The Watermark Rules — the 5-rule card for Claude's invisible watermark: what gets marked, what survives editing, and what to do before you hit send
The companion to the video. As of August 2, 2026, every new Claude model embeds an imperceptible watermark into the text it writes — worldwide, across the Claude app, the API, Claude Code, Cowork, and Tag, with no setting to turn it off — and generated PNG/JPG/SVG files carry C2PA signed content credentials in their metadata. This guide is the one-page system for living with it. INSIDE: (1) WHAT CHANGED — the EU AI Act Article 50 trigger, exactly which models and products mark content, and the transition status of older models. (2) HOW AN INVISIBLE TEXT WATERMARK ACTUALLY WORKS — it's not hidden characters (a real hexdump shows every byte clean); the signature lives in a statistical pattern woven through Claude's word choices, which is why copy-paste carries it and no cleaner tool can strip it. (3) THE CARD — all 5 rules on one screenshot-able page: Assume it's signed · Light edits keep the mark · Your voice erases it · A mark is not a verdict · Act like detection is already here — plus the full SURVIVES/WIPED table for both the text watermark and the file seal (copy-paste survives; deep rewrite, paraphrase, translation wipe it; a sentence or two is below the detection floor; screenshots strip file seals). (4) RULES IN DEPTH — including the two edges most coverage misses: the length floor and the mixing effect. (5) THE VERDICT PAGE — why a detected mark only means Claude MAY have touched the text (proofreading your own essay leaves the same trace), why no mark proves nothing, and why teachers, managers, and editors should never punish on a scan alone. (6) THE BEFORE-YOU-SEND CHECKLIST — the one question that runs the whole system ('Would a watermark hurt here?') plus three copy-paste disclosure lines for email, reports, and academic work. Every product claim verified against Anthropic's own documentation (support.claude.com, 'How Claude marks AI-generated content', August 13, 2026). Independent guide from Hyperautomation Labs — not affiliated with Anthropic.
Free. No spam. Unsubscribe anytime.