Many programmers are creating watermark removal tools after Anthropic rolled out Claude's text output watermarking feature.
Last week, Anthropic announced the adoption of an "invisible" marking mechanism, a form of watermark to distinguish content created by AI and humans. Specifically, the company established patterns in Claude's choice of words and phrases that normal users could not detect, but could be detected by someone with the encryption "key". The watermark will follow when text is copied and pasted, helping with traceability.
While many people praised the new move for increasing transparency, others said that the markup method affected the quality of Claude's output, not to mention that watermarks could be attached to the text even if they only used chatbots to proofread, translate or summarize. Some reported canceling their subscriptions to Claude because of this.
According to Business Insider, interest in the keyword "AI watermark removal tool" in the US on Google Trends increased by 60% compared to last week. Even many programmers and entrepreneurs have quickly developed AI label removal tools.
Guillaume Meyer, developer of AI startup Memo, launched the open source project "Watermarks Remover" a few days after Anthropic announced its plans. The tool removes metadata and hidden characters, then rewrites the text so that the meaning is preserved, thereby breaking word choice patterns that may contain watermarks.
It took Meyer about 5 hours to develop the first version and the tool currently has more than 14,000 "stars" on the GitHub developer platform, equivalent to 14,000 "likes".
"I'm all for attribution, but against watermarking. The two things are very different," Meyer told Business Insider. AI can leave watermarks when creating entire documents, he explains, but it can also do so when people only use it for light editing. He calls this "a false solution to a real problem".
In this week's post on
Tokyo-based software developer Ansh Aneja also deployed a watermark removal tool focused on Claude the same day Anthropic announced its plans, then released an open source version of MarkScrub. In a post on X last week, Aneja announced that the tool grew from 0 to 8,500 users in just one day.
Sharing with Wired, engineer Erik Hughes said it only took him 15 minutes to create a tool that works with Claude to remove invisible and similar-looking characters, rearrange sentences in paragraphs, and replace some words with synonyms.
According to Leon Chlon, a visiting researcher at Oxford University, another option to remove watermarks is to summarize Claude's feedback, translate it into another language such as Arabic, which is different from English, and then translate it back.
Anthropic also acknowledges that content that has been heavily edited, reworded or translated into other languages may no longer carry a watermark. This happens because the marking mechanism is built around Claude's choice of words and phrases.
In a blog post, the startup explains, large language models like Claude work by generating words one word at a time. When it needs to decide on the next word, the model selects from a list of suitable "candidates" based on the previous paragraph. For example, the next word in the sentence "It's cold today and...", is more likely to be "overcast" or "gray". Which word pattern is chosen usually does not matter to the reader, because the meaning of the sentence in both cases is essentially the same.
Anthropic's invisible marking technique takes advantage of such low-risk choices, which appear repeatedly in the output text, to create a pattern in Claude's responses. The choices are still made randomly, but the source of the randomness is different. Instead of using an arbitrary random number generator to display the next word, the watermark uses a "key" and some preceding words to decide. The person holding the "key" checks the word order and sees whether it is consistent with the choices Claude makes in the watermark. If so, the text was probably generated by AI.
Anthropic notes, the model will not always lean toward "overcast" or "gray." As with unmarked text, "gloomy" can appear in one sentence and "gray" in another, depending on the preceding words. The invisible marking technique also does not force Claude to choose an unreasonable word that the model did not consider.
An Anthropic spokesperson told Wired the company added identification marks to Claude's output to comply with the European Union's (EU) AI Act, and other companies are doing the same.
"Identifying AI-generated text is difficult, the new move will give people the tools to identify. Text from supported Claude models, including Claude Code, will contain invisible marks. It does not change the meaning, quality or readability of the response from Claude," an Anthropic spokesperson said, adding that the company plans to provide a text detection application programming interface (API) so users can be more proactive in text recognition. AI.
According to Konrad Kollnig, an associate professor at Maastricht University's Law and Technology Lab, neither the EU's AI Act nor the provisions of Anthropic explicitly prohibit people from creating or sharing watermark removal tools.
However, Kollnig warns that users can get into trouble if they use them to swap the origin, turning AI products into human ones. Anthropic's usage policy prohibits impersonation of others by presenting output from a model as if it were a human creation.
Anthropic is not the only company applying technology to help identify AI content. Previously, Google used SynthID technology to "stamp" artificial intelligence products, OpenAI also applied to images and sounds. However, Google began to relax when adding options for users to actively turn on and off watermarks on photos, videos, and music created by Gemini. Elon Musk's social network X also added the label "Created with AI" to content produced using this technology. Last week, music streaming service Spotify began labeling profiles containing artificial intelligence content with an "AI Persona" label.
