Big Idea

The argument in one line.

Fable 5.1's EU-mandated watermark doesn't stamp text with a hidden signature, it tilts the probability of certain word choices so a party holding a private key can statistically guess AI authorship, and the only real fix for someone who doesn't want that is running long-form output through an unwatermarked local model before publishing it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You publish long-form AI-assisted writing, blog posts, books, or scripts, and want to know whether it can be flagged as AI-generated.
You use Claude for content you don't want statistically traceable back to an AI model.
You're curious how LLM watermarking actually works mechanically, beyond the press-release language.

SKIP IF…

You only produce short posts, comments, or internal notes, where output is too short for watermark detection to say much of anything.
You don't care whether your writing is identifiable as AI-generated.

TL;DR

The full version, fast.

Anthropic's Fable 5.1 now watermarks every output, a requirement of the EU AI Act that every major lab except xAI has signed. The watermark isn't hidden code, it's a statistical bias: among several equally valid words, the model is nudged to favor certain ones more often. Only someone holding Anthropic's private key, not the public, can use that bias to estimate the odds a text was AI-written, and that key isn't available on request. The bias never changes facts, code, or meaning, just word choice. Because Claude, ChatGPT, and Gemini are all watermarked, none of them can rewrite the text clean; the workaround is running the output through an unwatermarked local model, such as one hosted with Ollama, for a full rewrite before publishing.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:25

01 · Watermarking

Fable 5.1 is the first Claude model to watermark all of its outputs, a change tied to an EU AI Act provision that every major frontier lab except xAI has signed, requiring a way to identify AI-generated text.

01:25 – 05:29

02 · How It Works

Modeled on Google DeepMind's 2024 SynthID paper, the watermark works by loading the dice on word choice: among several equally valid words, some get weighted higher so a party holding the private key can later estimate the odds a text was AI-written, without ever changing facts, code, or meaning.

05:29 – 07:57

03 · Removal Strategy

Since Claude, ChatGPT, and Gemini are all now watermarked, none of them can cleanly rewrite watermarked text. The fix is routing the output through an unwatermarked open-source local model, such as one run with Ollama, for a full rewrite, and the video flags that short outputs are unlikely to be long enough for the watermark to matter at all.

07:57 – 08:48

04 · Removal Prompt

The presenter shares a prewritten prompt that has Claude research the viewer's hardware, install Ollama, pick a suitable local model, and build a reusable skill that rewrites any output through that local model automatically.

08:48 – 10:13

05 · Outro

Closing take: the presenter argues watermarking concerns are somewhat overblown, since most AI-assisted writing is already obviously AI-written regardless of any statistical watermark, then points to his Claude Code masterclass.

Atomic Insights

Lines worth screenshotting.

Anthropic's Fable 5.1 watermarks every output it generates, in compliance with an EU AI Act that every major frontier lab except xAI has signed.
The watermark isn't a hidden signature stamped into text, it's a statistical bias that makes certain words more likely to be chosen.
Anthropic's approach is modeled on a 2024 Google DeepMind paper on SynthID's tournament-based watermarking for language models.
In an unwatermarked baseline, four equally likely words might each get a 25% chance of being picked; watermarking loads the dice so one jumps to 50% and another to 30%.
Only someone holding the private watermarking key, not the end user and not the public, can statistically infer that a piece of text came from a watermarked model.
There's no public site that can actually detect Anthropic's watermark. Anyone claiming otherwise is lying, because the key isn't publicly available.
The watermark is engineered to never change facts, code output, or overall meaning, only which of several equally valid words gets used.
Because Claude, ChatGPT, and Gemini are all watermarked, none of them can be used to rewrite watermarked text and strip the signal.
The only real way to remove the watermark is a full rewrite through an open-source local model, since local models aren't subject to the EU rule.
Short text, like a one-line comment, probably isn't long enough for watermark detection to say much with confidence; the concern mainly applies to long-form output.
The practical workaround is a two-step pipeline: draft with Claude as usual, then send the output to a local model for a full rewrite before publishing.

Takeaway

The watermark is a bias in word choice, not a hidden mark

HOW IT WORKS

Fable 5.1's EU-mandated watermark doesn't stamp text, it tilts word probabilities so a private key holder can guess AI authorship, and the only real fix is rewriting long output through an unwatermarked local model.

