The Tech Blog

I wanted to discuss SynthID today because AI image detection has reached a level where even pros aren’t able to detect if its real or not. Its bound to get worse over time. Google launched SynthID in 2023 for AI-generated images. Now the ecosystem is expanding with OpenAI and NVIDIA adopting the same.

So SynthID is an invisible watermark for AI-generated content. Google built it for images, text, audio and video, but the embedding method changes for each medium. For today, let us stick only to images. SynthID adds a weak hidden signal into image pixel data. You do not see it, but a detector trained for that signal can scan the image and return a confidence score.

Technically, SynthID-Image is described as a post-generation encoder-decoder system. The AI model first creates the image. A neural encoder then makes tiny, distributed pixel changes while keeping the image visually aligned with the original. A neural decoder later checks whether that hidden pattern is still present. Its a trade off basically. If the signal is too strong, image quality drops. If it is too weak, JPEG compression, resizing, filters or cropping can damage it.

Google’s paper gives the general design idea but does not publish the exact encoder architecture, secret keys, threshold logic used in production. Based on what I have seen and tested, it’s clear that the watermark is content-dependent and tested against real image edits like compression, resizing, cropping, rotation, blur, noise etc. The detector is trained to recover that weak signal without needing the original image.

BTW, you can now upload supported media to the Gemini SynthID Detector or OpenAI’s verification tool to check for these signals. So you really can’t pass off images as real from the top AI image model providers. The Chinese and open-weight models haven’t implemented them yet.

If there is interest in this, I will cover the latest reverse-engineering work on SynthID next, specifically how some folks are trying to reverse these watermarks. Let me know.