Gemini humanizer: 12 tools on the same fields, and what SynthID means for a rewrite

The order is by per-request cap, largest first, because a Gemini answer arrives as one block and a tool that takes it whole produces one rewrite rather than a stitched set. "Names Gemini" records whether the vendor's own product page lists Gemini as a source of text it rewrites, read on 9 November 2026. Plan figures come from each vendor's pricing or product page, checked on 8 November 2026. No column is a detector reading. The published run's 24 inputs include text generated with Gemini, and the Gemini subset column will be added to this table when the first run publishes.

None of those lines mentions a detector reading or a SynthID result, because none can.

What does Google say Gemini text carries?

Google's SynthID page states that SynthID embeds watermarks directly into AI-generated images, audio, text and video across its consumer products, that the marks are imperceptible to people and detectable by SynthID's technology, and that it has "expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience" (Google DeepMind, SynthID, read 9 November 2026). On method, the page says a language model assigns each candidate word a probability score and that "SynthID adjusts these probability scores to generate a watermark", which is not noticeable to the reader and does not affect output quality.

Two gaps on that page matter for a humanizer search. Google does not state which Gemini surfaces beyond the app and web experience carry the text mark, so text from the Gemini API or from Gemini inside other Google products is not described either way, and this page does not guess. And the public checking routes Google lists are for images, video and audio; it names no way for a member of the public to test a passage of text. So a Gemini answer may be marked, Google can read the mark, and you cannot. The general explainer, with the other deployed watermarks, is the AI text watermarks guide.

What does a humanizer do to a SynthID mark?

The reference design is the paper by Kirchenbauer and colleagues. Before each word is generated, a randomised set of "green" tokens is selected and softly promoted during sampling; a statistical test later counts whether a span of text has too many green tokens to be chance, with interpretable p-values and without access to the model, and the paper derives an information-theoretic framework for the watermark's sensitivity (Kirchenbauer et al. 2023, arXiv 2301.10226, read 9 November 2026). SynthID Text, as Google describes it, belongs to the same family: a nudge on word probabilities that leaves a countable pattern.

A count is the point. Every word a humanizer replaces is one token that no longer contributes to the count, so the signal falls in proportion to how much of the text changes, and the paper's sensitivity analysis ties detection strength to the number of tokens carrying the mark. A light edit that changes a word in ten leaves most of the count in place. A strong mode that rewrites most sentences pushes the count toward the chance rate. A pass that also drops a claim or invents a detail has changed the meaning along with the count, which is the cost this site's method rates first.

No vendor on this page publishes which of its mode names corresponds to which share of words changed, and no public instrument reads Google's key, so a vendor "SynthID remover" is a claim no reader can check; one of the twelve vendors advertises exactly that, read on 9 November 2026, and this page records only what Google documents. The last row points to the sibling pages: the Claude humanizer page sets out what Anthropic states Claude text carries, and the ChatGPT humanizer page does the same for OpenAI.

Do the detectors this site uses read the Gemini mark?

No. GPTZero's technology page says it incorporates research in detecting ChatGPT, GPT4 and Google Gemini output among other models, uses an end-to-end deep learning model with a sentence-by-sentence classifier, returns one of three document labels, human, AI or mixed, with a confidence category, and keeps its false positive rate at no more than 1 percent when evaluating AI versus human text (GPTZero, technology, read 9 November 2026). A classifier of that kind reads the statistics of the words on the page. A SynthID check reads whether the word choices follow a hidden rule, and only the key holder can run it. The two can disagree about the same text, and neither names the model that wrote it.

The published method scans every output on GPTZero and four other detectors, records each detector's own label on the run date with a screenshot, and scans every input before any tool touches it so the change is visible. It cannot and does not report a SynthID reading. No detector score should be the only evidence in a misconduct case.

How long is a Gemini answer, and which caps fit it?

The method uses 24 AI-written inputs across six genres, with the school essay paragraph set at 350 to 500 words, and that range is the yardstick for a typical long answer. A 500-word Gemini answer fits Phrasly, Clever AI Humanizer, HIX Bypass, StealthGPT, GPTHuman, Walter Writes AI and WriteHuman Basic in one request; it is four passes on QuillBot free at 125 words per input and seven on BypassGPT free at 80. A 2,000-word draft fits Phrasly paid, Clever AI Humanizer and HIX Bypass in one request and is three passes on Walter Writes AI's Starter plan at 750 words.

Caps from each vendor's pricing or product page, checked on 8 November 2026. Verva, Undetectable AI and Grammarly state no per-request cap, so they do not appear in the "one request" column; an unstated cap is a cap the vendor has not published, not a cap of zero. Splitting carries the extra cost noted above: shorter pieces, and terms that can drift between separate rewrites.