Anthropic, the developer of the popular artificial intelligence Claude, has announced the implementation of a mandatory digital labeling system for all generated content. This decision, initially dictated by the requirements of the European Union under the AI Act, will be applied globally, affecting users worldwide. The goal of the initiative is to increase transparency and combat misinformation; however, experts warn of new challenges regarding privacy and copyright.
Dual Identification System
To reliably identify generated material, Anthropic uses two complementary methods. The first is hidden watermarks directly within the text. When generating responses, the algorithm embeds a digital code invisible to the human eye into the text structure. The company emphasizes that this process does not alter the content, quality, or readability of the text, yet the watermark persists even when the material is copied, pasted, or partially edited.
The second method concerns visual files. When creating images (in .svg, .png, or .jpg formats), encrypted metadata is added according to the open C2PA (Coalition for Content Provenance and Authenticity) standard. This allows verification that the file was processed by the Claude service and helps detect subsequent modifications to the image.
Contradictory Data
Disagreements have arisen regarding the effectiveness of the new technology. On one hand, Anthropic representatives claim that their system makes the watermarking practically unremovable, a claim supported by sources stating that removing the mark is "almost impossible." On the other hand, the company's technical documentation contains a caveat: the presence or absence of digital watermarking is not absolute proof of the content's origin.
Furthermore, there is a risk of the watermark being lost. The marking may disappear or become unreadable in cases of aggressive editing, format conversion, or specific data manipulations. This creates a situation where the absence of a watermark does not guarantee human authorship, nor does its presence always unequivocally indicate full AI generation if the original text was only slightly edited.
Questions of Authorship and Ethics
An important aspect of the new policy is the clarification of authorship status. The presence of a Claude mark indicates only that the text was processed by the model—for example, during proofreading, translation, or summarization. This means that the ideas or original text may have belonged to a human, yet the system will still apply the AI tag. This creates an ethical dilemma: how to distinguish between fully generated text and text that the AI merely slightly improved?
The company is also developing detection tools that allow users and third-party services to check for the presence of Anthropic watermarks in texts or files. This should help platforms and media outlets identify content created with the help of neural networks more quickly.
Consequences for Developers
For developers creating their own products based on the Claude API, Anthropic will release additional technical instructions. This is necessary to help them comply with the transparency requirements of the EU AI Act. The implementation of watermarks becomes a mandatory standard, which may affect the architecture of applications using the Claude API and require developers to review their data processing protocols.
