Anthropic embeds invisible watermarks in Claude text outputs

The detail worth sitting with is that token-level watermarking works by subtly biasing the probability distribution over token choices during generation, meaning the signal is baked into the statistical texture of the output rather than appended afterward. That distinction matters because it means stripping the watermark requires degrading the text itself, which is a meaningfully stronger guarantee than metadata-based approaches.

This story belongs to a cluster of questions about AI provenance that the industry is being forced to answer as deployment scales. The Apple-Gemini integration covered here on the same day (The Decoder, September 15) is a useful contrast: Apple is outsourcing core assistant intelligence to Google partly because building and maintaining frontier models in-house is expensive, but that arrangement makes output attribution harder, not easier. If Siri responses are generated by Gemini, whose watermark applies? Anthropic's move implicitly pressures every lab that supplies models to consumer products to answer that question with something more durable than a terms-of-service clause.

Watch whether Google or OpenAI announce compatible or competing watermarking specs within the next two quarters. If a shared detection standard emerges before any major regulatory deadline in the EU AI Act's provenance provisions, that suggests the labs are getting ahead of the mandate rather than waiting to be compelled.