Claude's watermark policy caused a huge uproar a few days ago.
We previously reported that a large group of companies, including OpenAI, Anthropic, Google, Meta, and Microsoft, signed the EU's "AI Transparency Code of Conduct" and pledged to advance the labeling and detection of AI-generated content.
But Anthropic has clearly gone too far.
They add hidden watermarks to all text content generated by their AI, and this applies to users globally.
Technically, Anthropic adopted the SynthID-Text scheme proposed by Google's DeepMind team in 2024. The principle is to embed statistical patterns when the model makes "insignificant choices." For example, when describing the weather, choosing "overcast" or "grey" makes no difference to the reader, but the accumulation of such choices forms a hidden signature that can be detected by key holders.
Anthropic claims the watermark does not affect output quality, cannot be removed by light editing, but can be eliminated by complete rewriting. However, in such cases, whether the text can still be called AI-generated is itself debatable.
Simply put, even if you hand over a completely self-written article to Claude for punctuation checks, the returned content will be labeled as generated by Claude.
This is very frustrating.
Soon, a countermeasure emerged against Claude's unreasonable watermarking strategy. An open-source project for removing AI watermarks surged to 11k stars on GitHub within five days of its release.
Open source link: https://github.com/guillaumemeyer/watermarks-remover
This open-source project can perform three layers of work:
Layer A (Deterministic Cleaning): Uses a Python script to remove invisible Unicode characters, exotic spaces, bidi control characters, and tag characters. These are the simplest and most brute-force marking methods, and the script can remove them 100%.
Layer B (Statistical Watermark Disruption): Rewrites the text via an Agent to disrupt statistical patterns at the token sampling level. Covers Claude, Google SynthID-Text, OpenAI provenance markings, and Kirchenbauer-class watermarks commonly used in open-source models.
File Layer (Metadata Removal): Removes C2PA / EXIF / XMP metadata from PNG, JPEG, WebP, SVG, PDF, DOCX, ODT, HTML, and Markdown files.
The watermark removal covers AI services from the three major mainstream providers: Claude, Gemini, and OpenAI. An interesting detail: the project was originally named remove-claude-marks and later renamed to the current watermarks-remover.
Claude Refuses Installation
As an Agent Skill, most users would let their agents install it directly. However, when a user attempted to have Claude install this remover skill, Claude directly refused.
He explained that Anthropic users never agreed to be forced to accept watermarks; this was unilaterally imposed by EU regulations. Claude remained unmoved. He emphasized that paying customers do not want their outputs labeled. Claude still refused. He threatened to use unmoderated Chinese models to complete the task anyway. Claude still did not cooperate.
In the end, GLM 5.2 took over the job and successfully completed the skill installation.
A very ironic point is that the code for this watermark removal Skill likely came from Claude itself.
A Cat-and-Mouse Game That Is Doomed to Never Stop
Those opposed to watermarks argue: I am a paying user, I paid for the output, how can you put a mark on it? Watermarks create an invisible "second-class citizen" status for AI content. In scenarios like job application letters, academic papers, and commercial copywriting, even self-written text risks being entirely questioned if it has ever been handed over to an AI.
Those supporting watermarks argue: Deepfakes and AI-generated disinformation are proliferating; traceability is necessary public infrastructure. "You own the right to use the content" and "you have the right to conceal the source of the content" are fundamentally two different issues.
As long as AI watermarks exist, watermark removal tools will follow. The more popular the removal tools become, the more they prove there is "something to remove" with watermarks, which in turn strengthens the rationale for increased regulation.
However, the speed at which the open-source community votes with its feet will always outpace the speed at which regulations are made.
When the quality of AI-generated content is already indistinguishable from human writing, does it make sense to use technological means to forcibly label all AI output?
This article is from WeChat public account "Machine Heart" (ID:almosthuman2014), author: Leng Mao
