DEV Community

Dragos Roua
Dragos Roua

Posted on Originally published at dragosroua.com

How AI Watermarking Works? Can It Be Removed?

Short answer: yes, it can be removed.

Long answer, it’s a bit more complicated. To get the full picture, just watch the video, it’s roughly 5 minutes.

Key points:

  • Watermarking is implemented because EU regulations
  • It survives copy and paste, because it’s based on probabilistic token choices, not on invisible characters
  • Heavily editing the text basically removes this type of watermarking
  • If you want to avoid it entirely, using local, open source models is the safest way

In this video:

  • How Anthropic’s implementation (using Google’s SynthID) actually works
  • Why longer text is much easier to detect than short replies
  • The surprising side-effect: even human text you paste into Claude for proofreading can get watermarked
  • Two practical ways to remove or avoid the watermark

Timestamps:

0:00 – Intro: What AI watermarking is

0:55 – How the watermark is applied (token-level changes)

2:28 – Why short chats are safer than long-form content

2:48 – The proofreading trap

3:13 – Method 1: Heavy human editing (recommended)

3:48 – Method 2: Open-source models (with an important caveat)

5:14 – EU regulation context + final thoughts

Key takeaway: Watermarking is not baked into the model weights — it is applied after generation.

Local open-source models on your own hardware currently give the strongest protection.

If you found this kind of content useful, like and subscribe to my channel, it really makes me wanna publish more.

Top comments (0)