What SynthID Actually Is (And Why You Cannot Delete It)

Most claims about AI watermarking are exaggerated. People say ChatGPT stamps a hidden signature into your text; it does not, and what they are describing is a copy-paste artefact. So it is worth being precise about the one case where the word watermark is literally accurate.

Google DeepMind's SynthID for text is a real, documented, deliberate watermark applied to Gemini output. It is not a hidden character. You cannot find it, see it, or delete it, and the tools advertising a one-click SynthID stripper are selling you nothing at all.

How it works

When a language model produces text, each next word is sampled from a probability distribution over many plausible candidates. Usually several options are roughly as good as each other — "shows", "demonstrates", "indicates", "suggests" might all fit.

SynthID biases that choice. It applies a pattern to which of the near-equivalent candidates gets picked, in a way that is undetectable to a reader — the output still reads naturally, because every word chosen was a plausible word — but that leaves a statistical signature across the whole passage. A detector that knows the pattern can measure whether the sequence of choices in a body of text lines up with it.

Two things follow from that design, and both matter.

The signal is distributed, not located. It is not in any particular word. It is in the aggregate pattern across many word choices, which is why it needs a decent length of text to measure and why it degrades as text gets shorter.

The signal is the text. There is no separate payload riding alongside your words. The word choices are the watermark, so anything that preserves your wording preserves it.

Why every "SynthID remover" is a lie

Look at what the tools in this category actually do. They strip zero-width spaces, non-breaking spaces, word joiners, and byte-order marks — the invisible Unicode described in ChatGPT hidden characters.

Those characters are real and worth removing. They break find-and-replace, they skew word counts, they crash CMS blocks. Removing them is genuinely useful.

They have nothing whatsoever to do with SynthID. Deleting every invisible character in a Gemini response changes the visible word sequence not at all, which means it changes the watermark not at all. A tool marketed as a SynthID stripper that works by stripping characters is performing an unrelated operation and hoping you do not know the difference.

The same goes for anything promising to "remove" it with a synonym swap on a few words. The signal is measured across the whole passage; changing four words in eight hundred moves the measurement by a rounding error.

What does affect it

Rewriting. Genuinely rewriting — different words, different sentence boundaries, different rhythm — replaces the model's choices with someone else's. Google's own documentation acknowledges the signal degrades under heavy editing and paraphrasing, and that is the honest mechanism: you are not removing a mark, you are replacing the sequence it was measured over.

We will not put a number on how much. Nobody can, and here is why:

• The detector is Google's, evaluated on Google's terms, and neither its threshold nor its confidence bands are yours to inspect.
• Detection strength scales with text length in ways that differ per document.
• Any figure a vendor quotes was measured on their samples, against whatever version of the detector existed that week.

Anyone selling you a percentage on this is quoting a number they cannot verify about a system they do not control.

The layer that matters more anyway

Here is the part that gets lost in watermark discourse. For almost everyone reading this, SynthID is not the problem. Nobody submitting an essay is being run through Google's detector portal. What actually happens is that a human being reads the text and recognises where it came from.

Gemini has the most recognisable register of any major model. "Absolutely!" as an opener. "Let's dive in." Delve, landscape, tapestry, realm, a testament to. "In essence" and "ultimately" doing the work an actual conclusion should be doing. Every answer arriving as a bolded header with three bullets under it, whatever the subject happens to be.

A reader clocks that in a sentence and a half, with no detector involved and no watermark consulted. It costs you the same thing a detector flag would — the reader quietly revises their opinion of your work — and it happens far more often.

The useful thing about this is that fixing the visible problem and disturbing the statistical one are the same operation. Rewriting the text so it stops sounding like Gemini is exactly the rewriting that replaces the word sequence SynthID was measured across. You do not have to choose.

What we do, stated plainly

On the Gemini watermark remover and the homepage, pasting a draft and hitting Clean does two things:

1. Strips every invisible character the copy path left behind — a deterministic lookup, no model needed, no promises required.
2. Rewrites the phrasing so the text reads like a person wrote it, holding your numbers, dates, quotes, and citations fixed.

The first is complete and verifiable. The second changes the wording substantially, which is the only honest lever anyone has on a statistical watermark — and we will not tell you what that does to a specific score on a specific detector, because we do not know and neither does anyone advertising otherwise.

What we will say is that your text comes out clean in the byte stream and reads like you wrote it. That is the part we control, and it is the part that matters whether or not a scanner is ever involved.

Related: Gemini watermark remover, how AI detectors work, do you guarantee passing a detector.

Try PassMyEssay free → Clean the characters, rewrite the wording, keep your meaning exactly where it was.