European policy, specifically the “Code of Practice on Transparency of AI-generated Content”, is rightfully influencing AI companies to watermark their output to protect the public, and those watermarks are being applied worldwide.

Anthropic has signed the act, and a lot of the internet is beside itself because they’ll now have to write their own prose and marketing copy, or completely rewrite what Claude has been spewing forth when they prompt it. You see, Claude is now watermarking the text it generates in a way that can survive a simple copy and paste.

How will these people ever pass these documents as their own? How will marketing companies produce hundreds of articles from clients in a single morning? Will Fiverr become viable for article writers again? THE END IS NIGH!!!

For the moment, everyone appears to have forgotten that Google has been watermarking text since 2024 with SynthID. Anthropic has even claimed SynthID as a reference, but has not released an official spec.

How text watermarking works actually involves biasing the AIs output (token choice) towards what is dictated by a “secret key” or formula. Here’s an in-depth breakdown of how how AI text watermarking happens and now for my overly simplified version.

We start with a secret math formula or key, and the knowledge that many words have almost the same meaning. Then, when the AI is choosing the next word (token), we nudge it towards choosing words that are closer to the key instead of words that might otherwise be chosen, unless there is no other correct word to choose. When we have enough biased words appearing in the text, we can make a statistical confidence measure of whether or not AI generated the text.

For, ahem, “writers,” this actually results in lower quality output. It may be only negligibly worse, but the difference is certainly not zero degradation, as the output has a bias towards the formula so that the watermark can actually work.

What’s cracking me up is that the noisiest pushback is coming not from the legions of “authors” who suddenly discovered their hidden talents when they started paying for a Claude or ChatGPT subscription, but from lawyers and editors who are worried about “false attribution”.

They are describing a real and specific problem: false attribution. The watermark records that the model produced a sequence of tokens. It does not, and structurally cannot, record who produced the ideas, the argument, the research or the original draft. For a writer whose employer, publisher, journal or university reads a positive signal as settled proof of machine authorship, that gap is not academic.

Except, the proclaimed innocence and fear of false attribution does not pass the sniff test. Actually, it reeks of bullshit.

Take my workflow as an example. I draft my documents in WordPress, and when I’m close to my final revision, I paste them into Grammarly for spell-checking and grammar suggestions. Then I copy the revised text back into WordPress to add formatting, images, etc.

Occasionally, Grammarly will propose an edit to my sentence that includes removing or re-organizing words, yet it never really replaces my words, and even if it did, there wouldn’t be enough of them to generate enough statistical confidence to be flagged as AI-generated.

To generate that level of confidence, significant chunks, longer than rejigging a short passage, are required to have been generated by the AI.

For those who claim they wrote an article, legal document, or prose piece, and only had the AI fluff it (not a porn joke), or do some light editing, then there should not be enough consecutive text edited to get a statistical distribution of words created by the AI to trigger detection.

Lawyers especially might want to text a step back before their angry typing fingers expose their shenanigans. The more specific the style of writing, or the more fixed the syntax, the less effective the watermark is. The AI simply doesn’t have other words to choose from. And when it comes to the law, legal terms are well-defined and relatively specific in their usage.

This chart from the Explainx piece I linked above gives an idea of how deeply the watermarking “signal” is embedded in different types of text. I would assume that legal documents sit fairly close to technical writing on that scale.

Text Type	Entropy	Signal Carried
Discursive prose, essays, emails	High — many valid phrasings	Strong
Technical writing	Medium — terminology is constrained	Moderate
Source code	Low — identifiers and syntax are fixed	Weak
Structured output (JSON, tables)	Very low — schema-constrained	Near zero
A factual list or quoted text	Very low	Near zero

Of course, all of this worry about AI text watermarking is pointless. You can just throw the text into another AI that doesn’t watermark it and have it rewrite it. You can heavily edit the text, or go the whole hog and rewrite the text by hand from scratch.

All of which is probably more of a chore, especially when you add in the time you spend on source, fact and hallucination checking, than just doing the damn work yourself in the first place.