Gemini text cannot be verified by the public today: Google's SynthID Detector checks media, not text, so a style-based detector is the practical option. Facts below were checked on 2026-11-10.
Watermark and verification
- Google announced on 2024-05-14 that SynthID would watermark text generated in the Gemini app and web experience. Source: Google DeepMind, "Watermarking AI-generated text and video with SynthID".
- Google says the text watermark is less effective on responses to factual prompts, and that confidence scores can be greatly reduced after heavy rewriting or translation. Same source.
- Google's SynthID Detector opened to everyone, in English worldwide, on 2026-10-07. It checks whether an image, video or audio file was made with AI from Google or its partners. Text is not mentioned. Source: Google blog, SynthID Detector announcement.
- The Gemini app can verify images, videos and audio, and recognises only content made with Google AI tools. Limits include one file at a time and 100 MB per file. Source: Gemini Apps Help.
- SynthID Text is open source, but detection needs the key of whoever ran the watermarking. Source: Google AI for Developers, SynthID documentation.
- We found no public Google tool or API that verifies Gemini-generated text, and no Google statement endorsing or giving accuracy figures for third-party Gemini-text detectors.
Versions
- Gemini 4 Carbon: generally available in the API on 2026-11-10; rolling out to Google AI Ultra subscribers in the Gemini app.
- Gemini 4 Argon: generally available in the API on 2026-10-13; rolling out to Google AI Pro and Ultra subscribers in the app.
- Gemini 3.6 Flash: available to all Gemini app users since 2026-07-21.
- Gemini 3.8 Flash: listed in the API; app availability not confirmed.
Sources: Google blog, Gemini release notes and Google AI Studio documentation. More detail on the versions page.
This detector
- The detector is Veread, Fileverity's text model, and its accuracy is published on the accuracy page.
- 0.5% of human-written texts in seven languages were wrongly flagged.
- 98.3% of fully AI-written texts were detected, and 95.5% of AI-written and AI-assisted texts of every kind.
- 94.9% on AI models never used in development; 92.6% on AI rewrites of human text; 92.2% on human texts continued by AI.
- Minimum text length is 100 words.
- Family results are writing-style similarity, not proof of authorship. GeminiDetect does not read watermarks.
To run a check, use the Gemini AI checker; phrase-level tendencies are on how to tell if Gemini wrote it.