Surfacing on August 11, 2026, Anthropic says new Claude models will embed an imperceptible, machine-readable signal in generated text and attach signed C2PA provenance metadata to files — applied worldwide, and driven by the EU AI Act's transparency rules.
2026-08-11 · by Moe Ameen
Anthropic is preparing to mark content that Claude generates. Detailed in its support documentation and surfaced in reporting on August 11, 2026, the plan embeds an imperceptible, machine-readable watermark into text produced by new Claude models, and attaches signed provenance metadata to files Claude generates or processes using the Coalition for Content Provenance and Authenticity (C2PA) open standard. The marking is applied at the model level to models released on or after August 2, 2026, and covers the Claude Platform API, the Claude apps, Claude Code, Claude Cowork, and Claude Tag — plus supported models accessed through AWS, Google Cloud, and Microsoft Foundry. Anthropic says the markings apply worldwide, not only to users in the EU.
The text watermark is designed to be imperceptible to a reader, to leave the meaning, quality, and readability of a response unchanged, and to remain attached when someone copies and pastes the text or makes light edits. For files — including images such as .png, .jpg, and .svg — Claude attaches signed C2PA metadata that records that Claude generated or handled the file and can show whether that provenance data was later altered. Anthropic says it will publish technical details for detecting its watermarks, along with tooling so users and third parties can check content for supported Claude marks.
A crucial nuance drove much of the reaction: the mark shows only that Claude had a hand in a piece of content, not that a Claude model wrote all of it. Because Claude can proofread, edit, or translate writing a person authored, human-written text can pick up the signal too — a point critics latched onto after the news broke. The signal also weakens or disappears with heavy rewriting, paraphrasing, translation, or mixing with substantial human-written material, and a short passage may carry too little signal to detect reliably. C2PA file metadata can be stripped by format conversions, screenshots, or re-saving, and unmarked content is not proof a human made it.
The move follows the EU AI Act's Article 50 transparency requirements, which took effect on August 2, 2026 and push providers of generative AI to make synthetic output machine-readable and detectable; Anthropic signed the European Union's Code of Practice on transparency of AI-generated content. It lands amid an industry-wide scramble to label synthetic media — Suno adding audio watermarking, TikTok tagging billions of clips with Content Credentials, and Substack wiring in an AI-text scanner — as platforms fight to distinguish AI output from human work.
There is an immediate play while the story is hot. "Anthropic is watermarking Claude's output — what it actually means for creators" is a high-intent question this week, and fast recaps flatten the one detail that matters: the mark proves Claude *processed* something, not that it *wrote* it. Drop your take into [Kompozy](/) as a source and the engine fans the story into a [Blog Article](/glossary/output-buckets) for searchers, a Carousel breaking down what the watermark changes for anyone drafting with AI, a captioned explainer clip, and platform-native posts in your voice — generated and queued across the eight social platforms plus blog and email in one pass, published while the news is still current.
The deeper read matters more than the recap. As provenance becomes default, the differentiator is no longer whether AI touched your content — it will have, everywhere — but whether the content is distinctive and actually reaches people. That is what Kompozy is built for: the [Persona Brief](/glossary/persona-brief) governs voice and banned phrases so output reads like you and not median-prompt AI, face-locked [Persona Shorts](/glossary/persona-shorts) and [HyperFrames](/glossary/hyperframes) graphics give a recognizable identity, and [Autopilot](/glossary/autopilot) ships everything behind a per-post review gate where you can bake a consistent disclosure line into whatever you publish. An honest note, since it is the point: Kompozy runs on Claude alongside other models, so text it generates can carry these provenance signals too — the goal is not to dodge the mark but to build an on-brand body of work where disclosure is clean and the AI is a tool, not the tell.
Anthropic will embed an imperceptible, machine-readable watermark into text generated by new Claude models and attach signed C2PA provenance metadata to files Claude generates or processes. The signal identifies content that Claude had a hand in, and Anthropic says it will publish technical details and tools so people can detect supported Claude marks.
The marking is applied at the model level to Claude models released on or after August 2, 2026 — the date the EU AI Act's Article 50 transparency rules took effect. It covers the Claude Platform API, the Claude apps, Claude Code, Claude Cowork, and Claude Tag, plus supported models accessed through AWS, Google Cloud, and Microsoft Foundry, and applies worldwide.
No. Anthropic is explicit that the mark shows only that Claude had a hand in a piece of content, not that a Claude model wrote all of it. Because Claude can proofread, edit, or translate text a person authored, human-written work can carry the signal too — which is why the mark is best read as "Claude processed this," not "AI wrote this."
The text signal weakens or disappears with heavy rewriting, paraphrasing, translation, or mixing with substantial human-written material, and short passages may carry too little to detect reliably; C2PA file metadata can be lost to format conversions, screenshots, or re-saving. Because of that, an unmarked piece is not proof a human made it, and no single detector should be treated as definitive.