Anthropic will add imperceptible watermarks to text produced by Claude models launched in the European Union on or after August 2, 2026. The move follows the start of the EU AI Act, the world’s first comprehensive law on artificial intelligence.
Anthropic confirmed the watermarking plans on August 11. The company says the watermark will not change the meaning, quality, or readability of Claude’s output. Instead, it will sit inside the generated text as part of the output itself.
That design gives the watermark a useful feature: it can stay with the text after someone copies and pastes it elsewhere. Anthropic said, “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.”
What the EU AI Act requires
The EU AI Act took effect on August 2, 2026, bringing new transparency duties for AI providers, deployers, and some public agencies. AI systems must disclose their AI nature unless the surrounding context makes that obvious.
Providers must also design their systems so content generated or manipulated by AI carries machine-readable marks that allow detection. Anthropic’s Claude watermark fits this requirement for text generated in the EU under the company’s announced plan.
The rules cover more than written output. Deployers must label deepfake videos, images, and audio files, along with AI-generated or manipulated text that deals with matters of public interest. The European Commission has introduced icons for this content, though the icons are optional and can be used instead of labels created by platforms.
TikTok and Instagram have begun marking AI-generated and manipulated content with their own icons and disclaimers. Meta and xAI also label AI content, adding to a growing collection of disclosure methods across online platforms.
People must also be informed when they interact with emotion recognition or biometric categorization tools. The rule is meant to make the presence of AI clear before users rely on what the system says or does.
Fines, delays, and the protection gap
Companies that fail to follow the EU AI rules can face fines of up to 15 million euros, or 3 percent of global revenue. That top fine equals $24.2 million Canadian. EU institutions and agencies can face fines of up to 750,000 euros, or $1.2 million.
The law does not apply every major obligation at once. The most far-reaching high-risk duties for systems used in areas such as biometrics and border management have been postponed until December 2, 2027.
The delay also covers AI systems used in migration, asylum, and border management, all of which are classified as high risk. Critics say the decision leaves vulnerable populations without full protections for an additional 16 months.
Teresa Scassa, a professor at the University of Ottawa, said policymakers are trying to balance transparency with the risk that users will stop noticing warnings. “It’s clear that they’ve been working to try and balance different interests — for example, not to have to create disclosure fatigue by having too many disclosures so that people stop paying attention to them.”
That balance will shape how useful watermarks and labels become. A hidden machine-readable mark can help investigators and platforms detect content, but it does not give ordinary readers an obvious warning on its own. Visible icons and disclaimers address that gap, while optional standards may lead to different labels across services.
Canada watches Europe’s approach
The EU rules are already influencing AI standards beyond Europe, since companies often apply disclosure requirements worldwide. Canada launched a public consultation on AI transparency rules on August 6, 2026, with submissions due by September 23.
The consultation focuses on detecting AI-generated content, informing users, helping people understand AI systems, tracking incidents, and monitoring AI behavior. Canada’s AI strategy, launched in June 2026, includes watermarks for AI-generated content.
The Office of AI Minister Evan Solomon said, “Canada is closely examining the European Union’s AI transparency requirements that took effect this week, alongside work underway through the G7, OECD and other international partners.” Ottawa says Canada’s approach will reflect Canadian needs and values while drawing on international partners.
Brian McQuinn, an associate professor at the University of Regina, described the stakes in direct terms: “We are in an arms race around disinformation and misinformation with those actors who are using generative AI to misinform and attack the institutions of societies.”
The debate reaches beyond Canada and the EU. Washington is keeping its AI rulebook private, a decision that has caused dissatisfaction among smaller AI labs. Fortune reported on that situation as companies and policymakers weigh how much information AI developers should disclose.
Europe’s policy push is also part of a wider technology effort. Google opened a new AI centre in Berlin on March 5, 2026, adding to the region’s role in the global AI race.
Anthropic’s Claude watermark offers an early test of the EU’s approach: make AI content traceable without changing what users read. The result will depend on whether hidden marks survive editing, whether platforms recognize them, and whether visible labels remain clear enough for people to notice.
Based on
- Anthropic Watermarks Claude Text Output to Meet EU Transparency Rules — unite.ai
- Anthropic Rolled Out a Fix to Try to Stop Undetected AI-Generated Writing – Business Insider — businessinsider.com
- Europe’s new AI transparency rules are in effect. Should Canada copy them? – National | Globalnews.ca — globalnews.ca
- What came into force with the EU’s AI Act this week – and what didn’t | Technology News | Al Jazeera — aljazeera.com
- Washington is keeping its AI rulebook private. Smaller AI labs aren’t happy. | Fortune — fortune.com
