A look at how AI models encode a hidden, checkable pattern into everyday word choices — without changing what they say.
by Doris Booth
Founder of award-winning Authorlink.com, retired literary agent for bestselling authors. Early adapter and programmer of publishing technology.
Anthropic’s recent introduction of AI watermarking to its Claude platform has flummoxed the publishing industry. European Union’s transparency rules prompted Claude to launch the new mark on August 2, 2026. The company said in an interview with Euro News that it will work to add the mark to previous AI models retroactively.
Writers, agents, and publishers alike worry that if AI watermarks flag a finished text as inauthentic (whether or not wrongly labeled), reputations and incomes will crumble. The concern is widespread.
AI agents are embedded in almost every software program we access. Claude, Google Gemini, ChatGPT, and Meta’s Facebook each have more than one billion (yes, billion) active users a month churning out content with the help of AI. Three platforms alone collectively total about four billion individual users worldwide, equaling roughly eleven times the US population.
The Anthropic mark adds yet another test to the creative landscape. A literary work could be censored based on how the text was produced rather than on its quality.
Now more than ever, agents and publishers need to know whether AI merely assisted the author of a manuscript or generated most or all of the story itself.
Claude’s invisible statistical watermark can indicate the frequency with which an LLM writes text versus a human author. You, the reader or writer, cannot see or easily remove the mark. However, an analyst with a secret key can see how often an LLM was involved. Hopefully, the new approach will help publishing pros root out low-quality, mass-produced slop created with little human effort.
Unlike traditional ways of exposing fraud and plagiarism, a machine now mathematically calculates the “probability” that a work was AI-created. The detector uses the words Claude chooses along with a mathematical signature that can’t be broken or wiped clean without substantive changes to the text. The two-step process begins by recognizing a word or sentence that could have an arbitrary meaning, like “Today the sky appears [cloudy or grey]” AI knows which word it prefers, but it lets the human choose. One might think if the user doesn’t pick the preferred word, that would trigger the flag. The mark doesn’t care. It’s ok to pick the word you like. Behind the scenes, the tool has assigned a random number to the action, recording whether AI itself made a choice. Many numbers in the document show more AI activity than human interaction. More details are quoted from Claude below.
Interestingly, Claude’s enhanced AI watermarking tool (based on the SynthID-Text system) does not read actual content per se. The random mathematical algorithm measures statistical patterns in word choices to indicate that an AI model was involved, global economy authority Armando Roggio explains in a Practical Commerce post.
Claude is not the only platform employing digital sentries to weigh AI output against original human work. Google’s Gemini uses SynthID (developed by Google DeepMind) for text, images, audio, and video, alongside open C2PA metadata (for media files). Facebook parent Meta applies Stable Signature (for generative model tracing). Visible tags applied to generated content are labeled “Imagined with AI”. [1, 2, 3, 4]
Surprisingly, ChatGPT (OpenAI) has not yet embraced new watermarking out of concerns for false positives and robustness, or perhaps as a tricky business move. I’ll explain later. Moving from Claude to another platform like ChatGPT will break the watermarking pattern. One wonders if executives at OpenAI might bet that many Claude users will—at least temporarily—flock to ChatGPT to escape Claude’s statistical cop.
Wrongly accusing an author of fake or inaccurate information can taint the entire publishing industry and derail an author’s career.
Why Watermarking Matters for Writers
We all consume AI information from the Internet, knowingly or not. Watermarks muddy the lines between acceptable and unacceptable AI involvement in literature. The impact upon writers has grown more evident in recent months.
In July 2026, debut author Jerry Falade’s novel, Call Me, I’ll Hide the Body, won a $2 million deal with Macmillan’s Minotaur imprint. The work was suspected of being AI-generated, and Falade could not prove how the final manuscript was developed, though the author has vehemently denied the accusations.
More works have been flagged as signs of generative AI. In March 2026, Hachette Book Group cancelled the U.S. release of her horror novel Shy Girl and withdrew the book’s UK distribution. Ballard, too, denied writing the text with AI, claiming a freelance editor used AI without her knowledge.
Watermarking affects every writer’s creative process. Almost every agent or publisher who considers your work will run the text through an AI detection tool. While Anthropic touts its tool as better, mistakes or misjudgment and biases abound. AI still hallucinates and throws false positives. And, often mechanical rulings lack a deep understanding of the writer’s intent or pathways to the finished story.
So, to avoid being embarrassed, penalized or ruined, a novelist or essayist must build a strategy to defend the creative work.
