Claude’s text is getting an invisible signature. Anthropic will watermark outputs from its Claude models to meet the European Union’s AI Act requirements, adding machine-readable evidence that text came from an artificial system. Readers will not see the pattern, and Anthropic says the change will not affect output quality or user costs.
The watermarking feature is being introduced alongside C2PA support for Claude-processed images. Anthropic says both measures meet its obligations under the European Union’s AI Act, which requires synthetic audio, image, video, and text to carry machine-readable marks that enable detection as artificially generated or manipulated.
That turns provenance from a nice promise into a technical requirement. The internet, having spent years asking whether content was made by a person, now gets invisible tags instead. Progress has a funny way of looking like paperwork with better cryptography.
How Claude’s watermark works
Claude’s system uses a version of the open-source SynthID-Text approach developed by Google DeepMind. Google’s technology creates detectable patterns through wording probabilities, using low-stakes choices in text to leave a signature that can be identified only with a key.
Anthropic says the choices in watermarked text remain random, but the source of that randomness changes when watermarking is active. Instead of relying on an arbitrary random number generator, Claude bases those choices on a key and the words that came before them.
That pattern remains invisible to readers while staying detectable with the correct key. Anthropic describes the system as a way to identify generated text without changing what users read, which matters because an obvious marker would make the feature easy to remove and annoying to live with.
Anthropic states plainly: “Watermarking does not impact the quality of Claude’s output.” It also says the feature will not make Claude more expensive for users, so compliance will not arrive with a new line item on the bill.
Watermarks will help, but they are not magic
The watermark can survive light editing, though Anthropic says light editing may not remove it completely. A complete rewrite, where every word is replaced, will remove the watermark.
That limitation places the system somewhere between a durable fingerprint and a polite suggestion. It can help identify text that remains close to its original form, but it cannot prove authorship after someone rewrites the entire passage. Pangram offers AI detection approaches, yet no detector gets to repeal the basic fact that rewritten language leaves less evidence behind.
Code presents another boundary. The watermark will be less present in code because programming syntax leaves fewer arbitrary wording choices, although Anthropic says it can work in places where terms are flexible, such as comments within code.
Anthropic described that distinction directly: “Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code.” The system has room to operate where language offers choice; rigid syntax gives it less room to hide a pattern.
The announcement arrives as people show visible interest in identifying machine-made material. Over 1 million people have clicked LinkedIn’s AI slop button, a figure that suggests readers do not need another lecture about why provenance matters. They want a way to tell what they are looking at.
Anthropic is not planning to solve that problem alone. Other major model developers have signed the same Code of Practice and will implement their own watermarks, creating a shared direction for machine-readable identification across generated content.
Claude’s watermark will not stop text from being copied, edited, or rewritten. It will give detection systems a hidden signal when the original wording remains intact — a modest tool, but more useful than asking readers to guess.
Based on
