BIG CHANGE: Claude introduces hidden watermarks to his AI generated texts!

Adaptation to the European Act on Artificial Intelligence

The main movementč of this decision is the Act on Artificial Intelligence of the European Union, it is not Article 50 which refers to transparency. The regulation requires that providers of generative AI systems must ensure that their results are marked in a machine-readable format so that they can be identified as artificially generated. The goal is to provide clarity to users when communicating with artificial intelligence and to suppress the spread of deepfake content.

The need for such measures became apparent after numerous platforms faced criticism for unmarked AI content. For example, AI music platform Suno announced song tagging after a series of legal challenges, while newsletter service Substack teamed up with Pangram to tag AI-generated content. With this move, Anthropic joined other technology giants such as Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia, which also committed to complying with European rules.

How invisible tagging works

The water žig that Anthropic introduces is not visible to the human eye. Instead, the wordč is about an “imperceptible” signal that is embedded in the statistical patterns of the text itself, influencing subtle elements such as word choice or sentence order. The technology will be automatically applied to all models released after August 2, and the company has announced that it will extend support to older models as well. For files like images and documents, Anthropic will use the C2PA (Coalition for Content Provenance and Authenticity) open standard to add signed tamper-resistant metadata.

In an updated post on its support page, the company clarified how the ig will work: “Since the watermark ig is part of the text, it will travel with the text when copied and pasted elsewhere and may persist after certain changes. Markup will be applied at the model level, which meaning that it will be present regardless of which Claude product or platform the text comes from”.

The very idea of ​​hiding information within digital content, known as steganography, is not new, but its application to text presents a unique challenge. Unlike images, where data can be subtly altered without the human eye noticing, text is a much more complex medium, making it difficult to embed images without affecting quality. In response to public concerns, Anthropic has confirmed that it will also offer users a detection API, with the help of which they can check whether there is a watermark in the text.

User concerns and technical limitations

This announcement also caused significant concern among users. On social networks, they expressed fear that their work could be unfairly labeled as AI-generated, especially when they use tools to help; during proofreading or formatting. Critics claim that a system that produces a “weak positive signal and an insignificant negative signal” can be abused. This means that schools, employers and platforms could use the presence of AI as conclusive evidence of the use of artificial intelligence, which could lead to light charges. Anthropic themselves admit that the absence of an aquatic logo is not conclusive evidence that a text was written by a human, just as its presence does not fully explain the extent to which AI was involved.

Global trend towards transparency

Anthropica’s move is part of a wider global trend. China, for example, introduced regulations in September 2025 that state that all AI-generated content must be clearly marked, with visible tags for users and technical identifiers in metadata. In the United States, there is no single federal law, but there are numerous initiatives underway at the state level, such as the California Artificial Intelligence Transparency Act. At the federal level, bills have also been introduced that would charge the National Institute of Standards and Technology (NIST) with developing guidelines for water resources. The United Kingdom does not currently have specific laws requiring labeling, but the topic is actively being discussed.