On August 13, 2026, Ars Technica and TechCrunch reported that Anthropic will add invisible watermarks to all content processed by its AI models, including Claude. The stamp does not distinguish between text created from scratch and text the model only edited, and the measure is already generating mixed reactions among users.
The decision responds to the EU AI Act, which requires providers of AI systems to mark content generated or manipulated by their models. The rule applies to models launched after August 2, 2026, and gives until December to update existing ones. Anthropic confirmed that all its new models, offered globally and not only in the European bloc, will carry the stamp from day one: texts will carry invisible embedded watermarks and other files will include C2PA provenance metadata. For the Latin community using AI tools in offices and classrooms in Mexico and the United States, the measure adds a detection layer that did not previously exist.
User frustration has already become visible. TechCrunch documented dozens of Reddit comments this week from people who fear being discovered using AI at work or school. One user argued the stamp will catch the student who asked to reorganize a paragraph or the office worker who summarized a long document, while others defended the watermark as a way to track algorithm-generated material. There were also criticisms for applying the label to edited text, a case the European guidelines themselves exclude when the system serves an auxiliary function, such as grammar correction. Anthropic specified that the watermark will not work on some platforms or features that do not support it.
Anthropic said it will publish details on how to detect the watermarks to provide the technical support required by European law. Until then, the actual scope of the stamp will not be verifiable, and that will be the signal of whether the label distinguishes generated text from merely corrected text.
This article was written with artificial intelligence assistance from verified sources and reviewed by a human editor before publication.