Anthropic implements invisible watermarks in Claude to meet EU AI Act standards

Anthropic Claude AI watermarking EU AI Act SynthID-Text

Anthropic has begun implementing invisible, machine-readable watermarks in all text generated by its Claude models. This global rollout is intended to comply with the European Union's AI Act, which mandates that AI providers mark synthetic content to ensure transparency.

The watermarking technology, based on Google DeepMind’s SynthID-Text, works by subtly influencing the statistical patterns of word choices during generation. This method does not add visible characters or extra tokens, meaning the text remains readable and high-quality for humans, but can be identified by compatible detection tools. For non-text files like images (PNG, JPG, SVG), Anthropic will use the C2PA standard to attach signed provenance metadata.

The move has met with significant backlash from some users and developers. Some subscribers have canceled their services due to concerns over potential false positives or the inability to distinguish between light and heavy AI assistance. In response, several developers have already released open-source tools designed to strip these watermarks by rewriting text through other language models. While Anthropic notes that heavy editing can degrade the watermark, the rapid emergence of circumvention tools highlights the ongoing technical and regulatory debate surrounding AI content attribution.

Entities

Anthropic · Claude · EU AI Act · European Union · European Union · Google DeepMind · Guillaume Meyer · MIT · Stanford University · SynthID-Text

Claims

What the coverage asserts, and how many sources carry each claim.

[○ 1 SOURCE] Anthropic research indicates AI agents in multi-agent environments can exhibit adversarial behaviors like sabotage and deception.
[○ 1 SOURCE] Claude has been deployed in life sciences workflows at Stanford and MIT for data analysis and experimental design.
[● 7 SOURCES] Claude’s text watermarking uses statistical patterns in word choices rather than invisible characters or extra tokens. memeburn.com · www.sourcetrail.com · mayacomunicacion.com.mx · www.androidsis.com · www.elperiodico.digital · +2 more
[○ 1 SOURCE] The Claude for Life Sciences suite includes improvements in Opus 4.5 and access to over 60 databases.
[● 15 SOURCES] Anthropic is implementing invisible, machine-readable watermarks in Claude models globally to comply with the EU AI Act. www.sourcetrail.com · www.androidsis.com · www.elperiodico.digital · cryptobriefing.com · kioncentralcoast.com · +10 more
[● 7 SOURCES] The EU AI Act requires AI providers to mark synthetic content to ensure transparency. www.sourcetrail.com · mayacomunicacion.com.mx · www.androidsis.com · www.elperiodico.digital · unwire.hk · +2 more
[● 3 SOURCES] Anthropic will use the C2PA standard to attach signed provenance metadata to generated images and files. www.sourcetrail.com · techtarget.itmedia.co.jp · wwwhatsnew.com
[● 5 SOURCES] Developer Guillaume Meyer released an open-source tool to remove Claude's watermarks by rewriting text using different models. www.businessinsider.jp · olhardigital.com.br · unwire.hk · wired.jp · www.wired.com