Anthropic said this week that every piece of text or image generated by its Claude family of AI models will include an invisible watermark. The feature rolls out to Claude versions launched on or after Aug. 2 and applies across all Claude‑branded products, from the API to Claude Cowork and Claude Tag. By embedding hidden signals directly into the output, the company hopes to satisfy the European Union’s Code of Practice on Transparency of AI‑generated content, which obliges AI providers to inform users when they are interacting with machine‑produced material.
The watermark is not a visible tag that readers can see. Instead, it consists of subtle variations in spacing, word sequencing or other statistical patterns woven into the text. Those cues survive copy‑and‑paste actions, whether the user moves the content from Windows Notepad, macOS TextEdit or any other editor. For images, the watermark takes the form of signed provenance metadata embedded in file formats such as .svg, .png or .jpg. The cryptographic signature records the origin, creator and any subsequent modifications; tampering with the file breaks the signature and signals that the image may have been altered.
How the watermark works
Anthropic described the process as model‑level watermarking, meaning the markers are baked into the output regardless of which Claude interface or surface the user accesses. In practice, a detection system can scan the text or image file and reveal the hidden pattern, confirming that the content originated from Claude. The company has not yet released the detection tools to the public, saying it will provide more details in the future.
While the watermarks are designed to persist through routine editing, Anthropic warned that heavy rewriting, extensive paraphrasing, translation or mixing with other sources could erase or obscure the markers. Likewise, taking a screenshot of a Claude‑generated image and saving it as a new file would strip away the provenance metadata. Because of those limits, the presence of a watermark does not automatically prove that every word or pixel was created by the AI, nor does its absence guarantee the opposite.
Industry observers see the move as part of a broader push for AI transparency. Substack recently partnered with Pangram to label AI‑assisted posts, Suno announced similar disclosures for AI‑generated songs, and LinkedIn now lets users flag suspect AI content. Spotify’s AI Persona feature also tells listeners when music involves artificial intelligence. Anthropic’s watermarking joins those efforts, offering a technical solution that can be verified by machines rather than relying on user‑visible labels alone.
Critics note that AI detection tools have struggled with accuracy, especially for non‑native English speakers whose writing can be mistakenly flagged as synthetic. Anthropic’s approach sidesteps some of those pitfalls by embedding a provable signature at the source, but it also raises questions about how easy it will be for third‑party developers to build reliable detectors.
In a brief statement, an Anthropic representative declined to comment further on implementation details. The company said it will share detection guidance later and emphasized that the watermark will be applied globally, not just within the EU. As regulators worldwide tighten rules on AI disclosure, Anthropic’s step may set a precedent for other developers seeking to balance compliance with user experience.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.