AI provenance
Remove AI generation tags from your files
Files produced by AI image, video and text tools increasingly carry a record saying so. Some of it is an open standard, some of it is vendor-specific, and some of it is invisible characters hidden in the text itself. This page covers what those marks are, which ones can be removed, and which cannot.
The three kinds of mark
1. Metadata tags
The simplest kind, and the most common. An XMP field called
digitalSourceType set to trainedAlgorithmicMedia, an EXIF
Software field naming the generator, a PNG text chunk holding the prompt and
settings, or a video udta box naming the encoder. These sit in the file's headers.
They can be removed completely.
2. Content Credentials (C2PA)
A signed manifest describing how the file was made and by what. There is a fuller explanation here. It is embedded in the file and can be removed, though removing it is detectable in the sense that a verifier will report the credential as missing rather than as valid.
3. Hidden characters in text
Text can be fingerprinted with characters that render as nothing: zero-width spaces and
joiners, Unicode tag blocks in the U+E0000 range, which are a complete hidden ASCII
channel, bidi controls, and lookalike letters such as a Cyrillic о standing in
for a Latin o. These survive copy and paste, which is the point of them. They are
removed from text files, and from the visible body of documents when
Scrub text inside documents is on.
What this cannot remove
This is the part worth reading twice, because plenty of tools are vague about it and there is no good reason to be.
SynthID does not come off. Google's Veo, Imagen and their music models embed a watermark in the content itself rather than in the file headers. In video it is spread across frames and into the audio, and it is engineered to survive re-encoding, cropping, compression and filtering. There is no public method to detect or remove it. Cleaning a Veo clip here removes the metadata and leaves SynthID exactly where it was. A cleaned clip is not a mark-free clip.
Sora's logo moves. A fixed-region eraser can cover a static corner watermark, of the kind Runway, Kling, Luma and Pika use, and that option exists on the Video tab. Sora's mark drifts around the frame, so a fixed rectangle cannot catch it. Covering a moving mark needs per-frame detection and tracking, which is a computer vision problem and is not built here.
Pixel watermarks in images stay. Disturbing a watermark embedded in image pixels means running the picture back through a diffusion model and having it redraw, which needs roughly ten gigabytes of model weights and a GPU. That backend is not installed on this site, so the option is hidden rather than shown as a control that cannot work.
What you are told up front
Before anything is modified, the metadata is matched against a table of known generators, and each one declares its residual marks. If a file came from Veo, the report says so and says the SynthID signal stays, before you press Clean. If it came from a generator whose watermark is a static corner logo, the report says that too, and points at the region eraser. A green "cleaned" badge here means the metadata is gone, and nothing more than that.