Another questionable move from Company A has sparked widespread outrage online.
Just now, Anthropic announced: The new Claude will embed invisible watermarks in generated text.
Yes, not just metadata—these "ink marks" will become part of the text itself.
They will be copied and pasted along with the text to other places, and may still remain even after some degree of editing.
It's like a stubborn sticker that just won't come off...
Netizens have already started voicing their anger collectively.
The direct driver of this move is the EU's AI Act, which Anthropic has signed. This mechanism will apply to all models released on or after August 2, 2026.
For Claude models already online, Anthropic is also researching how to add this watermarking mechanism retroactively.
Oh, and one more thing—
This measure will be implemented globally, not limited to the EU.
The great "Watermark Era" is here.
Text Gets Watermarks Too
Recently, Anthropic signed the EU AI Act's Transparency Code of Conduct for AI-Generated Content.
This code was drafted by independent experts, evaluated by the European Commission and the AI Office, and has been signed by about 190 institutions, including Google, Meta, Microsoft, OpenAI, and others.
For Claude, this translates into two specific things.
First, text watermarking.
Supported Claude models will weave a kind of invisible marker directly into the text when generating content.
Yes, completely invisible. You can't see it, and it doesn't change the meaning, quality, or readability of the response.
What's even scarier is that it's hard to get rid of.
The watermark is part of the text, copying and pasting will carry it along, and even edited versions might not be safe.
Second, signed metadata.
When Claude generates supported file types like .svg, .png, .jpg, it will attach digitally signed provenance metadata.
This metadata follows the C2PA open standard, promoted by Adobe, OpenAI, Google, and others. It can indicate the file was processed by Claude and also detect if the file has been tampered with.
The coverage is across the entire product line. API, Claude, Claude Code, Claude Cowork, Claude Tag are all included.
Cloud customers accessing Claude via AWS, Google Cloud, Microsoft Foundry will also be subject to this.
And there are no regional exceptions; it's uniformly implemented worldwide.
Meanwhile, Anthropic stated it is developing detection tools, allowing users and third parties to check whether a piece of text or a file carries the Claude marker.
Detailed technical documentation will be released by the company later.
However, Anthropic also explicitly noted the limitations of this mechanism.
Detecting a mark only indicates the content was processed by Claude, it does not confirm Claude as the original author.
After all, users often use Claude for proofreading, translation, summarization, format conversion—the core ideas and data may come from humans; marked content could also be modified, excerpted, or mixed with other materials afterward.
Conversely, not detecting a mark doesn't mean the content definitely wasn't AI-generated.
It might come from an older model, be heavily edited or rewritten, the paragraph might be too short to identify, or metadata might be stripped due to format conversion or screenshots...
Basically, everyone is under suspicion now...
As of now, Anthropic has not disclosed details of the watermarking algorithm.
In other words, no one knows what the "invisible ink" looks like, or what else Company A might be able to do with it.
Company A's Labeling Tactics
So, is there really such magic that makes watermarks in text completely invisible?
Yes.
And the tricks might be more numerous than you think.
The most famous is something called Unicode.
You can think of it as issuing ID cards to almost every character and symbol in the world: 'A' corresponds to one Unicode number, '中' (Chinese) corresponds to one Unicode number, even '🙂' corresponds to one Unicode number.
But beyond that, there exists a set of characters that are basically invisible to the human eye but distinguishable by computers.
For example, the simplest "Hello"—underneath, one might be H-e-l-l-o, and another might be H-e-invisible character-l-l-o.
They look completely identical to the human eye, but when a program reads the Unicode code points, it can find that the second string has an extra special character.
Amazing.
I wonder if everyone remembers, Company A had already "gracefully followed the trend" and used the Unicode trick before.
On June 30th, a Reddit user discovered a shocking truth after reverse-engineering Claude Code—
Claude Code's packaged files contained a set of Trojan-like code. If it detected you were a Chinese user, it would add two strokes of invisible ink to your system prompt—'Today's date is 2026-06-30'.
1. Changing the date format hyphens to slashes, marking a hit for the Chinese timezone.
2. The single quote in 'Today's' was quietly replaced with three characters that have different Unicode code points but are visually identical.
The result? In any editor, any terminal, this line looks exactly the same as the normal version. But in the program, you are already tagged.
Yes, Company A has been using Unicode to secretly tag users in Claude Code.
The company has not revealed whether the new Claude watermark uses this approach.
Looking at it this way, text watermarks might not be so bad, right? At least you won't get your account banned...
Moreover, various AI detection tools are of mixed quality, often charging dozens per detection and giving an AI-generated probability number with no logical basis. Having an official labeling channel could at least save a lot of wasted money.
The main reason for everyone's anger is Company A's history of similar actions.
But thinking carefully, AIGC content does need a reliable detection system, just like OpenAI adding C2PA metadata and SynthID watermarks to images.
However, does embedding watermarks by restricting the model's sampling distribution really not affect generation quality?
Company A's statement is—
No, it doesn't.
I remain skeptical.
After all, since version 4.6, Claude has been increasingly incomprehensible.
One More Thing
Regardless, the "Avengers Alliance" against Anthropic has already begun to act!
YC founder Paul Graham has already hatched a new startup idea, just waiting for aspiring individuals to join!
Create a third-party regeneration function, rephrase the text, preserving the original meaning while removing the watermark.
Of course, a more lightweight solution, usable by everyone right now, has also emerged—
Kimi, remove that invisible watermark.
Reference Links:
[1]https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
[2]https://x.com/atharvabuilds/status/2086920300441268579?s=20
This article is from the WeChat public account "QbitAI", author: Jay
