{"ok":true,"source":{"repo":"https://github.com/guillaumemeyer/watermarks-remover","localClone":"/Users/mike/Code/watermarks-remover","reviewedCommit":"bcf497260af41b7af5ece3ca359e9631320f9c1a","license":"MIT","reviewedAt":"2026-08-24","stance":"Used as adversarial/tamper-risk research only. AI Detect does not run or expose watermark removal."},"records":[{"riskId":"layer-a-invisible-unicode","label":"Invisible Unicode and carrier cleanup","target":"Text with zero-width characters, bidi controls, tag characters, exotic spaces, or edit-based carriers.","mechanism":"Deterministic text normalization can remove visible-invisible carrier characters without needing a provider detector.","riskLevel":"high","detectionMeaning":"If this class of carrier is absent, the text may be clean, already normalized, or previously scrubbed.","whatAiDetectCanReport":["Whether obvious text-carrier artifacts are found when a detector is connected.","That absence of visible-invisible carriers is not a provider-watermark verdict."],"limitations":["A missing watermark or provenance signal does not prove human authorship.","A text-only check cannot tell whether a signal was never present, unsupported, removed, or weakened by normal editing.","Provider watermark verification can require private keys, official APIs, or compatible generation pipelines.","Modern Claude/SynthID-style text watermarks are described as statistical wording signals, not hidden Unicode."],"sourceUrls":["https://github.com/guillaumemeyer/watermarks-remover"]},{"riskId":"layer-b-statistical-rewrite","label":"Statistical text watermark weakening","target":"Token-sampling watermarks such as SynthID-style, Kirchenbauer green-list, keyed-Gumbel, or similar wording-distribution marks.","mechanism":"Substantial rewriting, paraphrase, or regeneration can move token choices away from the original marked distribution.","riskLevel":"medium","detectionMeaning":"A weak or missing statistical watermark can mean unsupported detector access, too little text, normal edits, or an intentional rewrite.","whatAiDetectCanReport":["Provider support status for public watermark verification.","A tamper-risk caveat on reports where watermark detection is unsupported or uncertain.","Separate AI-writing risk signals without treating them as watermark proof."],"limitations":["A missing watermark or provenance signal does not prove human authorship.","A text-only check cannot tell whether a signal was never present, unsupported, removed, or weakened by normal editing.","Provider watermark verification can require private keys, official APIs, or compatible generation pipelines.","The reviewed repo describes this path as best-effort rewrite, not guaranteed provider detector bypass.","Benchmarks against MarkLLM/SynthID-like schemes do not prove Google, Anthropic, or another vendor detector will fail."],"sourceUrls":["https://github.com/guillaumemeyer/watermarks-remover","https://github.com/guillaumemeyer/watermarks-remover/blob/main/docs/synthid-text-benchmark.md"]},{"riskId":"file-provenance-metadata","label":"C2PA and file metadata stripping","target":"Images, PDFs, Office files, HTML/Markdown, audio/video containers, and other files carrying provenance metadata.","mechanism":"Container cleaners can remove C2PA manifests, EXIF/XMP hints, document properties, YAML front matter, and similar metadata layers.","riskLevel":"high","detectionMeaning":"No C2PA manifest found means no local manifest was found; it does not prove the file was never AI-generated.","whatAiDetectCanReport":["Whether a browser-side C2PA manifest is present in supported file checks.","That export, screenshots, re-encoding, platform processing, or deliberate stripping can remove metadata evidence."],"limitations":["A missing watermark or provenance signal does not prove human authorship.","A text-only check cannot tell whether a signal was never present, unsupported, removed, or weakened by normal editing.","Provider watermark verification can require private keys, official APIs, or compatible generation pipelines.","Soft-bound provenance, pixel-domain watermarks, or remote credentials may remain outside local metadata checks."],"sourceUrls":["https://github.com/guillaumemeyer/watermarks-remover","https://c2pa.org/specifications/specifications/2.2/index.html"]},{"riskId":"media-pixel-watermarks","label":"Pixel and media watermark uncertainty","target":"Images, generated media, or audio/video that may include imperceptible in-content watermarks.","mechanism":"Removal risk depends on the watermark family, transform robustness, regeneration, compression, cropping, and available verifier.","riskLevel":"varies","detectionMeaning":"Metadata absence does not answer whether an in-content watermark or soft-bound provenance exists.","whatAiDetectCanReport":["C2PA/provenance metadata when present.","A limitation that pixel/media watermark support needs provider-specific verification."],"limitations":["A missing watermark or provenance signal does not prove human authorship.","A text-only check cannot tell whether a signal was never present, unsupported, removed, or weakened by normal editing.","Provider watermark verification can require private keys, official APIs, or compatible generation pipelines.","AI Detect does not currently run vendor image, audio, or video watermark verifiers."],"sourceUrls":["https://github.com/guillaumemeyer/watermarks-remover"]}],"updatedAt":"2026-10-19T15:31:11.796Z"}