Anthropic has started adding invisible, machine-readable markings to content created by its Claude AI models, giving the material a hidden digital signature that can be used to identify its origin. The move covers text as well as supported files and comes as governments and technology companies look for better ways to bring transparency to the rapidly expanding use of generative AI.
The watermark is not something users will see on the screen. Instead, Anthropic has embedded an imperceptible signal into Claude-generated text at the model level. The company says it is designed to survive copying, pasting and some minor editing, while leaving the content’s meaning and readability unchanged.
For anyone using Claude to write an email, report, article, essay or piece of code, the immediate experience is expected to remain the same. The difference is that the output now carries information that could potentially be used later to establish that it came from Claude.
That distinction matters as AI-generated content becomes increasingly difficult to recognise by appearance alone. Traditional AI detectors have often struggled with accuracy, particularly when AI-generated material has been edited by a person. Anthropic’s approach is different because it attempts to mark the content when it is produced rather than asking a detector to guess whether something looks machine-written.
The company says the system is intended to support AI transparency, rather than simply catch people using AI. But its potential applications are broad. Publishers could use provenance information when assessing submissions, companies could better understand the origin of documents, and educational institutions could have an additional tool when investigating suspected AI-generated assignments.
The move also comes amid growing concern about AI slop, a term increasingly used to describe huge volumes of cheaply generated, repetitive or low-quality online content. Search engines, social platforms and content publishers are dealing with an internet where the amount of machine-generated material is growing rapidly. A reliable provenance system could make it easier to distinguish AI-generated material from human-created work.
Claude’s watermarking is not limited to the chatbot itself. Because the marking is applied at the model level, it can also appear in outputs produced through Anthropic’s API and products such as Claude Code and Claude Cowork. The system can also carry over when supported Claude models are accessed through cloud services such as AWS, Google Cloud and Microsoft Foundry.
Anthropic is using a separate technology for images and other supported files. These outputs can contain digitally signed provenance metadata using the C2PA standard, an open framework designed to record information about the origin and history of digital content. This is increasingly important for AI-generated images, where conventional visual detection can be difficult.
The change is being driven largely by regulation in Europe. Anthropic has signed on to the EU’s Code of Practice on Transparency of AI-Generated Content, linked to the EU AI Act. From August 2, new Claude models launched in the European market have been required to support machine-readable marking of AI-generated content. Anthropic has chosen to apply the technology globally rather than restrict it to Europe.
That decision has not been universally welcomed. Some developers and users have questioned whether a permanent or persistent AI marker could create problems for people who use Claude only as an editing or brainstorming tool. A document that is substantially written by a human but lightly processed by Claude could potentially carry a signal indicating AI involvement.
There are also practical questions about how difficult the watermark will be to remove. Anthropic says the marking can survive ordinary copying and some editing, but it may not survive heavy paraphrasing, translation or extensive rewriting. Similarly, digital provenance metadata attached to files can sometimes be removed during processing or uploading.
Another unresolved issue is access to detection. Anthropic currently controls the technology needed to identify the text watermark, although it has said it plans to make detection tools available to users and third parties. That has prompted concerns about whether a single AI company should effectively become the authority on whether a piece of text originated from its own systems.
Still, the move represents a broader change in the AI industry. The focus is no longer only on making generative AI more capable, but also on making its output more traceable. Google DeepMind has pursued similar watermarking technology, while other major AI companies have backed broader transparency initiatives.
