BackWe Tested Google's New SynthID Detector: 100% on Originals, 18% on SynthID Removed AI Images
Google opened its SynthID Detector to everyone today. We ran Nano Banana images through it after a public tool stripped the watermark, then ran the same files through AI or Not's AI image detector.
Google today launched their SynthID Detector. This is a great service for anyone to check if an image was originally generated by Google, or any of its partners who utilize SynthID: OpenAI, NVIDIA, Kakao.
(Technically it opened to everyone today, after a limited release to media professionals last year.) We ran it against Google Nano Banana images that had SynthID removed, then ran the same files through AI or Not's AI image detector.
Key Takeaways
- The SynthID Detector identified 18% of the images after a public removal tool stripped SynthID from them. Without crediting the one "Small edit / Tampered" verdict, it is 16%.
- On intact originals it identified 100%.
- 82% of the removed images returned "SynthID was not detected." The tool's own text says that result does not rule out AI.
- AI or Not's AI image detector identified 100% of the originals and 100% of the removed images.
- Google's post says SynthID has watermarked over 180 billion images and videos.
Detection rates on Nano Banana images, before and after SynthID removal.
What Google Announced
Google DeepMind's SynthID post states: "Now, we're expanding access to everyone – with the tool available globally in English starting today." The tool lives at synthid.com and checks images, video and audio.
The numbers in the post:
- SynthID launched in 2023. Google says it has watermarked "over 180 billion images and videos, along with 240,000 years of audio content."
- Verification built into Search, the Gemini app and Chrome "regularly handle over 1 million requests daily."
- Partners named: OpenAI, NVIDIA and Kakao. Apple is listed as "soon."
- The pitch: "anyone can easily check if an image, video, or audio file was made with AI from Google or our partners."
Not all images are kept as originals and not all images are generated from a SynthID partner. So we put it to the test.
Does the SynthID Detector Work in Detecting AI Images?
Yes.. as long as it checks images made with SynthID baked in.
SynthID is a signal Google writes into the pixels of an image at the moment it is generated, and the detector looks for it. When the signal is there, you get a hit. We tested that on intact originals from the set: every one came back "SynthID was detected," with the text "This media was made or edited with Google AI."
A free check like this is now required by law: California's AI Transparency Act since August 2 (every public generator, as of SB 1000), and the EU AI Act's Article 50.
With the watermark intact, both tools identified every image they were given.
How We Tested the SynthID Detector
- Dataset: Google Nano Banana images generated for an earlier internal test (the unedited set). Each is a triplet: the original with SynthID, the same image with SynthID removed, and a screenshot of the synthid.com verdict on the removed version.
- Removal: a publicly available SynthID-removal tool published on GitHub.
- SynthID Detector: manual upload at synthid.com/image-detection on October 7, 2026. We saved all resulting screenshots.
- AI or Not: ran every image against the image detection API.
The per-image results for every triplet, with every screenshot, are available as a data download.
How Does SynthID Detection Do on Images Generated with Nano Banana but Edited with a SynthID Remover?
After SynthID removal.
18% came back as detections. Most said "SynthID was detected." One, the lonely lighthouse, said "Small edit / Tampered." The rest, 82%, said "SynthID was not detected."
The chef's kitchen scene is the honest one: a Nano Banana image that went through the removal tool and still got caught.
The chef's kitchen scene, one of the ones that still got caught. Source: Google SynthID Detector, synthid.com, 2026-10-07.
Most looked like this. Same picture, two verdicts:
The weathered captain. Left, the original with SynthID and its verdict. Right, the same image with SynthID removed and its verdict. Source: Google SynthID Detector, synthid.com, 2026-10-07.
The freckled portrait. Try to spot the difference.
The vintage Hollywood portrait. AI or Not's AI image detector called both versions AI.
Every image in the test with SynthID removed, with the SynthID Detector verdict and AI or Not's AI image detector verdict under each.
The SynthID Detector is working as designed here, and it says so on screen. For those, the heading was "SynthID was not detected," followed by: "It is unlikely that this media was made with AI from a SynthID partner." And then: "Note however, that this may have been generated by a non-partner organization, or at a time prior to a partner adopting SynthID."
That is a statement about the watermark. It says nothing about whether the picture is AI.
How Well Does AI or Not Do on AI Image Detection?
AI or Not's AI image detector identified every original and every removed image, 100% on both. It does not look for a watermark. It reads the pixels, so a even after a SynthID removal, the AI artifacts still show.
We saw it again when we looked at cropped Meta AI images: AI or Not identified 98%, Meta's own detector 35.3%.
For newsrooms, trust and safety teams, insurers and platforms: run the free watermark check, because a hit is a strong answer. A "not detected" is the cue to run the file through an AI image detector, like the one on aiornot.com, instead of filing it as "probably real."
FAQ
Does the SynthID Detector work?
Yes, on images that still carry SynthID. In our test it identified every intact Nano Banana original. After a public tool removed the watermark, it identified 18%.
What does "SynthID was not detected" mean?
It means the tool found no SynthID watermark in the file. Google's own text says it is "unlikely that this media was made with AI from a SynthID partner" and adds that it "may have been generated by a non-partner organization, or at a time prior to a partner adopting SynthID." AI or Not still detected 100% of those images.
Can SynthID be removed?
Yes, there are many commercial and open source options available to do so.
How reliable are AI image detectors?
A watermark checker, like SynthID, only has an answer while a visible or invisible watermark is on the file. AI or Not's AI image detector identified 100% of the AI-generated files in this test (the originals and the versions with SynthID removed).
How do AI image detectors work?
An AI image detector examines the pixels for traces that generative models leave behind. It needs no metadata and no watermark, which is why AI or Not's AI image detector gave the same answer on the originals and the removed versions.
Is an AI detection tool required by law?
Yes, a free detection tool is now required of AI generators: California's AI Transparency Act (SB 942) has been operative since August 2, 2026 and, since SB 1000 was signed on September 30, covers every generator the public can reach in California, while the EU AI Act's Article 50 has applied since the same date. The mandated tool verifies the provider's own label, which is the kind of check the SynthID Detector performs.
What do you get with AI or Not?
Instantly get your AI detection API to start building, and protecting.
AI detection covering images, text, video and audio.
All content checked gets deleted, instantly.
Start detecting AI for free, scale with pay-as-you-go.
AI to fight AI.
