Fable 5.1 Watermarks EVERYTHING, Here's How To Remove It

CChase AI
Computing/SoftwareInternet Technology

Transcript

00:00:00So yesterday, Anthropic launched Fable 5.1, and it is the first model they have put out
00:00:04that now watermarks everything it creates. And today, I'm going to show you how you can get rid
00:00:10of it. But first, let's talk about what this watermark even is, because it's a little bit
00:00:14of a misnomer to call it a watermark at all, why it's here in the first place, and how it works.
00:00:19Because if you understand those things, then the solution to get around this watermark will make
00:00:23a lot more sense. Now, why do we have to deal with this at all? Well, it's because this is
00:00:26in compliance with an EU AI Act that essentially all the major frontier labs have signed with the
00:00:31exception of XAI. And what this code states is that they need to, and they being the EU,
00:00:38needs a way to be able to tell if some sort of text or some sort of output was created by AI.
00:00:43And in order to do this, Anthropic has added a watermark to all of its outputs for models released
00:00:50after August 2nd, 2026, which just is Fable 5.1. Now, Anthropic claims this has no
00:00:56practical effects on the actual output it gives you, and that it's very limited in terms of what
00:00:59it's going to change if we're talking about code. But how does this thing actually work? And why did
00:01:03I call it a misnomer in the first place? Now, Anthropic's strategy for injecting these watermarks
00:01:07into its outputs is based on a paper that came out from Google in 2024 called Scalable Watermarking
00:01:13for Identifying Large Language Model Outputs. Now, this is a relatively technical paper. It goes into
00:01:19SynthID's tournament-based watermaking. But let me give you a simple mental model for you to sort of take
00:01:25on board. Now, as you know, large language models and their outputs are based on probability and there's a
00:01:29certain amount of randomness involved. If I give it a sentence that says, she is blank and it needs to
00:01:34fill it out, well, it's going to have a list of words it could possibly use to finish that sentence.
00:01:41And so let's say those words are beautiful, stunning, gorgeous, and pretty. Now, in reality,
00:01:48each of these words will have a different percentage or a different probability associated with them in terms of
00:01:54what the large language model will choose. A term you will often hear is called softmax, but you just need
00:01:58to understand all these words have a probability associated with them in regards to which one the
00:02:04large language model will choose. For this mental exercise, I want you to imagine all of these have
00:02:08the exact same percentage, right? So each of these words to finish the sentence has a 25% chance
00:02:14of being chosen. Now, normally before watermarking came into effect, it would just choose one of these.
00:02:20And so let's say it rolls the four-sided dice and it chooses beautiful. Cool, right? It had a 25%
00:02:26chance, all these had equal chance of being chosen. But now we work in a different system and that system
00:02:31is the watermarking system. With the watermarking system, we can still be in the same place where I say,
00:02:38hey, I need you to fill out the sentence, she is blank, and you have four options, right? And let's
00:02:42say, again, it's these four words and they have the same probability. When we add the watermarking
00:02:49construct, what it is going to do is it essentially is going to load this die. If this is some sort of
00:02:54four-sided die, well, it is then going to say, hmm, for watermarking purposes, beautiful is now going
00:03:01to be 50% and gorgeous is going to be 30%. So when we add a watermarking sort of system in place,
00:03:10it's going to choose certain words and more heavily weight them. And because certain words are more
00:03:15heavily weighted when the large language model chooses them, well, those words are going to show
00:03:19it more often, right? Using a watermarking system, more often than not, it's going to say she is
00:03:25beautiful. Now, you as the end user have no idea what these words are and which words are being
00:03:32weighted more heavily. But somebody who has the key, the watermarking secret key can tell. Now,
00:03:40they can't tell with absolute certainty because just like before, you know, there was always a chance
00:03:45it was going to say, you know, she is beautiful. You don't know that it was necessarily had a higher
00:03:49chance because of watermarking. But a person with the key is going to know which words were weighted
00:03:55more heavily. And it then is going to give them some sort of probability that tells that user who has the
00:04:01key, i.e. European regulators, okay, there's a 90% chance that this text, 100% chance, a 99% chance that this
00:04:09text was written with Claude, or this text was written with AI. That's how it works. Now,
00:04:15the idea here is that these words that it chooses to change and weight more heavily with watermarking
00:04:20aren't going to change the overall meaning of the sentence. It's not going to change your output.
00:04:24It's not going to change your performance. And when we talk about code specifically, it's not going to
00:04:28change code if it's going to affect what the output would look like. And it's not going to change facts.
00:04:34You know, if it says, hey, this historical event, you know, occurred on, you know, 810. Well,
00:04:41it's not going to change that wording, right? It's a fact. It occurred on 810. That's not going to be
00:04:45changed. But that is a simplified, again, simplified explanation of how this system works.
00:04:52Now, Anthropic hasn't totally come out with the exact algorithm for how that works, but we can
00:04:57extrapolate that from sort of the DeepMind papers. And what you also should have pulled from that is that
00:05:02having this key is the only way to know if it was written by AI based on the watermarking system.
00:05:08And this key is not public. To get access to the key, you have to request access. There's no public
00:05:12list. You probably have to be some sort of EU regulator. So anyone who comes out there and says,
00:05:16oh, hey, here's my website that can tell if it's watermarked or not, that's a lie that this does not
