Across the writing and publishing world there’s a huge degree of consternation about the use and abuse of AI in the writing of books. In the fiction space some high-profile authors have had their contracts, advances and books cancelled. In the non fiction, self publishing world Amazon is now limiting the number of books a person can upload to two per week (yes you read that right) in an effort to counter AI-generated slop and fraud Authors are seeing their books ‘stolen’ and put on Amazon with money going to someone else). LinkedIn is trying to get its members to call out what they think is AI-generated content, universities around the world are trying to work out if and how they detect AI-generated writing.
Separately, the European Union has enacted transparency mandates such as Article 50 of the EU AI Act. This Act started on 2 August and enforces mandatory transparency and disclosure rules and it applies universally across all risk tiers of artificial intelligence. Providers (read writers for our purposes) must disclose direct AI interactions unless obvious and ensure machine-readable marking for generative synthetic content.
AI companies have had what is commonly referred to as watermarks – an embedded code that survives rewrites, editing and more – although I having read more about this I think AI DNA would perhaps be more accurate. But let’s stick with ‘watermarking’ as that’s the most common term.
What is a watermark in relation to AI content?
One description is as follows: An AI text watermark is an imperceptible statistical pattern woven into word choices as a large language model generates text
As large language models (LLMs) like Claude, Gemini, Chat GPT predict the next word, they subtly bias statistical choices toward or away from certain acceptable synonyms using a secret rule or key. At almost every step, a language model has several near-equal word options that all fit. Watermarking nudges the model toward a specific subset of those options, following a hidden rule. Over a full passage of content, that nudge builds a statistical pattern.
A reader never notices, because each choice was already a natural fit. The meaning, tone, and quality stay the same. Only the distribution of choices shifts, and only a detector tuned to the rule can see it.
Because the pattern lives in the word choices themselves, it survives a plain copy and paste. Moving the text into an email or a document does not strip it. As declaude.org notes: they [watermarks] work because they don’t live in the characters at all. They live in the choices between words.
The watermark indicates that a model processed or wrote the text, but it cannot measure the extent of human contribution, original ideas, or minor editing.
Daniel Susser, a researcher and philosopher at Cornell University who studies AI ethics and policy notes that AI text watermarks don’t conclusively convey whether something was written by a person or by an AI model. Instead, the watermarking provides statistical evidence that suggests whether AI was involved. “That evidence is not a definitive binary, yes/no signal,” he said. “It’s a signal that has to be interpreted.”
The longer the content – like a LinkedIn Article, a white paper or a book – the ‘better’ the detection as there’s more text to sample.
What if I’ve written my book but used grammar and spelling tools?
A found mark really means it’s been ‘processed by’, not written by AI. Anthropic’s documentation notes that human text proofread or translated by Claude picks up the mark.
What if I’ve trained my AI on all my original content?
I know several people who have loaded all their writing, research papers, notes, presentations and books into their AI and then direct it to generate new articles and books. Would this content be watermarked? After all a human wrote all the original content.
The answer is YES as it has been processed using AI, and the subtle changes to what the human has written – the word choices the AI makes – can be ‘seen’ by those who have the ‘hidden key’.
Remember the watermark is only saying the content has been processed by AI, it’s not measuring the extent of human contribution, the provenance of the original ideas or a record of your writing process.
How should you disclose what AI you have used to write with?
There’s a current debate in publishing about how authors should disclose their interaction and use of AI. It’s by no means settled so it’s really going to be up to each person to find a descriptor that works for them. What I would say though is that it’s important that you disclose something. In this stage of the AI content/human content war no disclosure is tantamount to your writing being regarded as AI generated. Sad but true.
The three most common descriptors being used are:
- AI generated
- AI assisted
- Human authored
And, it’s important to make another distinction. AI detection and authorship are different questions. AI detection analyses characteristics of text. Authorship concerns who made the creative decisions behind the work. And this is where it gets pretty fuzzy as it then comes down to the extent to which you have used AI to write and how you have directed it. You might think you’ve got the balance right but a detection tool might have another view entirely. It gets even messier if you’ve loaded all your writing into your favourite AI (pre 2022 writing included) and you direct your AI to undertake a range of different reworks, revisions, structural changes and more…you are making the key decisions on your own writing…but it could now, potentially, be labelled as AI generated.
How good (or bad) are the AI detection tools?
There are various detection tools in the marketplace and some are better than others. I have not tested them. If you head across to reddit there’s huge amounts of information and commentary about the options that are out there. (type in watermarks in writing)
Ones I am aware of are Proudly Human, https://www.proudlyhuman.org/. I use this for my company history books. Some others are: Books By People https://booksbypeople.org/; a ‘human-authored certification by the US writers group Authors Guild https://authorsguild.org/human-authored/ ; Librinova, a French self-publishing company and literary agency. https://www.librinova.com
Are there any other options?
People have asked me about file provenance. File provenance is signed data attached to a file.
C2PA is a file provenance protocol, an open standard from the Coalition for Content Provenance and Authenticity. Backed by Adobe, Microsoft and others, C2PA records a cryptographically signed history of where a file came from and how it was edited. Anthropic uses both. Claude adds an invisible text watermark to generated text, and it attaches signed C2PA metadata to supported files like .png, .jpg, and .svg images.
A text watermark rides inside the language and survives a copy-paste. C2PA metadata sits alongside the file and is easy to remove.
And then there’s a whole world of tools purporting to be able to de-watermark your writing. Again, I have not personally tested any of these…as I don’t use AI to write.
Where does all this watermark detection and provenance proving leave us?
I find all of this rather dystopian, but here we are. We are in a world where trust is so low that we have to certify our humanness in writing (as well as images, videos and more).
The world now needs me to prove that I wrote what I wrote, that I am a human author.
Amazon’s KDP has approached it this way. KDP requires authors to disclose AI-generated text, images, or translations. It does not require disclosure of AI-assisted content. Amazon’s guidelines spell out the difference: if an AI tool produced the actual words, images, or translation that ended up in the book, that’s AI-generated, even after heavy editing.
If a human wrote the material and used AI to brainstorm, outline, edit, or check for errors, that’s AI-assisted, and no disclosure is required at all.
The distinction that matters isn’t whether AI touched the process anywhere, it’s what role AI played, and who made the decisions that shaped the final text.
I like the disclosure approach
If you utilise AI in your writing be upfront about it. There’s nothing at all wrong (and a lot right) about using a whole raft of AI tools to assist you in your writing. If you are using AI to write for you…even based on your own previous writing – say so.
If you’re like me and don’t use AI to write, get your writing certified.
And, keep a track of where, when and how you sourced and wrote your content. I do this as a matter of course in my company history research as I am sourcing information from non-digitised sources (historical paper documents, in person interviews, old video and audio recordings and more). I have to keep track of all this so I can show where I sourced information if its accuracy is questioned or tested.
- This article was written by me referencing the sources I have shared.
