Claude's Invisible Watermark: What It Can't Prove

Anthropic now embeds an invisible watermark in the text Claude generates. The news covered what happened. This covers how it works, and what a detected mark actually proves, which is a lot less than most people are assuming. The short version: most people assume it works by hiding characters between the words. It is in the word choices themselves. That one idea explains every property Anthropic describes, including why it survives a copy-paste and why a paraphrase destroys it. One flag, made in the video too: Anthropic has not published its implementation. The mechanism section is the established technique this belongs to and it fits every stated property, but it is inference, and it is labelled as inference on screen. And the case worth knowing about: write a paragraph yourself, paste it into Claude to fix the grammar, and what comes back carries the mark. Anthropic names proofreading and translation directly. A detected mark means the text may have been processed by Claude. It does not say who wrote it. CHAPTERS 0:00 The mark you cannot see 0:21 Your own paragraph comes back marked 0:36 What Anthropic actually shipped 1:12 Why it applies worldwide, not just the EU 1:27 The guess almost everyone gets wrong (and what breaks it) 2:08 A flag: this next part is inference 2:19 How word choice carries a signature 2:58 That is the whole trick 3:08 Where detection breaks down 3:44 What a detected mark does not prove 4:26 It fails in both directions 4:38 Nobody can detect it yet 5:05 The honest part: for most work, nothing changes 5:15 The one-liner WHAT YOU'LL BE ABLE TO SAY AFTERWARDS - What is marked: text gets an imperceptible watermark, files get signed C2PA provenance metadata - Why the hidden-characters theory is wrong, and what breaks it - How a keyed split of the vocabulary turns ordinary word choices into a measurable signature - Why the mark needs length before it can say anything at all - Why paraphrase and translation destroy it, and heavy editing wears it down - The two things a detected mark cannot establish, in Anthropic's own words - Why an unmarked passage proves nothing either A NOTE ON SOURCES No detection rates, accuracy percentages, or token thresholds appear in this video, because Anthropic has published none and inventing them would be the whole problem in miniature. Every factual claim traces to Anthropic's own help article. Where the video reasons past the source, it says so on screen. LINKS How Claude marks AI-generated content (Anthropic): https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content Anthropic's transparency hub: https://www.anthropic.com/transparency/voluntary-commitments #claude #anthropic #aiwatermarking #aidetection #euaiact #llm #ai #softwareengineering