00:05:21exist. Now, if you want to go deeper into how the text watermarks work, Anthropic also has their own
00:05:25article. And they really try to hammer the point home that this isn't going to affect the outputs.
00:05:29So the watermarking system, in essence, is all about certain words being chosen more than others.
00:05:35And this is a statistical algorithm that you, the end user, will never be able to figure out unless
00:05:40you have the key. So how do we fix this? Do we simply do light editing to sort of the text outputs
00:05:47that Claude gives us and hope that's going to work? Or do we need to essentially rewrite the whole thing?
00:05:52And how do we rewrite it? Because you can't use Claude to rewrite whatever this sentence is,
00:05:56because that will be watermarked. And you also can't use ChatGPT. You also can't use Gemini.
00:06:01So your options become pretty clear. You need to rewrite significant portions of the text while
00:06:09retaining the overall feel. And you need to use either XAI. So you either need to use Grok or more
00:06:16likely what you're going to do is use some sort of open source model, right? Some sort of Chinese
00:06:21model that you can download and run on your computer that A is going to be free and B is not subject to
00:06:28these watermarking laws. Now, what does that system look like in a practical sense? Well, you do all your
00:06:33normal outputs with Claude. And if you've created something, think some sort of long blog post, and you
00:06:37don't want it to be flagged with the watermark, we're going to take that output. You're then going to send
00:06:42it to some sort of local model. And Claude can do that for you. There's tons of really good local models
00:06:47out there that you can download via OLAMA. That local model is going to go ahead and do its thing.
00:06:51It's going to rewrite the text. You're going to give it a prompt that says, "Hey, I want you to,
00:06:55you know, retain the actual feel of this. You can insert your own voice notes, whatever." And then once
00:07:01it's complete, it's just going to send its exact response verbatim back to Claude. And you can do
00:07:05whatever you want with it at that point. Now, the one other variable you need to think about in this whole
00:07:09thing is the length of your outputs. Like how many tokens are we talking about? Because, and this is not a
00:07:14number that we can actually know, is how many tokens, how long does this output need to be
00:07:19for the detection API, if it is run, to say with any degree of confidence it was written with AI.
00:07:25Because if I'm just writing LinkedIn comments that just say, "She is beautiful," can the API really tell
00:07:31if that was written with Claude or not? No. So if you're someone who is very sort of scared about having
00:07:37things watermarked and attributed to Claude, well, just ask yourself, "Am I writing things that are long
00:07:41enough for it to even matter?" If we're talking like LinkedIn posts, comments under, you know,
00:07:46100 words or so, probably not. But if we're talking like giant blog posts that are thousands of words
00:07:51long, yeah, you are going to be hit with that watermark. And it's going to be very obvious that
00:07:56it was written with Claude. Now, the easiest way to get this all set up is to download Ollama,
00:08:00download a local model that makes sense for your particular hardware, then create a skill inside of
00:08:05Claude that when you call it, sends your output to that local model, it does its full rewrite,
00:08:10and then it brings it back. To make this easy for you, I've created an entire prompt that walks you
00:08:15and Claude through the entire process. You simply copy paste this into Claude. It's going to do some
00:08:20research on your hardware. It's going to figure out what local model makes sense for you. It's going
00:08:24to download and install Ollama if you haven't done it. It's then going to create a skill for you,
00:08:28and it's going to ask you questions so you can get the actual verbiage of how that rewrite sounds like.
00:08:35Basically injecting your voice, giving examples, that sort of thing. So I'll put a link down below
00:08:40to actually do this. And rather than walk you through this entire thing step-by-step for the
00:08:44next 10 minutes, just run this prompt and it will do it all for you. So hopefully that whole explanation
00:08:48for watermarking made sense. This isn't an actual watermark. There's nothing being written behind
00:08:53the scenes. There's no code there. It's just that specific words are going to be used more often.
00:08:57You're going to have no idea what those specific words are. The only way you can tell is if you have
00:09:02the key for the watermark, which you won't have unless you're some sort of regulator. So anyone who's
00:09:07telling you that they can tell it's watermarked is pretty much lying to you. So know that right away.
00:09:11And based on how this works, again, you can dive into that DeepMind paper if you want. The only way
00:09:16you're going to get around this is with heavy editing. It's not going to be light edits. You have to
00:09:20rewrite it all. And the only way you're going to be able to do that is with an open source model.
00:09:24And it's going to be the cheapest way. So if you want to get your hands on that prompt,
00:09:29you can find it down below. Hopefully this helps some of you out. To be totally honest,
00:09:33I think the whole thing is rather overblown in terms of, you know, things are getting watermarked
00:09:38of that nature. Most of the people you see using AI to do sort of their writing, it is painfully obvious
00:09:43already that it's AI because they have really no idea how to humanize it or make it sound like them.
00:09:49And unless we're getting into a place where, you know, everywhere you post,
00:09:52every single website has some sort of like watermarker detector and that becomes some like huge faux pas.
00:09:58Again, all we have to do is run this through a local model and you're kind of scot-free.
00:10:02So let me know what you thought. Make sure to check out Chase AI+ to get your hands on
00:10:06the Cloud Code Masterclass if this is something you want to dive into more.
00:10:10Other than that, have fun avoiding these watermarks.