Fable 5.1 is the first Claude model that watermarks every output it produces, and the word "watermark" undersells what's actually happening since there's no hidden mark or invisible code embedded in the text.
The watermark exists because Anthropic signed onto an EU AI Act requirement that every major frontier lab except xAI has also agreed to, and it applies to any model released after August 2nd.
Anthropic's implementation is modeled on a 2024 Google DeepMind paper on SynthID's tournament-based watermarking, which is worth treating as a simplified mental model rather than reading the full technical paper cold.
Without watermarking, if several words are equally valid to complete a sentence, the model effectively rolls dice and picks one at random with equal odds; watermarking loads that same die so certain words get chosen more often across many outputs.
The end user never sees which words were weighted or by how much. That information is invisible from the output alone, and someone holding Anthropic's private key can only estimate a probability, like 90% or 99%, that a stretch of text came from a watermarked model, never certainty.
The bias is engineered to leave meaning, performance, code output, and factual content untouched. It only nudges which of several equally valid words gets used.
The watermarking key isn't public and there's no request process open to ordinary users; any website or tool claiming it can tell you whether text is watermarked is lying, because the detection key required to do that doesn't exist publicly.
Because Claude, ChatGPT, and Gemini are all now subject to their own watermarking, none of them can be used to cleanly rewrite watermarked text without just re-watermarking it.
The workaround is to draft with Claude as usual, then send that output to an open-source model running locally, for example through Ollama, for a full rewrite. Light editing isn't enough, the rewrite has to be substantial enough to break the original word-choice bias while keeping the same voice and meaning.
Whether the watermark can be detected with any real confidence depends on output length. Short text, like a one-line comment, likely isn't long enough to test, while long-form output like a full blog post is long enough for the bias to become statistically detectable.
The most practical setup is downloading Ollama, picking a local model that fits your hardware, and building it into a Claude skill that handles the rewrite automatically instead of doing it by hand each time.
The whole watermarking concern may be somewhat overblown in practice, since most AI-assisted writing is already obviously AI-written regardless of any statistical watermark, simply because the writer never bothered to humanize it.

Glossary

Terms worth knowing.

SynthID: Google DeepMind's tournament-based text watermarking method, described in a 2024 paper, that Anthropic's own watermarking system is modeled on.
EU AI Act: European Union regulation requiring that AI-generated content be technically identifiable. Every major frontier AI lab except xAI has signed on to comply with it.
Text watermarking: A technique where an AI model quietly favors certain word choices over other equally valid ones, so a party holding a secret key can later estimate whether a piece of text was AI-generated.
Softmax: The function a language model uses to convert its raw scores for possible next words into a probability distribution it samples from.
Local model: An open-source AI model, run on your own hardware through software like Ollama, that isn't subject to a lab's watermarking requirements because it isn't operated by that lab.

Resources

Things they pointed at.

01:05linkScalable Watermarking for Identifying Large Language Model Outputs (Google DeepMind, 2024)

05:21linkAnthropic's "How Claude's text watermark works" article ↗

06:35toolOllama ↗

10:05productChase AI Plus / Claude Code Masterclass ↗

Quotables

Lines you could clip.

00:00

“Anthropic launched Fable 5.1 and it is the first model they have put out that now watermarks everything it creates.”

clean cold-open news hook, no setup needed→ TikTok hook↗ Tweet quote

05:05

“Somebody who has the key, the watermarking secret key, can tell.”

crisp reveal of the core mechanism→ IG reel cold open↗ Tweet quote

09:05

“Anyone who's telling you that they can tell it's watermarked is pretty much lying to you.”

punchy myth-busting line that stands alone→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

So yesterday Anthropic launched Fable 5 .1 and it is the first model they have put out that now watermarks everything it creates. And today I'm going to show you how you can get rid of it. But first let's talk about what this watermark even is because it's a little bit of a misnomer to call it a watermark at all.

Why it's here in the first place and how it works because if you understand those things then the solution to get around this watermark will make a lot more sense. Now, why do we have to deal with this at all?

Well, it's because this is in compliance with an EU AI Act that essentially all the major frontier labs have signed with the exception of XAI. And what this code states is that they need to, and they being the EU, needs a way to be able to tell if some sort of text or some sort of output was created by AI. And in order to do this, Anthropic has added a watermark to all of its outputs for models released after August 2nd, just is fable 5 .1 now anthropic claims this has no practical effects on the actual output it gives you and that it's very limited in terms of what it's going to change if we're talking about code but how does this thing actually work and why did i call it a misnomer in the first place Now, Anthropic's strategy for injecting these watermarks into its outputs is based on a paper that came out from Google in 2024 called Scalable Watermarking for Identifying Large Language Model Outputs.

