Anthropic just confirmed it’s using a version of Google DeepMind’s SynthID-Text to invisibly watermark everything Claude generates. The move isn’t optional — it’s compliance with the European Union’s AI Act, which requires machine-readable markers on synthetic text, images, audio, and video so platforms and regulators can identify AI-produced content. Anthropic is pairing the text watermarking with C2PA support for Claude-processed images, covering both output formats the model handles.
The underlying technique is clever in its subtlety. When Claude picks the next word in a sentence, many choices are essentially interchangeable — ‘overcast’ versus ‘grey’ after ‘The weather today was cold and…’ doesn’t change meaning. Watermarking hijacks those low-stakes decisions, using a cryptographic key plus preceding words to steer the random selection process. The result is a statistical pattern invisible to human readers but detectable by anyone holding the key. Anthropic insists the system won’t raise costs or degrade output quality, which makes sense given the approach only nudges choices when multiple options are roughly equivalent.
Claude won’t be alone here. Google’s Gemini has supported SynthID-Text since 2024, giving the search giant a head start on EU compliance. OpenAI, meanwhile, hasn’t detailed any text watermarking plans in its AI Act roadmap, though the law applies to them regardless. That gap is worth watching — OpenAI has historically been cagey about provenance tools, citing accuracy concerns with earlier detection methods. If Anthropic and Google can demonstrate watermarking without quality loss, pressure builds on OpenAI to follow suit or explain why it can’t.
The broader implication is that AI transparency is shifting from voluntary pledges to legal obligation. Watermarking won’t stop determined bad actors from stripping markers or using open-source models without them, but it establishes a baseline for the major commercial platforms. The real test comes after deployment: whether detection tools built on these watermarks actually hold up at scale, and whether EU regulators find the approach sufficient or demand more aggressive measures.
