In a significant move towards transparency in artificial intelligence, Anthropic has announced that its Claude AI models will soon embed watermarks in all generated text and files. This initiative aims to comply with the EU’s Code of Practice on Transparency of AI-generated Content and to inform consumers when they are interacting with AI.
The EU’s regulations require companies to include watermarks on AI content and to inform customers about the use of deepfakes, emotion-recognition, and biometric-categorization tools. Nearly 200 organizations have agreed to these measures, highlighting the growing importance of AI transparency.
Understanding AI Watermarking
Watermarking in the context of AI is a process that embeds markers in text indicating that the content originated from AI. These markers are invisible to readers but can be detected by computer systems. Examples include different spacing between words and numbers, word sequencing, and other signals that remain with the text even if it is copied and pasted across different platforms.
Anthropic has stated that Claude models launched on or after August 2, 2026, will include watermarking. This applies to all content created by Claude through various platforms, including Claude, Claude Code, Claude Cowork, and Claude Tag. The watermarks will be applied globally, not just within the EU.
The Process of Text Watermarking
Anthropic’s text watermarking process involves embedding markers in the text that are detectable by AI detection tools. These markers do not affect the quality or readability of the content. In internal testing, Anthropic found no impact on the creativity or readability of Claude’s text.
The watermarks will remain even if the text is copied and pasted from text editors like Windows Notepad or MacOS’s TextEdit. Human editing may not eliminate these marks, as they are applied at the model level. Anthropic’s tool is based on Google‘s SynthID system, which detects AI in text, images, video, and audio.
Watermarking Images and Files
For images and files, Anthropic will attach cryptographically signed notes in the metadata of files such as.png,.jpg, or.svg. This practice, known as C2PA, is standard in photo-editing software and camera manufacturers to record the origin of an image. If someone tampers with the metadata, the cryptographic signature will break, indicating tampering.
However, watermarking is not foolproof. The presence of a watermark does not necessarily mean the content was created by Claude, nor does the absence of a watermark mean Claude was not involved. For example, if someone rewords an essay using Claude, the output will have a watermark, but the content’s facts and details may come from a human writer.
Anthropic has also noted that content generated or processed by Claude might not have a watermark if the AI-generated content is too small or if the content has been heavily edited, paraphrased, translated, or mixed into other writing. Similarly, a screenshot of a Claude-generated image may not retain the metadata, making it appear as if it lacks a watermark.
The initiative by Anthropic to embed watermarks in AI-generated content is a step towards greater transparency and compliance with regulations. As AI technology continues to evolve, such measures will become increasingly important to ensure the integrity and trustworthiness of digital content.
