I Tested Google's New AI Image Detector
Tech journalist David Nield tests Google's free web-based SynthID Detector, finding it robust against basic edits but beatable with heavy pixelation or mixed content.
Stock photo for illustration only, not from the actual event
- Google launched a free web tool to detect SynthID watermarks across images, video, and audio.
- The detector successfully flags AI media from Gemini and ChatGPT even after cropping or filtering.
- Heavy pixelation and images with only 25% AI content can successfully bypass the tool.
- Users face a daily quota limit of approximately 10 files per media type.
David Nield, a technology journalist based in Manchester, U.K., has put Google's newly launched standalone SynthID checker tool to the test. Available for free on the web, the tool aims to help users determine whether media files were generated by artificial intelligence.
To combat the spread of unverified synthetic media, AI images produced by platforms like Gemini and ChatGPT come embedded with an invisible watermark known as SynthID. With Google's new online portal, anyone signed into a Google account can upload files to verify their origins without needing visible branding.
Stock photo for illustration only, not from the actual event
The checker supports an extensive array of file formats, including JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, and GIF for images, alongside video and audio formats. The SynthID standard itself is backed by major industry players such as Nvidia, OpenAI, Kakao, and Google.
During initial tests, Nield found that the detector quickly identified images downloaded directly from Gemini and ChatGPT, even correctly naming the specific AI tool used. It also accurately marked genuine photographs as AI-free, noting the absence of the SynthID watermark.
"I started with some easy tests. The detector correctly spotted that images created in Gemini and ChatGPT, and downloaded directly to disk, were indeed fake."
David Nield
Nield then pushed the limits by testing screengrabs, applying Photoshop oil painting filters, adjusting colors and contrast, and even cropping and combining two different AI images. The SynthID detector successfully identified the watermark through all of these transformations, proving the signature is woven deep into the file's pixels rather than relying on metadata.
Because the watermark is embedded directly into pixel data and frequency components rather than surface metadata, standard cosmetic edits cannot easily strip it away. However, this architectural reliance means that significantly degraded data or minimal AI contributions leave the detector with too little signal to work with reliably.
Despite its robustness, Nield managed to bypass the detector on two occasions: once by heavily pixelating an AI image to the point of rendering it useless, and another time by creating a composite image consisting of 75% real photography and only 25% AI-generated content, where the reduced AI footprint caused the detector to falter.
Source: Lifehacker
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment