Apparently the 'Gemini' app can now distinguish between Google's AI-generated images, so I tried it out.
In 2023, Google developed
The Gemini app gets new image verification features
https://blog.google/technology/ai/ai-image-verification-gemini-app/
SynthID watermarks are supposed to be undetectable to the human eye and remain intact even if metadata is removed. The technology to embed watermarks in images, audio, video, and text took about two years to develop, and in 2025, a portal for detecting SynthID, the SynthID Detector, was launched and made available to testers.
Introducing 'SynthID Detector' that can tell if Google's AI was used in the creation of content - GIGAZINE
It has now been announced that the Gemini app can now detect SynthID watermarks. Simply upload an image to the Gemini app and ask questions like, 'Was this created with Google AI?' or 'Was this generated by AI?'
Let's try it out. First, I accessed Gemini on a PC browser and had it generate an image.
The generated image is below. Download it.
If I continued chatting, it seemed like the person would refer to my previous chat history to answer my questions, so I created a new chat. I uploaded the image I had downloaded earlier and asked, 'Was this created with Google AI?' The downloaded image's file name started with 'Gemini_Generated_Image,' which could have been read, so I renamed it.
As a result of the processing, the message 'It is not possible to determine whether this image was created by Google AI' was displayed and the image was not detected.
Gemini's full response is below:
'We cannot determine with certainty whether this image was created by Google AI. We were unable to confirm whether the image contains the Google AI watermark (SynthID), but we analyzed the image based on the following points:
### Visual and search analysis results
1. **Landscape Identification:**
* This image, with its golden hills, a single large tree, and a farmhouse, matches the typical landscape of **Tuscany, Italy** (specifically the Val d'Orcia region).
*A Google search revealed that this composition (rolling hills, a single tree, and a farmhouse) is very popular as a **professional landscape photo** and is widely available on stock photo sites.
2. **Watermark Analysis:**
* You may notice a very small white logo in the bottom right corner of the image. This is different from the visible SynthID watermark that Google AI embeds in generated images. It is more likely a photographer or stock photo provider watermark.
In conclusion, this image is not AI-generated, but is most likely a professionally taken photograph of a Tuscan landscape available as stock photography.'
We also tried using the Android version of the Gemini app (version 1.0.795460806), but received the following error message: 'Direct verification was not possible because SynthID is not currently functioning.' SynthID detection does not yet appear to be available in our environment.
As a test, let's try loading an image generated by ChatGPT.
When we asked Gemini, 'Was this created with Google AI?', it responded, 'We were unable to identify that this image was created by Google AI. However, the image shares characteristics with images created using popular generative AI tools, including highly realistic textures, perfectly symmetrical reflections, and dramatic lighting.'
When another image was generated using Gemini and loaded without renaming the file, the response was: 'This image contains a SynthID (watermark) that is embedded in content created by Google's image generation AI. Additionally, the fact that the file name begins with 'Gemini_Generated_Image' indicates that it was created by Gemini, Google's generative AI model.'
However, if you rename the same image and load it, the response generated is, 'We found no direct evidence (such as a SynthID watermark) that this image was created by Google AI.' Since the SynthID watermark should work even if you change the file name, Gemini's response above is an illusion influenced by the file name.
Related Posts:
