On October 7, Google opened access to SynthID Detector, a service that checks content for SynthID watermarks. The detector identifies whether an image, video, or audio file was created or edited using AI. We ran several tests to see how the detector identifies images from different AI models.
The detector is based on Google DeepMind’s SynthID technology, which embeds digital watermarks that are invisible to humans into AI-generated content.
SynthID Detector lets you check content created by Google AI models and partners OpenAI, NVIDIA, and Kakao. Google’s listed supported models include Nano Banana, Veo, and Lyria. Apple is expected to join later.
- SynthID Detector is available at https://synthid.com/
For the test, we selected ChatGPT, Nano Banana 2.1, and Qwen-Image-2.1. To keep the experiment clean, we removed the metadata beforehand.
SynthID Detector Test on an Image From ChatGPT
We generated an image in ChatGPT using a prompt from our article: https://cpa.rip/prompts/ai-prompts-igaming/ and checked it in SynthID Detector:
- Result: SynthID detected. This media file was created or edited using OpenAI.
SynthID Detector Test on an Image From Nano Banana 2.1
In the new Nano Banana 2.1 we generated an image using a JSON prompt from the article: https://cpa.rip/prompts/nsfw-photos/ and checked it in SynthID Detector:
- Result: SynthID detected. This media file was created or edited using Google AI.
SynthID Detector Test on an Image From Qwen-Image-2.1
The next test used an image created with the local Qwen-Image-2.1 model:
Instructions for locally installing the free Qwen-Image-2.1 model: https://cpa.rip/ai/qwen-image-2-1/.
- Result: SynthID not detected. This media file is unlikely to have been created using AI from a SynthID partner. However, keep in mind that it may have been generated by a non-partner organization or before the partner started using SynthID.
SynthID Detector Test After Image Uniquification
First, we checked the image after running it through an image uniquifier, then checked it after applying blur and various graphic elements over the generated image:
- Result in both cases: SynthID detected. This media file was created or edited using Google AI.
SynthID Detector Test After Editing With Another AI Model
We took an image generated in ChatGPT, which SynthID Detector identifies as created with OpenAI, uploaded it to Nano Banana 2.1, and made targeted edits.
As a result, a new check in SynthID Detector showed created or edited using OpenAI, meaning that repeated editing in another AI model did not overwrite the original SynthID.
SynthID Detector Limits
The service itself does not provide precise information about its limits, but during testing, we hit a daily limit:
Comments