Key Takeaway

Anthropic's Fable 5.1 model implements statistical token watermarking to comply with EU regulations, which can be avoided by running generated text through an open source local model via Ollama.

Highlights

  • Anthropic launched Fable 5.1 on August 2nd, 2026, marking the first model that applies mandatory output watermarking.

  • EU AI Act compliance requires major frontier labs except XAI to implement text watermarking to identify artificial intelligence outputs.

  • Statistical watermarking works by loading the probability distribution of token choices to favor specific pre-selected words.

  • Deciphering text watermarks requires a secret key held exclusively by European regulators, rendering public verification tools ineffective.

  • Running outputs through an open source local model via Ollama bypasses Anthropic watermarking restrictions.

Timeline

Fable 5.1 Release and EU Regulatory Compliance

  • Anthropic released Fable 5.1 on August 2nd, 2026, featuring mandatory text watermarking.
  • Major frontier labs adopted watermarking to comply with the EU AI Act.
  • XAI remains the sole major frontier lab that has not signed the EU AI code.

Anthropic introduced a watermarking mechanism for all models released after August 2nd, 2026, starting with Fable 5.1. This change stems from compliance requirements under the EU AI Act, which mandates a method to identify artificial intelligence outputs. While Anthropic claims the watermark exerts no practical effect on standard outputs or code performance, the addition alters text generation patterns.

Mechanics of Statistical Token Watermarking

  • Statistical watermarking derives from Google's 2024 paper on SynthID tournament-based watermarking.
  • The algorithm alters the probability distribution of token choices to heavily weight specific words.
  • Secret watermarking keys remain restricted to authorized entities such as European regulators.

Large language models select output tokens based on probability percentages known as softmax. Watermarking modifies these probabilities by assigning higher weights to predetermined words during generation. End users cannot identify these weighted words without the proprietary secret key. Consequently, public websites claiming to detect watermarked text operate without valid access and provide unreliable results.

Bypassing Watermarks with Local Open Source Models

  • Short text outputs like LinkedIn comments lack sufficient length for detection APIs to confirm artificial generation.
  • Local open source models run via Ollama circumvent watermarking laws completely.
  • Automated scripts integrate local rewrites directly into Claude workflows to maintain the original voice.

Short texts under one hundred words generally escape reliable detection by watermarking APIs. For longer content like blog posts, heavy rewriting is necessary because light edits fail to remove the statistical footprint. Utilizing open source models via Ollama allows creators to rewrite text locally without triggering regulatory watermarks. Automated prompts configure this workflow directly within Claude to streamline the transition.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video