Now, this is a relatively technical paper. It goes into SynthID's tournament -based watermaking, but let me give you a simple mental model for you to sort of take on board. Now, as you know, large language models and their outputs are based on probability and there's a certain amount of randomness involved.

If I give it a sentence that says she is blank and it needs to fill it out, well, it's going to have a list of words it could possibly use to finish that sentence. And so let's say those words are beautiful, stunning, gorgeous, and pretty. Now, in reality, each of these words will have a different percentage or a different probability associated with them in terms of what the large language model will choose.

A term you will often hear is called softmax, but you just need to understand all these words have a probability associated with them in regards to which one the large language model will choose. For this mental exercise, I want you to imagine all of these have the exact same percentage, right? So each of these words to finish the sentence has a 25 % chance of being chosen.

Now, normally, before watermarking came into effect, it would just choose one of these. And so let's say it... rolls the four -sided dice and it chooses beautiful.

Cool, right? It had a 25 % chance. All these had equal chance of being chosen.

But now we work in a different system and that system is the watermarking system. With the watermarking system, we can still be in the same place where I say, hey, I need you to fill out the sentence. She is blank and you have four options, right?

And let's say, again, it's these four words and they have the same probability. When we add the watermarking construct, what it is going to do is it essentially is going to load this die.

If this is some sort of four -sided die, well, it is then gonna say, hmm, for watermarking purposes, beautiful is now going to be 50 % and gorgeous is going to be 30%. So when we add a watermarking sort of system in place, it's going to choose certain words and more heavily weight them. And because certain words are more heavily weighted when the enlarged language model chooses them, well, those words are gonna show up more often.

Using a watermarking system, more often than not, it's going to say, she is beautiful. Now, you as the end user have no idea what these words are and which words are being weighted more heavily. But somebody who has the key, the watermarking secret key, can tell.

Now, they can't tell with absolute certainty because just like before, there was always a chance it was going to say, she is beautiful. You don't know that it was necessarily. had a higher chance because of watermarking.

But a person with the key is going to know which words were weighted more heavily, and it then is going to give them some sort of probability that tells that user who has the key, i .e. European regulators, okay, there's a 90 % chance that this text, a 100 % chance, a 99 % chance that this text was written with Claude, or this text was written with AI.

That's how it works. Now, the idea here is that... These words that it chooses to change and weight more heavily with watermarking aren't going to change the overall meaning of the sentence.

It's not going to change your output. It's not going to change your performance. And when we talk about code specifically, it's not going to change code if it's going to affect what the output would look like, and it's not going to change facts.

You know, if it says, hey, this historical event, you know, occurred on, you know, 8 -10, well, it's not going to change that wording, right? It's a fact. It occurred on 8 -10.

That's not going to be changed. But that is a simplified, again, simplified explanation of how this system works. Now, Anthropic hasn't totally come out with the exact algorithm for how that works, but we can extrapolate that from sort of the DeepMind papers.

And what you also should have pulled from that is that having this key is the only way to know if it was written by ai based on the watermarking system and this key is not public to get access to the key you have to request access there's no public list you probably have to be some sort of eu regulator so anyone who comes out there and says oh hey here's my website that can tell if it's watermarked or not that's a lie that this does not exist now if you want to go deeper into how the text watermarks work anthropic also has their own article and they really try to hammer the point home that this isn't going to affect the outputs so the watermarking system in essence is all about certain words being chosen more than others.

And this is a statistical algorithm that you, the end user, will never be able to figure out unless you have the key. So how do we fix this? Do we simply do light editing to sort of the text outputs that Claude gives us and hope that's going to work?

Or do we need to essentially rewrite the whole thing? And how do we rewrite it? Because you can't use Claude to rewrite whatever this sentence is because that will be watermarked.

And you also can't use ChatGPT. You also can't use Gemini. So your options become pretty clear.

You need to rewrite significant portions of the text while retaining the overall feel. And you need to use either XAI, so you either need to use Grok, or more likely what you're going to do is use some sort of open source model, right? Some sort of Chinese model that you can download and run on your computer that A, is going to be free and b is not subject to these watermarking laws now what does that system look like in a practical sense well you do all your normal outputs with claude and if you've created something think some sort of long blog post and you don't want it to be flagged with the watermark we're going to take that output you're then going to send it to some sort of local model and Claude can do that for you.

There's tons of really good local models out there that you can download via Ollama. That local model is going to go ahead and do its thing. It's going to rewrite the text.

You're going to give it a prompt that says, hey, I want you to, you know, retain the actual feel of this. You can insert your own voice notes, whatever. And then once it's complete, it's just going to send its exact response verbatim back to Claude.

