Anthropic’s Claude watermark may exceed EU requirements, experts say

Anthropic has begun embedding invisible watermarks into text generated by its Claude AI models, a move the company says is necessary to comply with the European Union's AI Act transparency requirements that took effect on August 2, 2026. The watermark applies globally and across all Claude products — including the API, Claude.ai, Claude Code, and enterprise deployments — because the company says it has no way to apply the feature on a per-region basis.techcrunch+2

How the Watermark Works

Rather than inserting hidden characters or visible disclaimers, the system subtly guides Claude's word choices to create a statistical pattern detectable by specialized tools. Anthropic stated on its blog that the watermark "does not have any practical impact on the quality or content of Claude's outputs" and that "the difference between watermarked and un-watermarked text will not be distinguishable to readers". Because the signal lives within the text itself, it can persist when content is copied, pasted, or lightly edited.thenextweb+2

A detected watermark indicates only that Claude "may have processed the content," the company said — not that the AI authored it from scratch. Users who submit their own writing for proofreading or translation may still see a watermark attached to the output, though Anthropic told Business Insider News Corp ↗ 0.59% that lighter editing is "less likely to trigger the watermark" because "there's very little (if anything) for the watermark to attach to".anthropic+2

Broader Than EU Law Requires?

Experts told Business Insider that Anthropic's approach may exceed what the EU mandates. Article 50(2) of the AI Act exempts content where AI is used for "standard editing" that does not substantially alter the original text. Yet Anthropic's watermark can survive translation, summarization, and more extensive editing. Caroline De Cock, senior fellow at the Lisbon Council think tank, said Anthropic "is using a sledgehammer that hits simple proofreading and translation" where the law "called for a scalpel to catch deepfakes and mass synthetic generation".businessinsider

Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 200 other signatories including Meta ↗ 5.95%, Microsoft ↗ 3.25%, and OpenAI. Google Alphabet Inc. ↗ 1.28% already uses its SynthID technology to watermark AI outputs worldwide, while OpenAI has said it plans to expand labeling to text but has not specified a date.forbes+2

User Backlash and Industry Implications

The announcement triggered cancellations from some Claude subscribers and prompted developers to create tools to strip watermarks. A recurring concern is that the watermark could be interpreted as proof that Claude wrote a document, even when a human used it only for minor edits — raising questions about intellectual property ownership and workplace policies around AI use.forbes+1

Milos Rusic, CEO of enterprise AI company Deepset, cautioned against treating the watermark as definitive. "These systems can be useful evidence," he told Business Insider, "but I would be very cautious about turning them into an automated judgment about authorship or misconduct".businessinsider