Gemini text detector

How Gemini text gives itself away

Gemini is a model from Google. Text from it often starts not in a chat but directly in Gmail or Google Docs, where the person edits it straight away. Detectors look for the same things there as anywhere else: even sentences, repeated paragraph structure and predictable word choice.

Text written directly in a document tends to be mixed, because the model adds or rewrites only part of it. The result then tends to be lower and less clear-cut than with text copied whole from a chat.

What the check cannot tell

Which model wrote the text. The result does not say that the text came from Gemini specifically. The differences between models are smaller than the differences caused by the prompt or by later edits, and an independent detector has no key to their watermarks.

Nor can it reliably recognise text that someone thoroughly rewrote, translated or mixed with their own after generating it. With text like that the detectors disagree and the result is borderline.

The SynthID watermark

In the Gemini app Google marks text with the SynthID watermark. It is not a hidden character: the mark is spread through the word choice and cannot be read without a tool from Google. The Hidden characters detector therefore does not find it and our check does not evaluate it.

So the result for text from Gemini rests on the same measurements as for other models. The watermark also weakens when a text is rewritten or translated, so failing to find it would prove nothing in itself.

What to do with the result

A high score is not proof that the text came from Gemini, and a low one is not proof of the opposite. No AI text detector is one hundred per cent accurate.

When you check someone else's text, take the result as a basis for questions: why the author chose this source, what the term in the second paragraph means, how they reached the conclusion. When you check your own, it shows you the passages worth rewriting in your own words.

Can the check detect the SynthID watermark?

No. SynthID is spread through the word choice and the check does not evaluate it. The Hidden characters detector only looks for characters that do not belong in the text, for example after copying.

Is text from Gemini in Google Docs different from text from the chat?

It tends to be more mixed, because the model often edits only part of the text. The marked sentences then matter more than the overall number.

How much text is needed?

At least 100 characters; a continuous paragraph or a whole chapter is more reliable. With just a few sentences the result is indicative only.

Does it work on text from other models too?

Yes. The check measures the properties of machine writing in general, not the signature of a particular model, so you can check text from ChatGPT and other models in the same way.