If you use artificial intelligence to write an email, a report, an assignment, an article, or even a social media post, there is now something that may accompany your text without you seeing it: an invisible “fingerprint” indicating that it was generated or edited by artificial intelligence (AI).

Anthropic is the first of the major providers to actually implement such a system for Claude’s text. As of August 2, 2026, new Claude models incorporate an invisible, machine-readable watermark into their text as part of compliance with the European Union’s new transparency requirements.

Anthropic has chosen to implement the system worldwide, rather than only for its users in Europe. This is not a visible logo or an indication such as “Written by AI.” To the user, the text still appears completely normal. The watermark is a statistical pattern embedded in the way certain words are selected and can be identified by a specialized tool.

Why is Anthropic doing this? The answer lies in Europe.

Article 50 of the European Union’s Artificial Intelligence Act (EU AI Act), which applies from August 2, 2026, provides that providers of AI systems that generate synthetic text, images, audio, or video must ensure that the content can be identified as artificially generated or manipulated through machine-readable labeling.

The aim is to reduce the risks of misinformation, manipulation, fraud, impersonation, and consumer deception.

As the European Commission explains in its relevant announcement, Europe wants to make it somewhat harder for content that appears human-made to circulate without there being a way to determine that it was created by a machine.

This is particularly important at a time when the Internet is being flooded with what has become known as “AI slop”: vast amounts of cheap, hastily produced, and often low-quality content generated en masse by artificial intelligence tools.

The problem for Europe is not only that the amount of “junk” on the Internet is increasing. It is that it is becoming increasingly difficult for users to know what they are looking at, who created it, and whether they can trust it.

Who is affected by the new strategy

The change affects two very different categories of users equally. On the one hand, there are students, writers, and academics, for whom indiscriminate use of AI may raise issues of plagiarism, originality, or research ethics.

On the other hand, there are professionals such as lawyers, journalists, and marketing and communications executives, for whom AI has now become another everyday work tool — from research and drafting to translation and text editing.

The new labeling addresses both cases in practice through the same digital trace: a text produced with the assistance of AI may carry the same indication, regardless of whether it was used to circumvent an obligation or simply to complete professional work more quickly.

The question is whether this could lead to a new form of suspicion toward those who use AI legally and creatively as a work tool.

Will everyone do it? The change will not be limited to Claude.

The obligations under the AI Act apply to providers of the relevant systems in the European Union. Approximately 190 organizations have signed the European Code of Practice on transparency of AI-generated content, including Anthropic, Google, Meta, Microsoft, Mistral, and OpenAI.

Companies will have to ensure that the content they generate can be identified as the product of artificial intelligence. Google, for example, has already developed its own SynthID system and says it is working on the adoption of transparency and interoperable watermarking tools in cooperation with other companies.

At Anthropic, the watermark already exists, but it cannot yet be checked by everyone: Anthropic has not publicly released the detection tool for this particular watermark. This means that today, a third party cannot simply take a piece of Claude-generated text and independently check it to prove that it carries the specific watermark. Anthropic says the watermark can survive copying, pasting, and light editing. However, extensive rewriting, paraphrasing, or processing by another artificial intelligence model can weaken or even eliminate the statistical trace.

Tools are already appearing on the market that promise to make AI-generated text more “human” or to bypass detection systems.

The Great European Experiment

Europe is trying to solve a real problem: how to maintain trust in digital content when AI can generate virtually unlimited quantities of text, images, and video at minimal cost.

At the same time, however, Europe wants artificial intelligence to increase the productivity of its businesses and workers. The question, therefore, is whether the new rules will function as a necessary “trust signal” or whether they will add yet another layer of compliance to a technology that European businesses are trying to adopt as quickly as possible.

Anthropic maintains that its watermark does not increase costs or degrade the quality of responses. There is, however, already discussion about whether intervening in the way the model selects words could have even a slight impact on quality.

In any case, Europe is doing something that has not yet been done globally on a comparable scale: turning the detectability of AI into a regulatory requirement rather than an optional feature.