Publishing professionals must distinguish purely mechanical (fake) storytelling from original human thought. Writers have the obligation to help by clearly declaring they will apply AI as an assistant or an intern, not as a literary mimic or terminator of human thought.
Let us understand some basic do’s and don’ts before developing an ethical creativity plan for ourselves.
What Is AI Watermarking
I asked Claude directly to simplify its watermarking definition. Here, in part, is its answer:
When Claude writes text, it picks each word one at a time from a list of likely options. Often, several words would work equally well — like choosing between “overcast” and “grey” in “the weather was cold and ___.” When there’s no clear best choice, Claude picks somewhat randomly.
Watermarking changes how that random pick happens. Instead of a totally arbitrary random choice, the “randomness” is generated using a secret key plus the recent words in the text. The word Claude picks still looks completely natural — but now the pattern of choices can be checked against that key. If the pattern matches, that’s evidence the text came from Claude.
A few key points:
-
-
- It doesn’t force Claude toward one word every time (it won’t always pick “overcast” over “grey”) — the choice still varies based on context.
- It doesn’t make Claude use weird or unlikely words. It only nudges between options Claude would already consider.
- It’s invisible to a human reader, but detectable to someone with the key.
-
In short: watermarking hides an invisible, checkable signature in Claude’s word choices, without changing how the text reads.
In many cases, the mark travels with the document, unbroken by a user copying and pasting. To eliminate Claude’s statistical and hidden text watermarks, you must rewrite, translate, or heavily paraphrase the text. Simple metadata and hidden-character scrubbers cannot remove pattern-based word choice markers. [1, 2].
What Writers Want to Know About AI Watermarking
I offer my take on the following questions, and welcome other writers’ comments. Admittedly, I asked AI the most frequent questions writers ask. The answers are mine, gleaned from extensive reading and many years in publishing and technical arenas.
Q: Will my human-written draft be flagged as AI just because I used an LLM to check my grammar or tighten a sentence?
A: No. Using AI for outlining, brainstorming, and structuring is okay, but the core text must be written by a human. When a secret key is used to decode the watermarks, it could indicate extensive AI generation in the document based on how many flags it generates. It measures involvement, not contribution. If you ask Claude to rewrite a single awkward sentence in a long essay, Claude will mark that sentence as AI-generated. But the tool doesn’t know you wrote the other 99 percent of the document. Passages under 150-200 words, or minor proofreading edits of human text, often don’t contain enough data to trigger a detection.
Q: How much text needs to be rewritten to strip out or invalidate a questionable or accidental watermark?
A: If you are putting text into a standard commercial detector (like Turnitin or Originality.ai) that doesn’t use secret watermarks, it outputs an overall confidence percentage. [1, 2] I’m guessing that the scores might be similar for Claude’s Watermarking, though it uses a different method to calculate how heavily the document relies on AI-written text. Please let me know what you experience.
| AI Score Range | Standard Method: What it Means |
| 0% – 20% | Highly likely human. Normal human writing variation. |
| 20% – 50% | Mixed signals. Usually indicates a human wrote it but heavily used tools like Grammarly or an LLM to rephrase sentences. |
| 50% – 80% | Likely AI-assisted. Contains significant chunks of cut-and-paste AI text. |
| 80% – 100% | Strong AI signature. The text perfectly mirrors the uniform predictability and low variance of pure AI output. |
Q: Does Claude’s watermark mean the AI created the core ideas, or just that it touched the final paragraph?
A: The watermark does not critique the core idea per se. It tests patterns to signal how often decisions were made within the document. In other words, if you prompt Claude to write a 300-page novel in Ernest Hemingway’s voice, and you actually publish it, you can expect the AI-generated text to be exposed, banished or perhaps send you to jail for fraud.
Q: Should a Writer Hide or Declare their use of AI?
A: Never hide your use of AI. Instead use a template below to create your own policy, and carefully document the development stages of your manuscript.
Q: What breaks a text watermark? I am fairly confident that these solutions are solid.
A: Break the AI pattern by
- heavy editing or rewriting,
- short passages under roughly 260-200 pages, or minor proofreading edits on existing human text.
- extensively removing obvious AI words like delve, moreover, or in conclusion
- blending a few AI-generated sentences into a much larger human-written document
- copying the entire document to another AI platform
- ensure that a minimum of 30 percent of the text is rewritten, heavily modified or expanded by you.
