started · updated
Anthropic implements invisible watermarking for Claude AI models
Anthropic is implementing an invisible text watermarking system for its Claude AI models to comply with the European Union AI Act. The technology, based on Google DeepMind’s SynthID-Text method, works by subtly influencing the statistical probability of word choices during text generation. This creates a pattern detectable by specialized software but imperceptible to human readers.
Anthropic states that the watermarking process does not add hidden characters, increase computational costs, or degrade the quality, readability, or creativity of the output. The system is designed to indicate whether Claude processed the text, though it does not necessarily mean the AI wrote the entire content. The policy will apply to models released after August 2, 2026, and the company plans to release a detection API for authorized third parties.
The announcement has faced significant backlash from users, including developers and professionals. Some have reported canceling their subscriptions, expressing concerns that the watermark could lead to false accusations of AI usage even when Claude is only used for minor tasks like proofreading, translation, or summarization. Critics also argue that the watermark could complicate copyright and ownership claims for professional deliverables, such as code.
Entities
Anthropic · C2PA · Charles Hoskinson · Claude · Dario Amodei · EU AI Act · European Union · European Union · Google DeepMind · SynthID-Text
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 12 SOURCES] The system can only estimate if Claude was involved in the text and cannot distinguish between original generation and heavy editing. mashdigi.com · www.iclarified.com · slashdot.org · gigazine.net · www.ginjfo.com · +6 more
- [● 16 SOURCES] The technology is based on a version of Google DeepMind’s SynthID-Text method. mashdigi.com · www.iclarified.com · www.teknoblog.com · gigazine.net · it-online.co.za · +10 more
- [● 5 SOURCES] Anthropic will use C2PA-compliant metadata to watermark AI-generated images. mashdigi.com · www.patrickkphillips.com · www.newsbytesapp.com · dxmagazine.jp
- [● 16 SOURCES] The watermarking process does not add hidden characters, increase computational costs, or affect text quality. mashdigi.com · www.iclarified.com · gigazine.net · it-online.co.za · multiplayer.it · +11 more
- [● 20 SOURCES] Anthropic is implementing text watermarking for Claude models to comply with the EU AI Act. mashdigi.com · www.iclarified.com · rayhaber.com · slashdot.org · www.teknoblog.com · +14 more
- [● 17 SOURCES] The watermarking mechanism works by subtly influencing the statistical probability of word choices during text generation. mashdigi.com · www.iclarified.com · rayhaber.com · slashdot.org · gigazine.net · +12 more
- [● 7 SOURCES] Anthropic plans to release a detection API to allow authorized third parties to identify watermarked text. mashdigi.com · www.iclarified.com · gigazine.net · www.infomagazine.ma · es.digitaltrends.com · +2 more
- [● 7 SOURCES] The watermarking policy will apply by default to all Claude models released after August 2, 2026. mashdigi.com · gigazine.net · www.iclarified.com · www.patrickkphillips.com · www.infomagazine.ma · +1 more
- [● 4 SOURCES] The watermark can be detected even if the text is copied and pasted or slightly edited. www.businessinsider.jp · dxmagazine.jp · www.bright.nl · it-online.co.za
- [● 5 SOURCES] The watermark indicates Claude processed text, not necessarily that it wrote it. www.businessinsider.jp · www.ginjfo.com · www.diarioestrategia.cl · dxmagazine.jp · slashdot.org