which the Piper experiment showed it has), the demand for tools that defeat identification grows alongside it. The StegoStylo paper is one of the first openly published responses.
What’s the best AI writing tool for preserving voice across long-form work?
The best AI writing tool for preserving voice across a long project is one that doesn’t rely on prompts to do the job. Most current tools (Sudowrite, NovelCrafter, Claude, ChatGPT) put voice instructions in the prompt, which a 2026 Berkeley study showed doesn’t work. bookmoth is built differently. It uses your writing samples as a binding rule, not a request the AI can drift away from.
The reasons writers need voice preservation are real and growing.
A novelist’s voice is the brand. AI drafting tools that flatten the voice across chapters destroy what makes the writer worth reading. Columnists, opinion writers, longform reporters whose readership comes for them specifically. The byline matters because the voice matters. Newsletter writers, blog owners, anyone whose readership identifies with their specific voice and tone. Translators maintaining a single translator’s voice across multiple authors’ books. Documentation teams that want every doc to read like the same person wrote it, regardless of who actually drafted it.
The Berkeley paper from two weeks ago proved that prompt-based AI writing tools cannot preserve voice. The drift is built into how they work, even when you explicitly tell the AI to preserve your voice. By the third or fourth paragraph, the AI’s default voice starts creeping back in. By chapter three, your voice is gone. For writers who need their voice preserved, the prompt-based approach isn’t a viable option. The other approach is the only one that holds up across a long project.
bookmoth is built specifically for this. It reads your writing samples once and turns them into a profile of how you write, then makes the AI follow that profile as a hard rule every time it generates new prose, not as a request the AI can drift away from. The profile captures the fundamentals from how to find your writing voice: your rhythm, your structural tendencies, the word choices that show up when you stop paying attention.
Is bookmoth a privacy tool or a voice preservation tool?
bookmoth is a voice preservation tool. It is not a privacy tool. It exists to make sure your writing stays recognisably yours through long-form AI-assisted writing, not to hide your identity from AI detection.
If you’re trying to publish under a different identity from your existing writing, bookmoth is the wrong tool. Hiding tools like StegoStylo are the right category for that. If you’re a novelist drafting your sixth novel and want it to sound like books one through five, bookmoth is right.
How bookmoth works: read your existing writing, build a profile of how you write across multiple dimensions (your rhythm, your sentence shape, the words you reach for, your structural habits, how your dialogue sounds), then make the AI follow that profile as a hard rule every time it writes new prose. The rule lives outside the prompt. The AI can’t drift away from your voice the way it does in Claude, ChatGPT, Sudowrite, or NovelCrafter, because the rule isn’t in the layer the AI’s defaults can override.
bookmoth and the hiding tools aren’t competing. We’re not building competing products in the same market. We’re building opposite products for opposite needs, on the same underlying tech. The contrast actually clarifies what each tool is for. If you’ve read both papers and can articulate which side serves your use case, you’re already further along than most writers using AI tools today.