‘New models will mark AI-generated content from day one’: Claude will now hide an invisible watermark inside ordinary words — here’s how that’s even possible, and how EU rules could push OpenAI and Google to follow suit

As of today, Anthropic, the creator of Claude, has announced a groundbreaking step in the realm of artificial intelligence. Starting from day one, Claude will now embed an "imperceptible watermark" into its text generation, ensuring that any content it creates can be traced back to its AI origins. This move is unprecedented for Claude, as it extends beyond the watermarking of images, which are relatively simpler to embed, to encompass all text Claude produces.

The reason behind this decision is the EU AI Act, specifically Article 50(2), which mandates transparency in AI-generated content. This requirement forces providers like Anthropic to make their outputs machine-readable and detectable as AI-generated, a necessity that will take effect on August 2, 2026, with a transition period until December 2, 2026.

While Anthropic has not disclosed the exact algorithm used for this watermarking, my educated guess is that it involves statistical watermarking during token generation. Rather than using hidden Unicode characters or metadata, Claude's watermarking process subtly nudges the selection of preferred tokens more often than chance would allow.

This statistical fingerprint, distributed across hundreds of word choices, would survive minor edits and copying/pasting. However, the effectiveness and ease of circumventing this watermark remain speculative. The potential implications are significant. If Claude's watermarking becomes easily bypassed, or if other major AI players decide to follow suit, Claude might experience a significant loss of customers.

Moreover, the use of Claude for proofreading human-written text could lead to unintended consequences, with texts flagged as AI-generated even if they were not. While the EU AI Act sets a new precedent for transparency in AI-generated content, the repercussions of this move by Anthropic remain to be seen, particularly in the context of maintaining user trust and the ethical use of AI.