Anthropic has explained how it embeds invisible watermarks in Claude-generated text to comply with EU AI Act transparency rules.
The watermark works by replacing the random process that selects words with one governed by a secret key and context, leaving detectable statistical patterns without adding visible text or affecting performance.
The company is applying the watermark globally, not just in the EU, and will offer an API for detection, though the watermark can be erased by complete rewriting and does not work well on formulas, strict code, or single-answer facts.
What happened
Anthropic revealed how it embeds imperceptible electronic watermarks in text generated by Claude. The method, based on Google DeepMind's SynthID-Text, replaces the random number source used to select words with a secret key and preceding context, leaving a statistical pattern. The watermark adds no characters, so it does not affect text quality, speed, or cost.
Why it matters
Anthropic signed the EU AI Act's transparency code of conduct, requiring disclosure that AI-generated text contains watermarks. Although the regulation applies only in the EU, the company is rolling out the watermark globally across all models and products, including in Japan, because it lacks a reliable way to limit implementation by region. Other major AI developers are expected to implement their own watermarks under the same obligation.
What to watch
Detection requires an API Anthropic will provide soon, using a secret key to verify word sequences and measure Claude's involvement. The watermark works weakly in low-choice contexts (formulas, code, facts with one correct answer) and is defeated by complete rewriting, but survives light edits. Anthropic will deploy the feature to existing models over the coming months before the regulation's August 2, 2026 compliance deadline.
Ask the AI about this article →
Anthropic's decision to apply the watermark globally, rather than limit it to the EU, reflects a practical compliance challenge: the company lacks a reliable technical method to restrict the feature by region. Under the EU AI Act's transparency code of conduct, which Anthropic and roughly 190 other organizations signed, developers must disclose when AI-generated text contains watermarks. Rather than engineer separate systems for different markets, Anthropic chose to deploy the feature worldwide, including Japan, making the watermark a global standard for its Claude product.
The watermark's design rests on a fundamental constraint of language generation: when an AI model chooses the next word, most word choices do not meaningfully change the meaning or readability of the text. By embedding a secret pattern in those low-risk selections—without adding extra text or characters—Anthropic avoids the common criticism that watermarking slows AI or degrades output quality. The transparency comes at minimal cost. However, the watermark has clear limits: it cannot survive a complete rewrite, and it fails to embed itself in contexts where only one word makes sense, such as finishing a historical fact or completing a mathematical formula. For strict code, Anthropic limits watermarking to comments, where alternative wording is possible.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Analyst Ming-Chi Kuo says Nvidia has revived the Rubin CPX AI accelerator with a substantially redesigned arch…
A UK study by UK AI Security Institute and Limbic AI surveyed 6,474 British adults
Broadcom's Clayton Donley says companies are doing mission-critical work with AI agents quickly, but without t…
Bank of England governor Andrew Bailey warned that advanced AI poses risks to financial infrastructure in a le…
OpenAI released a new evaluation framework on July 17, 2026, urging companies to measure AI ROI by 'useful out…
As AI agents perform real business tasks, 'Agentic Identity' (giving each AI a unique employee-like ID) and 'D…
