← The AI Hype Audit — all 272 verdicts

BS BS: 'Pass all AI detectors' with a Claude watermark remover; the repo is real and says itself nothing can promise that, and the demo runs the 'humanize' pass inside Claude, which re-marks it Facebook · Sep 2026

The claim30-second reel, ~34K views, sent Sun Sep 6, 2026, titled 'How to remove Claude watermarks in 30 seconds.' Caption: 'Pass all AI detectors with this. Comment "water" for the full guide.' The screen recording shows the GitHub repo guillaumemeyer/watermarks-remover and its remove-ai-marks skill running inside Claude: Layer A strips hidden Unicode and EXIF/C2PA metadata, Layer B does a 'humanize pass' that rewrites the wording. Claude's own on-screen output reads: 'Nobody can prove "undetectable," and I won't claim it.'

The repo is real and, ironically, honest. guillaumemeyer/watermarks-remover (MIT, created Aug 11, 2026, ~20.9K stars, 2.4K forks, v0.7.0, pushed Sep 5) does two things: Layer A deterministically strips invisible Unicode and C2PA/EXIF/XMP metadata from 20-plus file types, and Layer B rewrites prose to try to break statistical watermarks. Its own docs say Layer B 'cannot claim official undetectable,' that 'no tool can honestly certify this fails the official check,' and that 'rewording degrades the copy.' Its ethics file lists academic fraud and circumventing disclosure rules as misuse. The caption promise is the BS. Anthropic's text watermark (default-on, worldwide, every model launched after Aug 2, 2026; see #181) is a SynthID-style pattern in word choices, not a hidden character, so Layer A does nothing to it. Layer B is the only attempt, and the skill itself says to run the rewrite with a model that is NOT the suspected origin, because output from Claude gets watermarked again. The reel shows Claude doing the humanize pass. That is the self-defeating loop from #181: you launder the mark by asking the marker to re-mark it. Nobody can prove 'passes all detectors' anyway: Anthropic's detection API is still unreleased, so there is nothing official to test against, and the commercial detectors disagree with each other. GPTZero claims 93.5 percent recall on paraphrased text, Turnitin ships a dedicated 'bypasser' flag, Originality claims 96.7 percent on paraphrased text, and independent tests show some humanizers slip through while some human writing gets flagged. Unreliable in both directions means 'pass all' is not a product anyone can sell. The reel is a lead magnet: 'water' feeds a DM into a guides site and a $59 a month Skool community. The honest route: strip your own EXIF if you want, use AI as a drafter, edit it for real, and disclose where disclosure is required.

What holds up

  • GitHub API, api.github.com/repos/guillaumemeyer/watermarks-remover (Sep 6, 2026): created Aug 11, 2026; MIT license; 20,935 stars; 2,412 forks; 17 open issues; last push Sep 5, 2026. Public, active, matches the URL in the reel.
  • Repo README + skills/remove-ai-marks/SKILL.md (Sep 6, 2026): Layer A = 'deterministic' Unicode/metadata removal; Layer B = 'best-effort' statistical rewrite; 'Until vendors ship public detectors and keys, no tool can honestly certify this fails the official check'; 'cannot claim official undetectable'; 'Rewording degrades the copy'; 'Layer A does not remove token-sampling watermarks'; the rewrite should use 'a model != suspected origin.'
  • Repo references/ethics.md: flags 'Academic fraud or misrepresenting AI assistance where disclosure is required' and 'Circumventing lawful transparency or platform disclosure rules' as misuse; 'A removed mark does not mean the content was never AI-assisted.'
  • support.claude.com article 16266773 'How Claude marks AI-generated content' (re-verified Sep 6, 2026): imperceptible watermark 'woven directly into the text itself,' 'may persist through some editing'; files get signed C2PA provenance metadata; models launched on or after Aug 2, 2026 mark at launch.
  • TechCrunch, Aug 15, 2026: Anthropic uses Google DeepMind's SynthID Text approach, 'low-stakes choices' in word selection, 'detectable to anyone who has a key'; light editing likely preserves it, 'a complete rewrite where every word is replaced' removes it; Anthropic 'plans to release a watermark detection API,' no timeline.
  • BleepingComputer, Aug 13, 2026, 'AI watermark removers flood the web. Almost none can prove they work': Anthropic's detector isn't public, so tools test against ordinary AI detectors instead; the repo author himself says real removal needs heavy rewriting with a second model. Anthropic's usage policy bars presenting results 'as human-generated.'
  • GPTZero (Jan 26, 2026): 93.5% recall on paraphrased text across 12+ humanizer tools. Turnitin press release (Aug 27, 2025): 'AI bypasser detection' shipped specifically to flag humanizer-modified text. Originality.ai: 96.7% on paraphrased RAID content (Oct 2025). Independent 2026 tests show GPTZero dropping to 40% or below on some humanized samples, and Turnitin admits ~1% false positives on genuine human work.
  • Funnel (Sep 6, 2026): guides.tenfoldmarketing.com routes readers to 'AI Builders - Ten Fold' on Skool: $59/mo or $29/mo annual, 81 members, 'pricing increases after 150 members,' Claude Code Mastery modules, '56+ custom skills.' 'Comment water' is the DM hook into that ladder.

What doesn’t

  • The caption promises what the repo's own docs and Claude's on-screen output both say cannot be promised.
  • The demo runs the 'humanize' rewrite inside Claude, which the skill itself says not to do; Claude's rewrite ships with a fresh watermark.
  • Nobody, including the tool's author, can test against Anthropic's real detector because it hasn't been released.
  • The 30-second 'trick' is a lead magnet for a $59/mo community, not a guide.

The catch

Stripping Unicode and metadata is real and takes seconds, but the text watermark lives in the word choices, and the only way to break it is a heavy rewrite by a different model that degrades the copy and still can't be verified. If you then hand it to a school, client, or platform as human-written, the tool's own ethics file calls that misuse and Anthropic's usage policy calls it a violation.

How to actually do it

  • If privacy is the goal, strip EXIF/C2PA from your own photos and files with any metadata tool (or Layer A of this repo). That part is legitimate and verifiable.
  • Treat Claude as the drafter. Rewrite the draft yourself in your own words, add your facts, quotes, and opinions. That is the only 'humanize pass' that holds up and it makes the writing better, not worse.
  • Disclose AI use wherever a rule requires it: school policies, client contracts, YouTube's altered-content label, Amazon KDP's AI declaration. The mark proves processing, not authorship, so disclosure costs you nothing.
  • For SEO, ignore detectors and follow Google's actual rule: scaled AI pages 'without adding value' get demoted. Fewer, better pages beat laundered bulk.
  • Ignore any 'comment X for the guide' funnel that promises 'undetectable.' If they could prove it, they'd publish the test.

You can clean metadata in 30 seconds; you cannot buy 'undetectable,' and the one detector that matters for Claude text isn't even public yet.

Confidence
High
Posted by
Tenfoldmarc (AI-tips Facebook page)

See the original claim →

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.