How Writers Can Maintain Integrity Against AI Scrutiny
Here’s a suggested strategy to openly disclose how and where you use AI in your work
- In a submission to an agent or editor, write a short, honest statement of how you view and use AI as a project for accuracy and refinement, not as a replace to human expertise or creative thought.
- Briefly state how you only use these activities to optimize your workflow:
- research and brainstorming ideas or story angles
- structure and outline
- text clarity or tone (minimally)
- internal summaries, case studies, or raw research
- Eliminate these activities in the manuscript
- Never copy and paste full AI output unless used as a direct quote
- Don’t use Application Programming Interface (API)/tools to generate mass pages at once
- Don’t rely only on AI generated and citations or quotes without fact-checking the source.
- When querying an agent or publisher, briefly include your experience, expertise, authoritativeness, and trustworthiness (EEAT). AI favors these credentials.
- You, the human, must verify all statistics, claims, and links within your story.
New submission guidelines in the age of AI
All creative content now should be clearly identified as AI-assisted (human written with AI input) or AI-generated. (AI produced from human prompts). Publishers want writers to be more transparent in how a manuscript was produced. The old guidelines, such as creating your manuscript with correct margins, font size and font faces, headers and footers, simply are not sufficient to appear professional.
The way that AI reads a document involves consistent tags or labels that show how different pieces of information are connected. You as a writer will need an organized framework called “schema” to describe your work so that a machine can find, understand, and distribute the contents to agents, editors, and the consuming public. You don’t need to write computer code to be effective. You only need to understand some standard labels and their general order. It is important to tell AI machines the nature of the content you submit by using carefully labeled tags. This practice is becoming more standard in machine language. You can see a list of standard attribution tags for publishing (different from marketing ) here: schema.org. However, don’t stress. You need only use a few. I’ve listed a few important tags below. You can also use a tool like Google Rich Results Test to improve your work’s machine readability. The easiest way to comply with today’s complicated technical standards is to feed part of your story or perhaps a summary into Google Rich Results Test and see what sort of output you receive. Just be sure to ask Google not to track your content or use it for AI training.
Below is a simplified template (schema) for general web search to help you develop a professional document. This is only a rough start. I welcome your feedback on your use of such tools.
If you did not use AI at all in your work, just state “none.” And be aware that the schema for academic work will be different. Bracketed text is only an example.
Content type: [creative work]
Headline/Book Title:
Author Name:
Description/category: [such as adult trade fiction, historical, memoir]
Transparency Statement: [This manuscript/content was originated by the author. AI was only used for administrative/editorial support. The core creative work and voice are exclusively mine. All narrative structure, character evolution, thematic elements, and final prose were entirely conceived and executed by the human author.]
Disclosure of AI Use:
Brainstorming and research, example: [I used Claude to outline historical terminology for Chapter x]
Editing and Line Polish, example: [I used ChatGPT, Grammarly or Microsoft Word Editor to proofread final drafts for passive voice and syntax clarity]
Administrative, example: [I used AI to format the book proposal, synopsis, and create the initial query pitch letter].
Organizational Policies Toward AI
Library of Congress. I am directly quoting the Library of Congress statement below, with a link to the site to ensure accuracy.
The U.S. Copyright Office (part of the Library of Congress) maintains that copyright law requires human authorship and does not protect works generated solely by artificial intelligence. [1, 2]
Core Policy and Registration Guidance
-
- Prompts Are Not Authorship: Text or visual prompts given to an AI generator are considered unprotectable ideas rather than expressive execution, meaning prompts alone do not grant the user authorship of the output. [1]
The Author’s Guild
The Authors Guild says “when you claim authorship in a work, it means that the words, thoughts, and imaginings are yours.” You’ll find recommended best practices on the Guild site.
Amazon Kindle Publishing Policy on AI-generated and AI-assisted content
“You must disclose content (text, images, or translations) that is created by an AI tool. You do not need to disclose AI-assisted work, such as using AI to brainstorm, check grammar, or refine text you already wrote yourself.” [1, 2]
How do you use AI?
Join the conversation. Feel free to use the comment box below for feedback, and you are welcome to share my thoughts. with friends and colleagues. Please credit Authorlink.com by including a link to this article.
About this article and my policy:
I used Claude Pro to research and properly format this article, but the thoughts are my own. The accompanying article image is from iStock. My philosophy as CEO of Authorlink.com is that writers must always be transparent about the use of AI, and that an experienced human must always initiate the core content, with mandatory human fact-checks of AI involvement in written material. Please keep writing and strive for authenticity in your work and your life. Stay human.
