Anthropic just started embedding invisible watermarks into everything Claude writes. Including your code.
Starting with every model released after August 2, Claude weaves a machine-readable signal into its output that survives copy-paste and applies across every surface: Claude, Claude Code, Claude Cowork, the API. No opt-out. The trigger is the EU AI Act's Transparency Code, and every major lab is doing some version of this by year's end.
The backlash came from an unexpected group: not students, but lawyers and academics who use Claude to copy-edit their own writing, now finding their human drafts flagged because Claude touched three sentences of them.
For engineering teams, three things actually matter here:
If Claude Code touched a function, the watermark travels into your repo, and possibly a client's, indefinitely. Any "no AI-generated code" clause just got a lot more testable.
The watermark proves Claude touched the text. It doesn't prove you followed a real process, or that the output matches what was approved.
Detecting AI involvement answers one question. Regulators and clients want three: who asked for it, under what constraints, and whether it matches what was approved. A spec-driven paper trail answers all three. A watermark answers one, and only sometimes.
Top comments (0)