Removing an AI mark is not one uniform operation: invisible characters and file metadata can often be stripped deterministically, while statistical text marks are embedded in word choices and may require substantial rewriting. This project separates those cases—and makes clear that a rewrite cannot certify a text will pass a vendor’s detector.
What Sets It Apart
- Different layers for different carriers: deterministic scripts target invisible Unicode, while file cleaners handle provenance metadata across formats including images, office documents, PDFs, and audio/video containers. This lets users remove inspectable metadata without automatically changing the prose.
- Statistical text cleanup is explicitly best-effort: rewriting can disturb token-sampling signals, but it changes wording and may reduce fidelity, voice, or precision. Optional research harnesses can test against matching watermark schemes; they are not universal vendor-detector oracles.
- Agent integration is separated from the cleaning engine: the skill calls a self-hostable Python service over HTTP, while command-line and service interfaces support other workflows. Optional pixel-watermark backends are external and may regenerate or alter media.
- Reports distinguish evidence from uncertainty: the project describes deterministic findings separately from best-effort work and residual signals, avoiding a blanket claim that content is watermark-free.
Who It Suits—and the Trade-off
A good fit for people auditing or cleaning files they own or are authorized to process, especially when privacy or metadata hygiene matters and they can review changes. For text, start with lossless invisible-character cleanup if preserving the original wording is important. Look elsewhere—or retain the original—if you need guaranteed evasion of a provider’s detector, unchanged prose after statistical-watermark removal, or a general-purpose tool for concealing authorship.
