Tools AI Minute Newsroom 2026-08-15

Google says the invisible marks are the real safeguard. The tool that strips them gained 8,700 stars in three days.

On 12 August we reported that watermarks-remover, an MIT-licensed project for stripping AI provenance marks, had appeared with roughly 130 stars on its first day. It now has 8,883 stars and 908 forks. In between, on 14 August, Google made its visible AI badge optional outside the EU and South Korea and pointed to SynthID and C2PA — the invisible layers — as what remains. Those are the layers this project targets. It removes invisible Unicode characters from text and strips C2PA, EXIF and XMP metadata from PNG, JPEG, WebP, SVG, PDF, DOCX, ODT, HTML and Markdown, naming Claude, Gemini/SynthID-Text and OpenAI schemes explicitly. Its README is unusually candid about the limits: pixel-domain removal is an external, unbundled backend described as "a regenerating remover, not a guarantee"; C2PA soft binding, which can re-link a stripped file to a remote manifest, is out of scope; and it states that "no tool can honestly certify that vendor detectors will fail."

Why it mattersThe gap this exposes is between what a provenance system does and what people assume it does. Metadata is the easy layer and it comes off with one command — which means a missing C2PA badge tells you nothing about whether a file was ever signed. The pixel layer is genuinely harder, and the honest answer in this repository is that removing it means regenerating the image and may still leave traces. Anyone arguing that watermarking settles the question of AI disclosure should read the residual-risk table in a tool built to defeat it; it is more precise about what actually survives than most of the announcements are.

#Image Generation

✓ Verified · 3 sources

Read in the app — free, in 9 languages

Google says the invisible marks are the real safeguard. The tool that strips them gained 8,700 stars in three days.

Related stories