Developers share tools to remove Claude’s invisible text watermarks

In short: Soon after Anthropic confirmed that Claude will add invisible watermarks to its text, developers began sharing ways to remove or weaken them.

What's going on

Anthropic says its Claude AI models will embed an invisible, machine-readable watermark in AI generated text. Think of it like a hidden stamp that people cannot see, but software can look for.

Within hours of that confirmation, developer Guillaume Meyer published code that aims to strip those watermarks by rewriting the text. His approach uses another AI model to make multiple rewrites, swapping in synonyms and lightly reorganizing sentences. The goal is to change the word pattern enough that the watermark no longer shows up.

Other developers have posted similar methods. One tool removes invisible or look-alike characters and also shuffles sentence order and replaces words with synonyms. Another method described in the report is to condense the text, translate it into a very different language, and then translate it back. Anthropic has said that heavy editing, paraphrasing, or translation may prevent its watermark from surviving.

Anthropic says it is adding watermarking to comply with the European Union’s AI Act. The rules require AI model providers to label synthetic content, meaning content made by AI, so machines can detect it. Companies that do not comply could face fines of up to 3 percent of annual turnover.

What to watch

It is not yet clear how well these removal tools work, because Anthropic has not released its watermark detection tool for the public. Anthropic says it plans to ship a text detection API, meaning a tool other software can call like a simple service, which could make it easier to test what edits still keep or break the watermark.

Source: Wired