Anthropic Will Start Tagging Claude's AI-Generated Text and Images

Anthropic has pledged to embed invisible watermarks and origin information into text and images generated by its Claude AI models, aiming to comply with new European Union transparency rules that took effect on August 2nd. AI products that launched before that date have a four-month grace period, giving them until early December to comply (The Verge).

The system works in two ways. For text, Anthropic weaves a hidden watermark directly into what Claude produces, without changing the meaning, quality, or readability of the response (The Verge; Claude Help Center). Think of it as a digital stamp baked into the text itself. Because it is applied at the model level, it travels with the text when copied and pasted elsewhere and may survive at least some editing, according to Anthropic's documentation. For images, the company applies C2PA, a standard from the Coalition for Content Provenance and Authenticity already used by Adobe, OpenAI, and Google. Generated files also include digitally signed information about where they came from, where the file format allows it (Claude Help Center).

The marks are invisible to human readers but are designed to make Claude-generated content detectable by people and online platforms that have the right tools. Anthropic says it is working to let users and third parties detect the watermarks and origin information embedded in Claude output, with technical documentation promised for a later date (The Verge).

Rollout is phased. New Claude models will carry the markings from day one. For existing models, Anthropic describes support as a work in progress. The markings are applied globally across supported Claude products, including the Claude Platform API, the consumer interface, Claude Code, Claude Cowork, and Claude Tag (The Verge).

Anthropic is upfront about the limitations. The company acknowledges that its marking systems are not perfect and that content without detectable marks could still have been produced by an AI. This is a notable concession: the absence of a watermark cannot be treated as proof that a human wrote something, which makes it hard to build a simple pass-or-fail system for distinguishing AI content from human content.

The two approaches are technically different. C2PA for images works by attaching signed metadata to a file, recording its origin and any changes. It is well-understood and already used across the industry, though it can be stripped when files are re-encoded or screenshotted. Text watermarking at the model level works differently. Anthropic has not yet published how it works, but the claim that it survives copy-paste and some editing suggests the signal is hidden in word choice or text patterns rather than wrapped around the file as metadata. That could make it harder to remove when text moves between environments, but also harder to verify independently without Anthropic's own detection tools, which are not yet available.

The EU rules require providers of general-purpose AI models to mark AI-generated content in a machine-readable format. The August 2nd effective date triggered those obligations, and the grace period gives existing products until early December to comply. Anthropic's announcement positions Claude to meet that deadline, though the phased rollout means coverage will not be complete on day one.

Across the broader landscape, C2PA for images is being adopted by major model providers, which makes checking image origins simpler. Text watermarking is more fragmented. OpenAI has discussed watermarking for ChatGPT output but has faced tension between detection usefulness and competitive concerns; a watermark that third parties can detect could be used to filter or penalize AI-generated content at scale, which may not suit every provider's interests. Anthropic's decision to ship both layers signals a willingness to accept that trade-off, at least for now.

In this author's view, the practical value here is a detection path for Claude output that does not depend on the goodwill of the person publishing the text. If the watermark holds up under real-world copying and editing, and if Anthropic follows through with accessible detection tools, platforms operating in the EU will have a concrete signal to work with rather than software that guesses at authorship. The caveat is that Anthropic itself disclaims infallibility, and the system's value depends on adoption beyond Anthropic's own products.