Anthropic’s Claude Gets Invisible Watermarks: EU AI Act’s Ripple Across Tech Markets

Anthropic has updated Claude models to embed an invisible watermark in every piece of text generated from August 2, a move that globally enforces compliance with Article 50 of the EU AI Act.

How the Invisible Watermark Works

The text watermark subtly biases the model’s word choices using a statistical key; individual selections appear normal, but over sufficiently long passages a detectable pattern emerges and persists when the text is copied.

Signed Provenance Metadata for Digital Files

For supported formats such as .svg, .png, and .jpg, Claude attaches C2PA (Coalition for Content Provenance and Authenticity) compliant signed metadata. This label signals that the file was processed by Claude and enables detection of any subsequent tampering.

Penalties Under the EU AI Act

Article 50 mandates machine‑readable tagging of AI‑generated outputs, with fines up to €15 million or 3 % of a company’s worldwide annual turnover—whichever is higher—for non‑compliance.

Industry Reactions and Competitive Landscape

■200 companies have signed the EU Code of Practice on Transparency, indicating pre‑emptive compliance.

■Google has been watermarking images since 2023 and is now extending the practice to text, audio, and video.

■OpenAI possesses watermarking capability for ChatGPT but has withheld deployment due to false‑positive concerns and competitive risk.

■xAI remains outside the EU code, potentially positioning itself as a differentiated alternative.

■%3 of corporate revenue at risk of fines.

■200 firms signed the EU Transparency Code.

■2026 marks the global rollout of the watermark.

■2 months until detection tools become publicly available.

Deniz Arslan – While the move enhances transparency of AI‑generated content, it also adds a new compliance cost layer for startups and large tech firms alike. Watermarks that linger on code or edited content could create a perception of “untouchability,” prompting investors to reassess risk profiles; early adopters may mitigate regulatory exposure, whereas laggards risk both reputational and financial penalties.