Testing the SynthID Detector

With all the talk about watermarks, I began to wonder when they would serve a practical use.

So I tested the SynthID detector at https://synthid.com/.

Note that SynthID only works on media. So I don’t know what it would do if I uploaded a PDF of a Bredebot post.

AI image detection

But I did upload this image, which you’ll see in less than seven days.

Google Gemini.

SynthID identified this as created by AI at some point.

AI video/audio detection

It got a little more granular when I uploaded the Sí video, a Canva-created mp4 file with Canva-generated text, two instances of a Google Gemini image, and Google Lyria music. Check the bottom of the report below.

From SynthID.

For the video, SynthID said that there were AI-generated images in the middle of the video. (It missed the final part of the video, which mixed human-generated text with a smaller version of the AI image.) SynthID did detect the synthetic nature of the audio.

Not truly granular, but better than nothing I guess.

So what?

But what’s the use case, other than as a bludgeoning tool to whack people who use AI in any form?

Actually there is one. Let’s say I post a video with a picture of a Mexican businessman. Is that a real picture?

Or you hear a tune in the background. Did a human write it?

This tool showed that neither the Mexican businessman nor the background music were of human origin. So what?

Any results need interpretation—preferably human interpretation.

Or maybe not.

2 Comments

Leave a Comment