Anthropic announces invisible watermark for Claude texts – Critics warn of false markings

How the watermark works

Critics, however, warn of the opposite effect. Because the watermark operates at the level of individual words and phrases, it can also affect human texts that were subsequently revised by an AI only in parts. "Users who, for example, submit their own written text to the AI for editing will in future have to live with the risk that their texts will be completely branded as AI-generated – even if only individual phrases and words were polished across the entire text," the report warns.

According to this, authors, students, and journalists face an increased risk of having their works wrongly classified as AI-generated – with potential consequences for grading, copyright, and credibility. Editorial workflows in which AI is used only as an aid for proofreading could also be affected.

Reactions from providers and Brussels

Anthropic is one of a number of major providers that have responded to the EU rules in recent months. In addition to the US corporation, Microsoft and the French company Mistral have also announced or already introduced similar labels. Brussels is closely monitoring implementation and reserves the right to take further steps if the requirements are not met.

The move comes just one day after several broadcasts reported further news from around the world, including rescue operations following an earthquake in Colombia, an eruption of Mount Etna causing flight cancellations in Sicily, the defusing of aerial bombs in Neu-Ulm, a Bundesrat initiative by Lower Saxony regarding child media stars, weather reports from Bavaria, and further medals for German athletes.

The EU Commission emphasizes that the labeling requirement is part of a comprehensive package to regulate AI. The goal is to protect consumers from deception while not stifling innovation.

What the label means for users

Anthropic stated that it will roll out the watermark gradually and simultaneously work on methods to reduce misclassifications of human texts. Independent testing bodies are to monitor the detection accuracy in the coming months.

Observers see a fundamental conflict in the debate: on the one hand, AI-generated content should be unambiguously identifiable; on the other hand, the labeling must not become a tool that casts blanket suspicion on legitimate writing.