And you can do whatever you want with it at that point. Now, the one other variable you need to think about in this whole thing is the length of your outputs. Like how many tokens are we talking about?

and this is not a number that we can actually know, is how many tokens, how long does this output need to be for the detection API if it is run to say with any degree of confidence it was written with AI? Because if I'm just writing LinkedIn comments that just say she is beautiful, can the API really tell if that was written with Claude or not?

No. So if you're someone who is very sort of scared about having things watermarked and attributed to Claude, well, just ask yourself, am I writing things that are long enough for it to even matter? If we're talking like LinkedIn posts, comments under, you know, 100 words or so, probably not.

But if we're talking like giant blog posts that are thousands of words long, yeah, you are going to be hit with that watermark. And it's going to be very obvious that it was written with Claude. Now, the easiest way to get this all set up is to download Ollama, download a local model that makes sense for your particular hardware, then create a skill inside of Claude that when you call it, sends your output to that local model, it does its full rewrite, and then it brings it back.

To make this easy for you, I've created an entire prompt that walks you and Claude through the entire process. You simply copy paste this into Claude. It's going to do some research on your hardware.

It's going to figure out what local model makes sense for you. It's going to download and install a llama. If you haven't done it, it's then going to create a skill for you and it's going to ask you questions so you can get the actual verbiage of how that rewrite sounds like basically.

injecting your voice, giving examples, that sort of thing. So I'll put a link down below to actually do this. And rather than walk you through this entire thing step by step for the next 10 minutes, just run this prompt and it will do it all for you.

So hopefully that whole explanation for watermarking made sense. This isn't an actual watermark. There's nothing being written behind the scenes.

There's no code there. It's just that specific words are going to be used more often. You're going to have no idea what those specific words are.

The only way you can tell is if you have the key for the watermark, which you won't have. unless you're some sort of regulator. So anyone who's telling you that they can tell it's watermarked is pretty much lying to you.

So know that right away. And based on how this works, again, you can dive into that DeepMind paper if you want. The only way you're going to get around this is with heavy editing.

It's not going to be light edits. You have to rewrite it all. And the only way you're going to be able to do that is with an open source model.

And it's going to be the cheapest way. So if you want to get your hands on that prompt, you can find it down below. Hopefully this helps some of you out.

To be totally honest, I think the whole thing is rather overblown in terms of, you know, things are getting watermarked of that nature. Most of the people you see using AI to do sort of their writing, it is painfully obvious already that it's AI because they have really no idea how to humanize it or make it sound like them.

And unless we're getting into a place where, you know, everywhere you post, every single website has some sort of like watermark or detector and that becomes some like huge faux pas. Again, all we have to do is run this through a local model. you're kind of scot -free so let me know what you thought uh make sure to check out chase ai plus to get your hands on the cloud code master class if this is something you want to dive into more other than that have fun avoiding these watermarks

The Hook

The bait, then the rug-pull.

Anthropic's newest model quietly watermarks every piece of text it writes, and the presenter opens by admitting even calling it a "watermark" is a bit of a misnomer, then spends the next ten minutes explaining what it actually is and how to get around it.

Frameworks

Named ideas worth stealing.

02:50concept

The Loaded Dice Model of Watermarking

A mental model for token watermarking: instead of giving every valid next word an equal chance, the system quietly reweights the probabilities (e.g. 50%/30%/10%/10% instead of an even 25% each) so a party holding the secret key can later infer, with some statistical confidence, that a text came from the watermarked model.

Steal forExplaining probabilistic AI-text detection to a non-technical audience

06:25list

The Rewrite Bridge

Write your first draft with Claude as usual
Send the raw output to an unwatermarked local model (e.g. via Ollama)
Have the local model do a full rewrite that keeps your voice and meaning
Publish the local model's rewritten output, never the original watermarked text

The practical fix for anyone who doesn't want long-form AI writing statistically flagged: never publish a watermarked model's raw output directly, always route it through an open-source local model first.

Steal forAny content pipeline producing long-form AI-assisted writing meant for public distribution

CTA Breakdown

How they asked for the click.

VERBAL ASK

10:05product

“make sure to check out chase ai plus to get your hands on the cloud code master class”

single low-key verbal mention at the very end, no on-screen link or hard pitch; the removal prompt and resources are pointed to only via "link down below" in the description

MENTIONED ON CAMERA

05:21linkAnthropic's "How Claude's text watermark works" article ↗

06:35toolOllama ↗

10:05productChase AI Plus / Claude Code Masterclass ↗

Storyboard