Anthropic has announced it will start embedding invisible watermarks in the text generated by its Claude large language model and chatbot. This move aligns with the new European Union rules aimed at enhancing AI transparency. The San Francisco-based AI powerhouse made this statement clear, highlighting that the new watermarking policy will encompass a wide range of services provided by Claude. This includes everything from its consumer app to the Claude Platform API for developers, Claude Code, Claude Cowork, Claude Tag, and even versions accessed through major cloud providers like AWS, Google Cloud, and Microsoft Foundry.
Now, here’s the kicker: the watermark isn’t just a local affair. Anthropic clarified that this marking will apply to every region where Claude is offered, not merely the European Union. This means that whether you’re in Paris, New York, or Tokyo, Claude’s outputs will carry this watermark. The watermark itself will be integrated either directly into the text or as metadata, but don’t get too comfortable with it. The company warned that this marking can be easily stripped through various processes like format conversion or screenshots. So, a document that had a Claude mark might lose it entirely after being edited or converted… tricky, right?
This initiative comes on the heels of the EU’s AI Act, which mandates that AI systems generating synthetic content, like deepfakes, must clearly mark their outputs as artificially created under Article 50. Regulators anticipate that providers, like Anthropic, will adopt a multi-layered marking strategy. This strategy is expected to combine digitally signed metadata, imperceptible watermarking, and sometimes even fingerprinting or logging as a backup, because, let’s face it, no single marking technique is foolproof.
Starting from 2 August 2026, the transparency obligations under Article 50 will kick in, and non-compliance could lead to hefty fines—up to €15 million or 3% of a company’s total global annual turnover, whichever is higher. A significant penalty, for sure! Anthropic has made it clear that a detected watermark is merely a signal, not definitive proof. It only indicates that the text may have passed through Claude, leaving the full history of its creation a bit murky. Many users rely on the assistant for proofreading, translating, or summarizing content that originated elsewhere, so a watermark might show up on text that Claude did not initially produce.
And it doesn’t stop there. The reverse scenario is also a possibility; heavy editing, translations, or using older models that lack marking support could mean that genuine Claude outputs go undetected. Anthropic isn’t navigating these waters alone, either. Google DeepMind has rolled out its own marking tool, SynthID, for AI-generated text, images, and audio, while OpenAI has been discussing watermarking strategies but has taken a slower approach to broader deployment.
So, what’s next for these AI giants as they dive into the complexities of compliance and transparency? We’ll be keeping a close eye on how this all plays out in the coming months…
Kaynak: Orijinal Haber