Your Manuscript, AI, and the Invisible Ink

Michael Ireland

Sep 21

7 min read

What Every Author Needs to Know About the New AI Watermarks

This blog is subject to our Notice to Readers.

If you've been anywhere near a writers' group lately, you've probably heard the rumor: AI tools are now secretly stamping everything they touch. Some people are whispering about it like it's the plot twist in a thriller. Others are selling apps that promise to scrub your manuscript clean.

Let's take a breath together. Yes, something real has changed. No, your book is not doomed. And the fix is refreshingly simple. It's the kind of habit a good editor would recommend even without the AI part.

Here's what's actually going on, what it means for you, and how I handle it with my own clients.

What changed this summer

On August 2, 2026, a new transparency rule under the European Union's AI Act came into effect. It requires companies that make AI writing tools to mark what their tools produce so it can be identified as AI-generated. Anthropic, the company behind Claude, signed on and began building an invisible watermark into the text its newer Claude models write. Even though the rule is European, the watermark applies everywhere Claude is used, including here in North America.

Now, when most of us hear "watermark," we picture a faint logo stamped diagonally across a photo. This isn't that. There are no hidden characters, no secret code tucked between the words, and nothing you could spot if you squinted really hard at the screen. Instead, the watermark lives in the pattern of word choices Claude makes as it writes. Think of it like a handwriting quirk that only a special machine can recognize. Readers can't see it, it doesn't change the meaning, and it doesn't identify you. It only says "Claude had a hand in this," not that it was your hand was holding the pen.

(Nobody is reading your memoir chapter about your awful Aunt Doris and reporting you to anyone. Promise.)

The letter and the envelope

There are actually two kinds of marks to understand, and the easiest way I know to explain them is with an old-fashioned letter.

The letter is your words, the actual text of your book. The watermark I just described lives here, but only in words Claude itself wrote. If Claude drafts a paragraph, that paragraph carries the pattern. Your own words don't, no matter how many times Claude has read them.

The envelope is the file itself, your Word document. Every file carries hidden information you don't normally see: who created it, when it was saved, and what software touched it. AI companies can also add a provenance stamp to certain types of files their tools produce. It's basically a label on the envelope that says "this file came from an AI tool," and it can be there even if nearly every word inside is yours.

So here's the key question: whose envelope is your manuscript in?

The golden rule: your master file stays yours

Every author should have a master file, the one true copy of the manuscript that all your changes go into. (Some of us name ours things "FINAL," then "FINAL FINAL," then "FINAL FINAL REALLY THIS TIME." I call mine PF for Penultimate Final and UF for Ultimate Final.)

Here's the good news that should let you exhale: uploading your manuscript to Claude so it can read and review it does not mark your manuscript. Claude works from a copy, and the file sitting on your computer never changes. If you've already uploaded chapters for feedback, you're fine.

What matters is what you do with the file that comes back from Claude.

If you ask Claude to "fix my manuscript" and it hands you a new, corrected version, that new file came from Claude. If you adopt it as your master and keep working in it, you've moved your book into Claude's envelope. That's the one habit you should break if you can.

Instead, ask Claude for a "human fix" file. That's simply a separate list of suggested changes showing where each issue is, what the text says now, and what it should say. Then you open your own master and make the changes yourself. Yes, it takes a little longer. It's also how editing worked for a few centuries before AI showed up, and you'll know your book better for having made each change by hand.

If you only remember one line from this post, make it this one: treat anything that comes back from Claude as a suggestion, and editorial note, never as your manuscript.

What about small fixes and new writing?

A few commas and corrected typos give a watermark detector almost nothing to work with. There's no creative word choice involved in spelling "definitely" correctly. (Although, let's be honest, "definately" has fooled the best of us.) Short snippets of a sentence or two usually don't contain enough of a pattern to measure either.

New writing is different. If Claude drafts a transitional paragraph, a chapter summary, or a back-cover blurb and you paste it in word for word, that text carries the mark. Light editing won't remove it. Substantially rewriting it in your own words will, and honestly, that's what you should be doing anyway. It's your book. It should sound like you, not like a very polite robot.

You may see apps advertised that claim to "remove" AI watermarks. I'd save your money. As of this writing, nobody can prove those apps work, because the official detection tools haven't been released to the public yet. More importantly, they miss the point, which brings me to my next topic.

A clean file is not a secret

Keeping your master file clean is about good file habits and protecting your ownership of your work. It is not about hiding anything.

Many publishers, agents, and contests now ask whether AI was used in writing or editing a manuscript. Those questions are about what happened, not about what a detector can find. So before you submit, read the contract or submission guidelines carefully and answer honestly. If you used AI to check your punctuation, say so. Many publishers see AI-assisted proofreading very differently from AI-written chapters, but they want to hear it from you rather than discover it later.

What about ChatGPT and other tools?

Anthropic, the owner of Claude, has been clear about how its marking works. Other companies are moving in the same direction under the same European rules. As of this writing, Google already marks some text from its Gemini models, and OpenAI marks images and audio made with ChatGPT and has said it intends to do the same for text.

The details will keep changing, but your habit doesn't need to. Whatever tool you use, keep one master file that no AI has ever handed back to you, take the AI's suggestions as notes, and make your changes by hand.

What about ProWritingAid?

Many of you use ProWritingAid (PWA), and it's one of the questions I get asked most. The good news is that most of what PWA does is fine. Its grammar, spelling, punctuation, and style suggestions, along with big-picture reports like Chapter Critique and Manuscript Analysis, analyze your writing and suggest changes rather than writing new text for you. Accepting a suggested comma or a corrected spelling in your own master doesn't leave a watermark, for the same reason I mentioned earlier: there's no word choice for a watermark to live in. And because PWA works right inside your own document, your envelope stays yours.

The two features to watch are on PWA’s Sparks and Rephrase. These use generative AI to write new sentences for you, and PWA hasn't said publicly which company's AI powers them. If it's one that watermarks its output, any sentence you accept could carry the mark. Retyping the suggestion word for word doesn't change that, because the mark lives in the words themselves, not in how they got onto the page. (I know. Sneaky, right?) (Note: This is new since August 2026).

So treat Sparks and Rephrase the same way you'd treat a Claude draft. If you like the idea, take it, then write it in your own words. It'll sound more like you anyway. And if a publisher asks whether AI helped shape your text, an idea borrowed from Sparks or Rephrase is worth mentioning, even if every word is yours.

One last PWA tip: AI detectors, such as Turnitin, aren't the same thing as watermarks. They guess based on writing style, and PWA itself warns that they sometimes flag writing that’s been edited with PWA even when no AI writing features were used. That's one more good reason to be upfront about the tools you've used.

The short version

Uploading your manuscript for review is safe. Don't adopt a file an AI hands back as your master. Ask for a human fix list and make the changes yourself. Rewrite anything an AI drafts in your own voice. And be honest when a publisher asks.

That's it. No panic required.

One final note. If you’re using any AI to work on your manuscript, be sure your personal Settings and Preferences are set so that your AI app does not use your document to train AI, does not share your document with others, and does not use location metadata. You’ll find these options in your settings. In Claude, it’s under Privacy – Preferences. Turn off the options at “Location Metadata” and “Help Improve Our AI Models.”

If you'd like (human) help with your own manuscript, I'd love to hear from you. You can get in touch through my website anytime.

Happy writing, and remember, back up that master file!