What happened
Anthropic turned on invisible cryptographic watermarking across the full Claude lineup - Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 - on Monday, applying it to every text output regardless of region or deployment surface, CryptoBriefing reported. The watermark is statistical, not visible, and is embedded during token sampling. It survives light paraphrasing and copy-paste but degrades under aggressive rewriting, consistent with the approach Google DeepMind published under the SynthID-Text banner in a 2024 Nature paper. Anthropic had previously offered watermarking as an opt-in for enterprise Bedrock and Vertex customers. Monday's change makes it default and non-optional. The company did not publish a detection API alongside the rollout, meaning third-party verification is not yet available outside a small set of partner integrations. Detection endpoints are expected before the EU deadline, per the report.
The timing is not coincidental. Article 50 of the EU AI Act requires providers of generative systems to mark synthetic content in a machine-readable way, with enforcement kicking in on September 2, 2026. Providers who miss the window face fines scaled to global turnover. Anthropic's global rollout - rather than an EU-only carve-out - signals a bet that a single compliance surface is cheaper to run than region-gated logic.
Why it matters
Provenance is the piece the generative-AI stack has been missing since 2023. Every serious policy conversation - from the EU AI Act to the US executive order rescinded in early 2025 to the UK's AISI voluntary framework - has circled the same question: can a downstream reader tell whether a piece of text came from a machine. Anthropic's answer, delivered globally and by default, is the strongest yes any frontier lab has shipped.