Anthropic’s help center now carries a support article titled “How Claude marks AI-generated content.” It describes plans to embed an invisible watermark in text that Claude generates and to attach signed provenance metadata to files it produces. As someone who handles generated content daily, this is not an abstract matter for me either, so this article organizes the underlying legal framework and the state of each company’s response, based on official sources.

Background — EU AI Act Article 50 and the transparency Code of Practice

The starting point is the EU’s AI Act (Regulation (EU) 2024/1689). Article 50(2) obligates providers of AI systems that generate synthetic content to mark their outputs in a machine-readable way.

Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.

This provision entered into application on August 2, 2026. The article only requires technical solutions to be “effective, interoperable, robust and reliable as far as this is technically feasible” and does not prescribe concrete implementations. To fill in the practical side, the European Commission developed the Code of Practice on Transparency of AI-Generated Content, whose final version was published on June 10, 2026. Signing the Code is voluntary, but the Commission and the AI Board have confirmed it as an adequate voluntary tool for demonstrating compliance with Article 50(2) and 50(4), so signatories can rely on adherence to the Code rather than case-by-case assessment.

Anthropic’s support article opens by stating that the company signed the Code “as a provider of both generative AI models and generative AI systems,” positioning the article itself as an explanation of how it plans to put those commitments into practice.

How Claude marks content — two complementary techniques

Claude’s marking combines two techniques of different natures.

1. Watermarks embedded in text

When a supported model generates text, an imperceptible watermark is woven into the text itself. In the article’s own words:

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.

Because the watermark is part of the text, it travels along when text is copied and pasted, and may persist through some editing. It is applied at the model level, so it is present regardless of which product or surface the text comes from. Unlike file metadata, this approach carries no external information that can simply be stripped off.

2. Signed provenance metadata attached to files (C2PA)

When Claude generates a supported file type, such as .svg, .png, or .jpg, it attaches signed provenance metadata. This follows the C2PA (Coalition for Content Provenance and Authenticity) open standard used across the industry for recording content provenance. If a signed label is present, it signals that the file was processed by Claude and lets you detect whether the file has been tampered with.

On detection, work is under way to enable users and third parties to detect Claude’s watermarks and metadata, with details to be shared in forthcoming technical documentation.

What is covered

The scope the support article lays out is as follows.

  • Models: Claude models launched on or after August 2, 2026 support marking at launch. The law includes a transition period for earlier models, and work to add support to them is in progress
  • Products: Output from supported models is covered everywhere Claude is used—Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Embedded watermarks apply to all generated text; provenance metadata applies where Claude supports processing files
  • Cloud partners: Embedded watermarks apply when supported models are accessed through AWS, Google Cloud, or Microsoft Foundry. Signed provenance metadata may not be supported on every platform, depending on each platform’s features
  • Regions: Though the framework originates in EU law, marking applies wherever Claude is offered, worldwide

The limits of a mark — what it shows and what it does not

In its “Limitations” section, the support article is explicit about how far a mark’s meaning reaches. The framing is useful for reading this class of provenance technology in general, not just Claude’s.

When a mark is detected, what can be said is only that the content may have been processed by Claude—not its full provenance.

  • Claude may not be the original author. If Claude is used to proofread, translate, summarize, or convert files, the output carries a mark even when the underlying ideas, text, or data came from elsewhere
  • The content may have changed afterward. Marked content may have been modified, excerpted, or combined with other material after Claude processed it

The absence of a detectable mark is no proof that content was not AI-generated either. Claude-generated content may carry no detectable mark if, for example:

  • It was generated by a model released before marking was supported
  • The text has been heavily edited, paraphrased, translated, or mixed into other writing
  • The passage is very short, leaving too little text for a reliable signal
  • A file’s metadata was stripped through format conversion, re-saving, screenshots, or other means
  • It was produced through a platform, feature, or file type where that marking type wasn’t supported

In other words, the presence or absence of a mark is designed as a probabilistic signal that is conclusive in neither direction. The article closes by urging businesses that deploy Claude in their own products to independently assess what Article 50 requires of them.

Where Google and OpenAI stand

As responses to the same Code of Practice and the same application date, I also checked the other companies’ official materials.

Google deploys its watermarking technology SynthID. According to Google DeepMind’s official page, watermarks are applied to text generated by the Gemini app and web experience, and SynthID is also built into images and video (Gemini) and audio (the music-generation model Lyria and NotebookLM’s podcast generation). A verification portal, SynthID Detector, lets users check uploaded files for watermarks. For text, the technique is described as adjusting probability scores as words are generated.

OpenAI explains its approach in the help center article “Provenance signals (Content Credentials, SynthID) in OpenAI-generated content.” Supported images generated with ChatGPT, Codex, and the OpenAI API carry both C2PA metadata and SynthID watermarks; generated audio carries SynthID watermarks. Two things stand out: OpenAI has adopted SynthID, a technology that originated at Google, and it clearly treats metadata (richer information, but easily removed) and watermarks (more durable through edits, but less informative) as complementary signals of different natures. A verification tool at openai.com/verify and a verification API are already available. For text, OpenAI says only that, consistent with its commitments under the European Commission’s Code, its goal is to expand provenance signals to all modalities including text—so as of this writing, images and audio lead the way.

The current picture: Google has already shipped text watermarking, while Anthropic and OpenAI are implementing it. Meanwhile, the “reading” side is getting built out alongside the “writing” side. As covered in an earlier article, Cloudflare has begun experimenting with its WebMCP Content Credentials pack, which reads the C2PA metadata of images on a page inside the browser and hands it to agents. Between the writers (the AI companies) and the readers (browsers, CDNs, verification portals), a distribution channel for provenance information is taking shape.

Summary

  • Anthropic signed the Code of Practice on Transparency of AI-Generated Content under EU AI Act Article 50(2) and published a support article laying out its plan for machine-readable marking of Claude’s outputs. Article 50(2) entered into application on August 2, 2026
  • The method combines two techniques: an imperceptible watermark woven into generated text at the model level, and C2PA-compliant signed provenance metadata attached to generated files (.svg, .png, .jpg, and so on)
  • Models launched on or after August 2, 2026 support marking at launch, with support for existing models in progress. Coverage spans all products—API, Claude, Claude Code, and more—and applies worldwide, not just in the EU
  • Detection of a mark shows only that Claude may have processed the content; it neither identifies the original author nor proves the content unmodified. Conversely, the absence of a mark does not prove content was not AI-generated. It is a probabilistic signal, conclusive in neither direction
  • Google already watermarks generated text with SynthID; OpenAI runs C2PA plus SynthID for images and audio and has announced expansion to text—the major providers are falling into step under the same Code of Practice

References