Archive · 23 Aug 2026
Developers Race to Remove Anthropic’s Invisible AI Watermarks
Anthropic plans to put hidden computer-readable patterns into text created by future Claude models.
People would not be able to see these patterns, but special software could detect them.
Soon after the announcement, developers made tools that try to remove the patterns.
Some tools rewrite sentences, change words, or translate the text and back again.
However, nobody can be sure these tools work until Anthropic releases its detection software.
Anthropic says the watermark would only suggest that Claude probably handled the text.
It would not prove that Claude wrote all of it or that another AI did not create it.
Critics worry that detectors could wrongly accuse students, researchers, or job applicants of using AI.
Anthropic plans to embed invisible, machine-readable watermarks in future Claude-generated text.
Developers have released tools that rewrite, reorder, translate, or alter text to try to remove the marks.
Guillaume Meyer’s GitHub tool attracted more than 100 contributors after Anthropic’s announcement.
Anthropic says its watermark would show Claude was likely involved, but not whether Claude wrote or edited the text.
The watermarking effort is linked to European Union transparency requirements and concerns about false positives.
- Who
- Anthropic, developers creating watermark-removal tools, and people concerned about AI-content detection.
- What
- Anthropic is adding invisible watermarks to future Claude outputs while developers are attempting to remove them.
- When
- Anthropic announced the plan earlier this month; the article also says the relevant EU rules went into effect on August 2, 2026.
- Where
- The effort concerns Claude-generated content and European Union AI transparency requirements; tools have appeared online, including on GitHub.
- Why
- Anthropic says the watermarks support transparency obligations under the EU AI Act, while developers are responding to concerns about detection errors and false accusations.
Transparency and detection advocates
Watermarking critics and circumvention developers
Purpose of watermarks
Transparency and detection advocates
Anthropic says marking Claude’s output helps meet EU transparency obligations and gives people better tools to identify AI involvement.
Watermarking critics and circumvention developers
Critics argue that blanket watermarking and detection can produce false positives and unfairly affect students, researchers, or job applicants.
Reliability of identification
Transparency and detection advocates
Anthropic says machine-readable marks can help identify content from supported Claude models without changing its meaning, quality, or readability.
Watermarking critics and circumvention developers
The watermark cannot distinguish between Claude writing content and Claude heavily editing it, and it cannot establish that the text was not produced by another AI model or originally written by a human.
Removal tools
Transparency and detection advocates
Watermarking proponents can use detection systems to identify marked content, although Anthropic’s detector has not yet been released.
Watermarking critics and circumvention developers
Developers are building tools that rewrite text, swap synonyms, reorder sentences, remove unusual characters, or translate text in attempts to evade detection; their effectiveness remains uncertain.
- Watermarked content
- Future supported Claude models, including Claude Code, are expected to add invisible watermarks to text outputs.
- Detection meaning
- Anthropic says a watermark would indicate that Claude was likely involved in processing content, not necessarily that Claude wrote it.
- Main technique
- The method is based on Google DeepMind’s SynthID-Text approach, which patterns word and phrase choices.
- Meyer tool
- A watermark-removal tool by Guillaume Meyer drew more than 100 GitHub contributors, according to the article.
- EU requirement
- Article 50(2) of the EU AI Act requires providers to label synthetic audio, images, video, and text so machines can detect them.
- Potential penalty
- The article says non-compliance with the EU rules could lead to fines of up to 3 percent of annual turnover.
- Planned Anthropic service
- Anthropic says it plans to release a text-detection API.