TLDR: Anthropic now embeds an invisible watermark in text generated by new Claude models and signs image files using the C2PA standard, a move tied to the EU AI Act's transparency rules. Google, OpenAI, Meta, Microsoft and others have made similar commitments, so this is quickly becoming standard practice across the industry rather than something unique to Claude.
Why this matters: As AI writing and images spread across the web, schools, employers and platforms increasingly need a way to tell what came from a model versus a person. Watermarking gives that a technical foundation, but it is not foolproof. Knowing how it works, and where it falls short, helps you understand what these labels can and cannot prove.
Claude starts marking its own text
Anthropic confirmed this week that Claude models launched on or after August 2, 2026 automatically embed a machine readable watermark into the text they generate. The signal is invisible to readers, does not change the meaning or quality of the writing, and is designed to survive copy and paste. The change applies across Claude.ai, the API, Claude Code, Claude Cowork and Claude Tag, and Anthropic says it applies worldwide rather than only to users in the EU.
For images and other supported files, Anthropic is using a different approach: signed provenance metadata based on the C2PA standard, an open framework already used by Adobe, Google and Microsoft to record how a piece of content was created or edited.
Why now: the EU AI Act
The push comes from Article 50 of the EU AI Act, which took effect August 2 and requires AI providers to mark generated content so it can be identified downstream. Anthropic signed the EU's Code of Practice on AI generated content transparency, and according to reporting from Quartz, it is far from alone. Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have all made similar pledges under the same code.
That last point is worth underlining for anyone assuming this is a Claude specific feature. Google has used its SynthID watermarking system in image and text tools for a while, and C2PA metadata is already common practice among major AI image generators. Anthropic's announcement effectively brings Claude's text output in line with where the rest of the industry is heading.
What the watermark does not prove
Anthropic is upfront that a detected watermark is not conclusive proof of authorship. Claude may have simply proofread or translated a human's own writing, and heavy editing, format conversions or asking a different model to rewrite the text can strip the signal entirely. The company has also not yet released a public detection tool, so for now there is no way for an outside reader to check a piece of text against the watermark directly.
Still, this is a meaningful first step toward giving platforms and regulators a consistent signal to work with, even if it is not a perfect one.
Source: TechCrunch