SynthID is not a character you can delete
This is the single most important thing to understand, and it is the thing the rest of this market is quiet about.
Google's SynthID for text does not work by inserting hidden characters. It works by adjusting the probability distribution the model samples from when it picks each word, nudging it toward certain choices in a pattern that a detector holding Google's key can recognise statistically across a long enough passage. The watermark is the sequence of words. There is no extra byte in the file.
The consequences follow directly. You cannot detect SynthID without Google's key, so no third-party tool can tell you whether text carries it. That includes this one, and every tool that claims otherwise. You cannot remove it by stripping characters, because there are no characters to strip. Deleting invisible characters from SynthID-marked text leaves the watermark completely intact.
The only thing that disturbs a statistical watermark is changing enough of the actual word choices: rewriting or paraphrasing the passage substantially, ideally through a different model. Even then, nobody outside Google can verify whether it worked, because verification needs the key. Any tool showing you a confident "SynthID removed ✓" is showing you a decoration.
What Gemini text does carry that we can remove
Gemini output regularly contains invisible Unicode characters on top of SynthID, and those are ordinary data we remove completely and verifiably. Zero-width characters, unusual space variants, sometimes private-use codepoints picked up in formatting.
It also carries the usual typographic fingerprints: em-dashes, curly quotes, the rule-of-three habit. The checker below reports those and can normalise them.
So the useful framing is: run the check, get a complete answer about the character layer, and treat the statistical layer as a separate problem that character tools do not touch.
Gemini images: metadata comes off, pixels do not
Images from Gemini, Imagen or Nano Banana ship with a C2PA manifest and often an XMP block naming the generator, plus a SynthID mark embedded in the pixels themselves.
The metadata layer we remove completely and losslessly. We drop the C2PA manifest, the EXIF block, the XMP packet at the container level, copying the compressed pixel data through byte for byte so there is no quality loss. After that, a Content Credentials verifier finds nothing.
The pixel-domain SynthID mark survives all of that, and it survives cropping, resizing, re-compression, even a screenshot, by design. Removing it means regenerating the image through a diffusion model, which by definition changes what the image looks like. When we detect signs that an image came from a SynthID generator, the tool says so explicitly rather than letting you assume the file is now clean.