Claude is secretly watermarking your text, and it is almost impossible to remove

Say goodbye to anonymous AI-generated text. Anthropic will start to add special invisible watermarks to text generated by its latest AI models.

Claude AI on a smartphone
Claude is the first AI model to include a watermark in AI-generated text. | Image by Anthropic
0
Say goodbye to anonymous AI-generated text. Anthropic will start to add special invisible watermarks to text generated by its latest AI models. The change comes after the European Union AI Act came into effect on August 2, requiring transparency for AI-generated content.

Future Claude models will add invisible watermarks to your AI-generated texts


Anthropic posted an update to its official support page, stating that all new Claude models launched after August 2 will support something called "machine-readable marking" from the day they launch.



This "watermark" will be embedded into the generated text, but it will be invisible to the users. Anthropic also says that the marking won't affect Claude's response in any way. This "watermark" will work at the model level, applying to Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag, and all platforms that use these AI tools.

Recommended For You
Do you think adding a watermark is a solution for AI-generated content?
13 Votes


Anthropic is rolling the change worldwide



Even though EU regulations prompted the move from Anthropic, the company says the change will apply worldwide.

The "watermark" won't be tied to the way you access Claude. Models accessed through AWS, Google Cloud, or Microsoft Foundry will also have the tag. This "watermark" will stay with the text even if it's copied and pasted and will survive editing, clearing formatting, and other tricks meant to remove it.

Not an AI-checker tool



Anthropic stresses that this new "watermark" is not meant to be a definitive test of whether someone has used AI for creative work or not. A positive result for the "watermark" means that Claude has tinkered with the content. This includes proofreading, grammar checks, translation, summaries, etc.

Vice versa, If the watermark is missing, this doesn't definitively prove that the content is created by a human. Combining Claude-generated content with other AI agents or heavy editing can destroy the "watermark."

Will hidden watermarks finally kill fake human work?



I don't think so, but it's definitely a step in the right direction. Anthropic said that it will publish specific mechanisms and tools for detecting the "watermark," but at the moment these detection mechanisms remain a mystery.

The EU's transparency code, described in Article 50 of the AI Act, requires all AI-generated content to be clearly marked.

Anthropic isn't the sole AI company to comply with these new regulations. Google announced that it will sign the EU Act, and Meta also confirmed it will participate. The EU AI Act includes a six-month transition period for past AI models, so all versions will eventually have to comply with the regulation. For now, the new rules are in effect for all new AI model versions launched after August 2nd.

Do you think adding a "watermark" to AI-generated text will solve the ongoing crisis with people passing AI content as their own?

And in case you're interested in everything AI, check out these:
Six-month unlimited plan is now 57% off
$90
$210
$120 off (57%)
Mint Mobile is now allowing you to get whichever plan you like for either three, six, or 12 months for just $15/mo. If you go for the six-month unlimited service, for instance, you'll now have to pay just $90 upfront instead of $210.
Buy at Mint Mobile
Recommended For You
COMMENTS (0)