.@AnthropicAI’s new AI watermarking doesn’t really make sense. Watermarking actually gets less useful as AI gets more useful. Claude will soon add machine-readable marks to generated text and signed metadata to supported files to comply with the EU AI Act. But this means Claude will also mark work it only edited, translated, summarized or reformatted. So something written entirely by Claude and something written entirely by a human but proofread by Claude will carry the same AI signal. That might work when AI is a separate tool you occasionally use, but it breaks down once AI is embedded in how most writing, coding and creative work gets done. At that point, knowing AI was involved tells you very little. What people really want to know is how much of the work it actually did. And that's a problem we still don't really have a solution for.(Sahara AI 🔆)