What SynthID actually is (and why no tool can remove it from text)

If you have searched for a way to remove SynthID, you have probably found a dozen tools promising to do it. Most of them are selling you a zero-width character remover with a different label on the box.

This is what SynthID is, why that mismatch happens, plus what you can actually do.

The one-sentence version

SynthID does not add anything to your file. It changes which words the model picks, or which pixels the image generator produces, in a pattern that a detector holding Google's key can recognise statistically.

Everything confusing about SynthID follows from that sentence.

How the text watermark works

When a language model generates text, it does not pick the single most likely next word. It samples from a probability distribution across thousands of candidate tokens, which is why the same prompt gives different answers each time.

SynthID intervenes in that sampling step. Using a secret key, it generates a pseudo-random function that scores candidate tokens, then nudges the model toward tokens that score well. The nudge is small enough that quality does not visibly suffer. Google ran a large-scale live comparison on Gemini traffic and found no meaningful degradation in user-rated response quality.

Across a short passage that bias is invisible. Across a few hundred tokens it becomes a measurable statistical signature: the text uses more high-scoring tokens than chance would produce. A detector with the same key can recompute the scores and measure the correlation.

The watermark, then, is the word sequence. Not a character in the word sequence: the sequence itself.

Why this makes removal tools impossible

Three consequences follow, and each one kills a category of product being sold today.

You cannot detect it without the key. The scoring function is keyed. Without Google's key you cannot compute which tokens were favoured, so you cannot measure the correlation. Any third-party tool claiming to tell you whether text carries SynthID is producing a number with nothing behind it. Google runs a detector portal for its own ecosystem; there is no public API and no key available to anyone else.

You cannot remove it by cleaning characters. There are no characters to clean. Strip every zero-width space, every tag character, every odd whitespace variant from SynthID-marked text and the watermark is entirely intact, because none of those were carrying it. This is the specific claim to watch for: a tool that removes invisible characters is doing something real, but it is not doing this.

You cannot verify removal. Even a genuinely effective attack, such as rewriting the text substantially through a different model, leaves you unable to confirm it worked, because confirmation needs the key you do not have.

What actually disturbs it

Since the watermark is the word choice, changing enough of the word choices degrades it. Paraphrasing through a different model is the standard approach, and the research literature confirms it reduces detection confidence.

But note what that costs. You are no longer removing a mark from your text; you are producing different text. And you still cannot verify the result. Anyone selling "SynthID removal" as a service is selling a paraphrase with a confident progress bar attached.

Truncation also matters: the signal needs length. Google has said detection is unreliable on short passages, roughly under a couple of hundred tokens. A few sentences probably do not carry a readable mark.

SynthID for images is a different mechanism

The image version embeds the watermark in the pixels: a pattern distributed across the image, applied during generation, designed to survive the transformations images normally go through. Cropping, resizing, JPEG compression, colour adjustment, screenshotting: it is built to survive all of them.

This is the part that catches people out. When you strip the C2PA manifest and EXIF data from a Gemini or Imagen image, the pixel watermark is untouched. That strip is a real and complete operation, and this tool does it losslessly. Content Credentials verifiers will report nothing. Google's SynthID detector will still read the mark.

Removing it means regenerating the image through a diffusion model, which changes what the image looks like. There are research techniques for this. They are not a metadata operation and no metadata tool performs one.

Where invisible characters fit in

They are a completely separate phenomenon that gets conflated with SynthID constantly.

Invisible Unicode characters such as zero-width spaces, tag characters, odd whitespace genuinely do appear in AI output, including Gemini's. They are real, they are detectable by anyone, and they are removable completely. We find and remove them, and you can verify the result by re-running the check.

They are just not SynthID. Two different layers, two different answers:

	Invisible characters	SynthID
Where it lives	Extra characters in the file	The word or pixel choices
Who can detect it	Anyone	Only Google
Removable?	Completely	Not by any file operation
Verifiable after removal?	Yes	No

The practical upshot

If you want to know whether your document contains hidden characters, that is a question with a definite answer, and you can get it in about ten seconds for free.

If you want to know whether it carries SynthID, nobody outside Google can tell you, and you should treat any product claiming otherwise as evidence about that product rather than about your document.

Being clear about which question you are asking is most of the work.