AI Watermarks: "Processed" Is Not "Written"

Since 11 August, everything Claude produces carries an invisible mark. It attests that a model processed the text — never that it wrote it. Sophie and Professor Omni unpack a distinction almost nobody will catch, and that is about to fuel a lot of false accusations.

In this episode

• Why proofreading your own text is enough to get it marked — and why the verdict is binary, indifferent to proportion
• The right image: it's not a stamp, it's a regional accent. What survives, what erases it
• The case of code: the entropy dependency is documented, but it's an expectation, not a measurement — plus the protocol to falsify it
• The real risk, neither legal nor technical: perception, the day the detection interface goes public

References

• Full article: AI Watermarks: "Processed" Is Not "Written"
• Concepts: Article 50 of the EU AI regulation, signed provenance metadata, entropy and watermark robustness

24h Action

Take the last public text you signed. Ask yourself whether you could say out loud, without discomfort, how it was produced. If the answer is no, the mark isn't your problem.