Anthropic adds invisible watermarks to text generated by Claude models
Anthropic said models released after August 2, 2026 will automatically watermark generated text and files. Because the watermark is embedded in the text, it travels with copy and paste. The reason is straightforward: the EU AI Act transparency rules that took effect the same day require generated content to be marked in an identifiable way.
The watermark sits inside the text
Per a TechCrunch report dated August 11, 2026, Anthropic will automatically watermark both text and file output from all models released after August 2. The company plans to extend support to older models as well.
The watermark is applied at the model level, so it is not tied to a particular interface. The products covered are listed as the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. For files, Anthropic uses the C2PA open standard.
The behavior is described this way in the report: because the mark is part of the text, it travels with the text when copied and pasted elsewhere and may persist through some editing. How much editing it takes to remove the watermark is unclear, and TechCrunch notes it asked the company for comment.
Why now and who it covers
The driver is regulation. Per TechCrunch, the EU AI Act transparency rules took effect on August 2 and require AI companies to mark AI generated or edited content in a way other systems can identify.
The scope is not limited to Anthropic own interfaces. Per a report from The Register the same day, the marking also covers third party providers of Anthropic models, naming AWS, Google Cloud and Microsoft Foundry.
There is no geographic split either. The Register reports that marking will apply to output from supported models wherever Claude is offered, worldwide. A change made to comply with European regulation therefore affects every user, Türkiye included.
TechCrunch lists Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia among the companies that have committed to the same EU rule. This is not a single company decision but a simultaneous compliance move across the sector.
Marked content means measurable content
The most concrete outcome is this: detecting AI generated text stops being guesswork. Until now whether a text came from a model was judged by statistical detection tools, and those tools got it wrong. An embedded mark is a different category. That assessment is ours.
Second, the question of persistence. The source says the mark travels through copy and paste and may survive some editing, but how much editing removes it is unclear. The Register also notes that open source removal tools already exist for the C2PA standard on the file side. The system is not absolute against a determined removal attempt.
Third, the practical consequence for content teams. A text being detectably AI generated does not mean using it is wrong. But it is no longer something that can be hidden. Teams that build their processes on that assumption will not be surprised.
Fourth, the worldwide application. Even where the EU AI Act does not apply directly, the output of the tool you use will carry the same mark. That is the technical reality regardless of what local regulation says.
What it means for businesses in Türkiye
The assessment below is not in the sources, it is our reading.
AI use in content production is widespread here and often goes undisclosed. This change makes that choice technically risky: because the text is marked, undisclosed use becomes detectable after the fact.
The topic also intersects with the commercial advertising regulation that took effect on August 1, 2026, which requires disclosure of AI use in advertising. Now the output itself carries a mark as well. The two need to be handled together, and that connection is our comment.
The practical suggestion: rather than trying to hide AI use, tie it to a written company policy. Document which content types it is used in, at which stage a human reviews it and how it is disclosed to the client. That is a safer position than concealment on both regulatory and reputational grounds.
One technical note: the presence of a mark does not mean the content is poor. What matters is review. A text that a human has checked and whose facts are verified is not a problem because it carries a watermark.
The UNALSOFT take
In the content workflows we build for clients on the agentic AI side, AI use is not something hidden but a defined step. Our approach is this: the model produces, a human verifies, the source is cited. This blog runs on the same rule. Every story names its source and no detail that is not in the source gets written. A watermarking requirement does not change that approach, it only raises the cost for those who prefer concealment. Over the long run we expect the teams building transparent workflows to be the ones who benefit.
Is your AI use policy written down?
Let us make your content workflow auditable together.