Google’s AI watermark
SynthID detector: what’s possible
SynthID is Google’s invisible watermark for AI images, video, audio and text. It can only be read by Google. GPTTrace can’t detect it — but it can read the labels Google puts next to it, and check the content itself.
Free · no sign-up · your file never leaves your device
Drop an image, click to choose, or paste
JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device
Drop an audio file or click to choose
MP3 · WAV · M4A · FLAC · OGG · Opus — the first 60 seconds are analysed on your device
Drop a video or click to choose
MP4 · MOV · WebM — 8 frames are sampled and checked on your device
SynthID in one paragraph
Google DeepMind introduced SynthID in 2023 to watermark images from its Imagen model, and has since extended it to video (Veo), music and speech (Lyria and other audio models) and text (Gemini). The watermark is embedded by a neural network during generation and is invisible to people. A matching detector network can later recognise it, even after the content has been cropped, compressed or re-recorded. The detector is the part Google keeps to itself.
What GPTTrace checks instead
Google rarely relies on SynthID alone. Its products also write standard metadata, and that metadata is open to everyone:
- IPTC digital source type. Google labels AI images with trainedAlgorithmicMedia, which many apps display as “Made with Google AI”. GPTTrace reads it and treats it as proof.
- Content Credentials. As a C2PA steering member, Google has been adding signed manifests to content from its products. GPTTrace reads the signer and the actions.
- Visible marks. Consumer Gemini images and Veo videos have carried small visible watermarks; a suspiciously cropped corner is a hint.
When the metadata has been stripped, GPTTrace’s neural classifier and forensic checks examine the content itself, independent of any watermark. Those checks work on images from any generator, not only Google’s.
A practical approach
If you suspect an image, clip or song came from a Google model, run it through GPTTrace first: if the IPTC label or a Google-signed credential is present, you have your answer in seconds without uploading anything. If not, use Google’s own SynthID verification for a definitive yes or no on Google content, and treat the GPTTrace classifier result as the general-purpose check for everything else.
The bigger lesson is that watermarks and labels only cover the companies that use them. Open-source models running on someone’s own computer add neither, which is why content analysis remains necessary.
Why this page doesn’t pretend
Search for “SynthID detector” and you will find sites that claim to detect it. They can’t, unless they are Google. We would rather explain what is actually possible and give you the checks that do work, than show a fake “SynthID: not found” that would mislead you into trusting an image.
What Google Imagen / Gemini and Google Veo and Google Lyria / MusicFX leave in a file
These are the traces GPTTrace checks for. “Some files only” means the trace is often missing — a re-save, screenshot or social-media upload removes metadata — so its absence proves nothing.
| Generator | Where | What to look for | How reliable |
|---|---|---|---|
| Google Imagen / Gemini (Nano Banana) | Pixels | SynthID invisible watermark — only Google’s own tools can read it | Documented by the vendor |
| Google Imagen / Gemini (Nano Banana) | IPTC / XMP | DigitalSourceType = trainedAlgorithmicMedia (“Made with Google AI”) | Documented by the vendor |
| Google Imagen / Gemini (Nano Banana) | C2PA manifest | Content Credentials on images from recent Gemini and Pixel versions | Some files only |
| Google Imagen / Gemini (Nano Banana) | Pixels | A small visible sparkle watermark in the corner on free-tier Gemini images | Commonly seen |
| Google Veo | Frames | SynthID watermark in every frame (not readable outside Google) | Documented by the vendor |
| Google Veo | Frames | Visible “Veo” watermark on most consumer exports | Commonly seen |
| Google Veo | C2PA / XMP | Content Credentials or IPTC DigitalSourceType on some exports | Some files only |
| Google Lyria / MusicFX | Audio | SynthID audio watermark — only Google’s tools can read it | Documented by the vendor |
How accurate is it? Our measured numbers
We test GPTTrace on labelled image samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.
image check: AUC 0.938 (cross-validated)
679 labelled samples (399 AI, 280 human), run 2026-10-08. At the “Likely AI” line it caught 64% of AI samples and wrongly flagged 5% of human ones.
| Source | Truth | Samples | Result at “Likely AI” |
|---|---|---|---|
| gemini-nano-banana | AI | 40 | 40% caught |
| midjourney-v6 | AI | 40 | 68% caught |
| midjourney-v5 | AI | 40 | 73% caught |
| flux-dev | AI | 40 | 13% caught |
| flux-schnell | AI | 40 | 48% caught |
| sdxl | AI | 40 | 100% caught |
| gpt-image | AI | 40 | 30% caught |
| kling | AI | 39 | 97% caught |
| leonardo-stablecog | AI | 40 | 98% caught |
| bitmind-imagine-mix | AI | 40 | 80% caught |
| fullsize-photos | Human | 40 | 10% wrongly flagged |
| open-images-photos | Human | 40 | 5% wrongly flagged |
| lfw-faces | Human | 40 | 0% wrongly flagged |
| caltech-objects | Human | 40 | 3% wrongly flagged |
| coco-photos | Human | 40 | 0% wrongly flagged |
| ffhq-faces | Human | 40 | 0% wrongly flagged |
| celeba-faces | Human | 40 | 18% wrongly flagged |
Data sources and method: methodology & accuracy.