Anthropic Adds Invisible Watermarks to AI-Generated Claude Content
Anthropic, the artificial intelligence company behind the Claude model, says content generated by newer versions of Claude will include an invisible digital watermark designed to identify AI-generated text and images.
How Anthropic’s AI Watermarking Works
Text produced by Claude models launched after August 2 will contain an embedded watermark indicating that the content was generated by AI. Images created by Claude generally include metadata with a digital signature showing that the model processed the file.
Claude’s text watermarking system uses algorithms that subtly influence how the model selects words. Across a large body of text, these changes create a statistically detectable pattern. Anthropic says the watermark does not affect the meaning, quality, or readability of Claude’s responses and may remain visible to detection systems even after the content has been edited.
The San Francisco-based company plans to apply these watermarks to Claude-generated content worldwide.
Anthropic’s Watermarks and the EU AI Act
The move comes as providers of advanced AI models prepare to comply with the European Union’s AI regulations. Under the EU AI Act, providers of frontier AI systems must make AI-generated content detectable or risk fines of up to €15 million, approximately US$17 million, or 3% of their annual global turnover.
AI models released after August 2 must meet the requirements immediately, while models already available on the market have until December 2 to comply.
AI Watermarks Are Not Definitive Proof
The presence of a Claude watermark does not necessarily reveal how the model was used. Anthropic says a detected watermark provides a signal that Claude generated or processed the content, but it is not conclusive evidence.
For example, a person may have used Claude only to summarize, translate, or refine an idea originally created by a human. Likewise, the absence of a watermark does not prove that text was not generated by AI.
Because Claude’s watermark relies on subtle patterns in word selection, the signal may disappear in very short passages or after extensive paraphrasing and rewriting.
AI “Humanizer” Tools Can Remove Watermarks
Researchers remain concerned that AI watermarking may not prevent people from using generative AI to produce fraudulent or low-quality academic papers, sometimes referred to as AI slop.
Reese Richardson, a metascientist at Northwestern University, said motivated users could potentially remove watermarks by running AI-generated text through another model or using an AI “humanizer” tool.
However, Nihar Shah, a computer scientist at Carnegie Mellon University who studies research evaluation, said watermark detection could still help identify some misuse if AI companies provide reliable verification tools with low false-positive rates.
Potential Uses in Academic Publishing
AI watermarks could help academic journals, universities, and conference organizers enforce policies restricting the use of AI-generated content. The 2026 International Conference on Machine Learning, or ICML, used watermarking in one of its peer-review tracks.
Conference organizers added watermarks to papers circulated for review and generated evidence when AI was used to prepare peer-review reports. The process reportedly identified 506 reviewers who violated the conference’s AI-free policy.
Shah, who helped develop ICML’s watermarking process, said the experience suggests that while some users may deliberately attempt to evade detection, others may simply copy and paste AI-generated content without removing its digital trace.
The Future of Invisible AI Watermarks
Anthropic’s system represents one of the growing efforts to make AI-generated content easier to identify. Although invisible watermarks are not foolproof, they could provide valuable signals for content platforms, educators, publishers, and regulators seeking greater transparency around AI-generated text and images.
Source: www.nature.com