Google has begun testing a digital watermark, SynthID developed by DeepMind, to identify images created by AI, in order to fight disinformation.
The way it works is by embedding changes to individual image pixels, so the watermarks are invisible to the human eye, detectable by computers.
DeepMind, however, have claimed it’s not “foolproof against extreme image manipulation.”
It’s becoming a lot harder to tell the difference between real and AI generated images, and AI image generators are becoming more mainstream, through Midjourney which has over 14.5 million users.
These allow individuals to create images within a matter of seconds, using simple text instructions, which has led to queries regarding copyright and ownership globally.
Google already has Imagen, an image generator, and creating/checking watermarks currently only applies using this tool.
Watermarks are a logo/text added to an image for ownership, and make it harder for the image to be copied and used.
However, these types aren’t suitable for the identification of AI images, as they can be edited or cropped.
Now companies have started using “hashing,” which creates digital “fingerprints” of known videos of abuse, enabling them to identify and remove them quickly before they spread. However, these can become corrupted through cropping/editing.
The system used by Google to create an invisible watermark, allows individuals to identify whether the image is real or AI generated.
Head of Research at DeepMind, Pushmeet Kohli claimed the system modifies images very subtly “that to you and me, to a human, it does not change.”
He said even when the images has been cropped/edited, the software can still identify watermarks.
“You can change the colour, you can change the contrast, you can even resize it… [and DeepMind] will still be able to see that it is AI-generated.”
He noted it was an “experimental launch” DeepMind requires people to use it so they can learn more about how robust it is.
Last month, Google was one of the seven leading companies in AI to sign up for a voluntary agreement, ensuring the safe development/use of AI.
Mr Kohli said this move reflected the commitments, however, other including Claire Leibowicz, from campaign group Partnership on AI, claimed there needs to be more coordination between businesses.
“I think standardisation would be helpful for the field. There are different methods being pursued, we need to monitor their impact – how can we get better reporting on which are working and to what end?”
“Lots of institutions are exploring different methods, which adds to degrees of complexity, as our information ecosystem relies on different methods for interpreting and disclaiming the content is AI-generated.”
Microsoft and Amazon were other among those pledging to watermark AI generated content.
Meta has also published a research paper claiming watermarks will be added to generated videos for similar demands.
Watermarked AI-generated images have been banned in China since the beginning of this year.
