Anthropic’s new watermark for Claude’s generated text has drawn a lot of controversy. Over the weekend, the Dario Amodei-led AI firm has observed many of its loyal subscribers cancel their memberships, allegedly due to the AI text watermark impacting their workflow, as seen on social media feeds and forums. Anthropic has now explained how its system works and even given out a cheat sheet on how one can avoid the AI watermark on Claude-generated text.
In compliance with Article 50 of the European Union AI Act, Anthropic has officially explained how it embeds machine-readable watermarks into text generated by Claude. While the system allows platforms to trace AI-generated copy back to the model, Anthropic candidly acknowledged the technology’s core limitation – it remains simple to defeat.
What is Claude’s text watermark?
Unlike photos or videos, which hide invisible digital certificates inside image files, plain text contains only letters, numbers, and spaces. There are no hidden tracking codes or invisible Unicode characters added to the output.
However, the watermark is an imperceptible statistical pattern hidden inside word choices. Here’s how it works:
The mechanism: When writing a sentence, an AI often faces multiple natural-sounding options. For instance, finishing the sentence “The weather today was cold and…” could easily end with “overcast” or “grey”.
The secret key: Normally, the model picks between equally good choices using a simple random generator. Anthropic’s new system replaces standard randomness with a secret mathematical key that factors in the preceding words.
The result: With the secret key embedded, the output reads completely naturally to humans. However, someone holding Anthropic’s detector key can check the text to see if the word sequence matches the specific mathematical choices Claude makes under that key.
How to bypass or remove the watermark in Anthropic style
Because the watermark relies entirely on the exact order of words, altering the sequence breaks the mathematical chain the detector looks for. Anthropic has shared several scenarios where the signal naturally weakens or disappears:
Paraphrasing, or light editing: If you consider swapping a few words for synonyms, or altering sentence structures, or even converting paragraphs into bullet points, it changes the preceding context, immediately breaking the statistical pattern.
Redoing Claude’s text with another AI: Copying Claude’s text into a different AI tool (or asking another model to rewrite it) replaces the word choices with a brand-new randomness pattern, wiping out the original signal entirely.
Translating back and forth: Translating the text into another language and back into English resets the vocabulary sequence.
Short passages or factual writing: Claude’s secret watermark needs a long string of open-ended text to leave a strong signal. On short answers, factual lists, or strict grammar edits, there aren’t enough word-choice options for a watermark to register in the first place. Hence, if you keep your answers short or in a basic ‘factual style’, your text should be free from watermarking.
Our recommendation:
Regardless of these tips and tricks to break the secret key in Claude’s generated text, we recommend using Claude’s generated text only as reference material for your work, and not copy-paste its suggestions directly into your work. This way, your work retains your originality while taking all the help from the advanced algorithms of the AI.