← The AI Hype Audit — all 181 verdicts
BS “Get rid of Claude's new invisible watermarks” — the watermark is real; the remover re-inks itself
The claimHow to get rid of Claude's new invisible watermarks so the things you make aren't AI-detectable — “if you want to use Claude properly and not get watermarked, use this Claude skill to get rid of it.”
The premise is real news, and that's what makes this one dangerous: as of early August 2026, Anthropic does embed an imperceptible watermark in Claude-generated text at the model level — across claude.ai, the API, and Claude Code — plus digitally signed C2PA provenance metadata on supported files. Anthropic's own help center confirms it, with no opt-out; Fortune, Forbes and Gizmodo describe a SynthID-Text-style statistical bias in word choices under a key Anthropic holds, driven largely by EU AI Act Article 50. Now the twist that kills the product: a Claude skill runs INSIDE Claude. The watermark is applied at generation time, inside the model — so any “de-watermarking” rewrite the skill produces is itself freshly watermarked on the way out. You cannot use Claude to wash Claude's watermark out of Claude's output. What actually degrades the signal, per Anthropic's own limitations note, is heavy HUMAN editing, paraphrasing, translation, or mixing with other writing — none of which a skill provides. At best it could strip C2PA metadata from files, which any exif tool does and which tampering the C2PA signature is built to reveal. And Anthropic notes the inverse too: absence of a mark never proves content is human. “Not AI detectable” is a mirage from both directions.
What holds up
- The premise is true: Anthropic confirms an imperceptible model-level watermark in Claude-generated text, default-on across all Claude products, no opt-out — plus signed C2PA metadata on supported files (support.claude.com article 16266773)
- The mechanism is real and covered by Fortune (Aug 11), Forbes (Aug 13) and Gizmodo: a SynthID-Text-style statistical bias keyed by Anthropic
- Applies to models shipped after Aug 2, 2026 — this genuinely is new
What doesn't
- A Claude skill runs inside Claude — its “cleaned” rewrite is re-watermarked by the same mechanism the moment it's generated; the product defeats itself by construction
- The key lives with Anthropic, not in your context window — no prompt, skill, or instruction can switch the watermark off
- Per Anthropic's own limitations, only heavy human editing, paraphrasing, translation, or mixing degrades the signal — exactly what a skill doesn't do
- Stripping C2PA metadata is trivial with any exif tool and is precisely the tampering the signature is designed to expose
- Anthropic also notes absence of a mark never proves human authorship — so “not AI detectable” fails in both directions
- A skill sold explicitly to defeat AI-content disclosure is a usage-policy problem dressed as a productivity hack
The catch
The skill asks the printer to erase its own ink — every rewrite Claude produces leaves the factory freshly watermarked, so the only thing being removed is money from the buyer.
How to actually do it
- Accept the fact half of this video: Claude output is watermarked, default-on, no opt-out — plan for that instead of around it
- If the work matters, edit it until it's yours — heavy human revision is the one thing Anthropic says genuinely fades the signal, and it also makes the writing better
- Disclose AI assistance where your context calls for it; the regulation driving this (EU AI Act Art. 50) is moving toward disclosure being assumed anyway
- Never buy a “bypass” product whose mechanism runs inside the system it claims to bypass — that one test kills this whole genre
The watermark story is true; the $-skill built to remove it re-watermarks its own output — snake oil with a factual label.
- Confidence
- High
- Posted by
- an app-builder creator selling a “watermark removal” Claude skill
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.