Anthropic will add machine-readable watermarks to all Claude outputs to comply with the EU AI Act. Goal: help platforms detect AI text and fight misinformation.
Users are uneasy. If a student paraphrases Claude’s essay, will it still be flagged? If a journalist uses Claude to draft and then rewrites 80%, is it "AI content"? False positives could ruin reputations.
Analytically, watermarking works for images. For text, it’s messy. Text is infinitely rewritable. Metadata is stripped when copied to WhatsApp. Current detectors already have 10-15% false positive rates and have wrongly accused students.
The deeper issue: misinformation spreads because of incentives, not because of lack of labels. A viral WhatsApp forward is believed because it confirms bias, not because it lacks a watermark.
What we actually need:
1. AI literacy in schools: teach source verification.
2. Platform accountability: Meta, X, YouTube must label viral AI content, not just producers.
3. Legal clarity: Who is liable when AI content causes harm?
4. Open standards so watermarks don’t become a proprietary lock-in.
Anthropic’s move is well-meaning. But if done badly, it will create a market for "AI washers" that remove watermarks and penalize honest users. The line between tool and author is already blurry. Law must catch up before tech does.