Seeing is...

 Michael – Thursday

It’s been quite few years since you could unequivocally add “believing” at the end of that phrase. About ten years ago when I was still teaching image processing at the University of the Witwatersrand in Johannesburg, I had a number of students who did projects on how to identify fake images i.e. images that had been doctored so that they would show something that never actually happened. There was a small but developing discipline called image forensics to try to separate the true from the fake.

Teddy bears working on new AI research underwater with 1990s technology

Some of those fake images were ridiculously crude. A Chinese newspaper ran a picture showing a local dignitary opening a new road, but the dignitary’s feet missed the surface of the road by several inches. Two political leaders who had never met shook hands, yet they cast shadows in different directions.

Even as more sophisticated image-merging techniques were developed, it was possible to determine with a pretty high level of certainty that the images had been tampered with. Basically, when images are doctored, fingerprints remain. When the images come from different sources, the cameras may leave evidence. When two images are merged, their relative compression may well be different.

However, all this is predicated on your starting with an image and then doctoring it. What if you create the whole image from scratch? One’s first reaction is that it will be obvious that it’s not a photograph. Well, that’s true if a human draws it, but how about a computer? The images in this blog were created by OpenAI’s DALL-E software. Do you still think that? You don’t have to draw anything for the software, just tell it want you want…

Girl with pearl earring in her kitchen

Since the software was released about six months ago, over one million users are generating millions of images every day. So seeing is no longer believing. Basically every image has to be treated as potentially fake.

That’s not quite true, at least for DALL-E. You can’t ask the software to produce a picture of President Biden doing something or other, pictures of politicians and celebrities are banned. The reality however, is that what OpenAI can do other deep-learning software companies (and government agencies) will learn to do also. And DALL-E now allows one to upload images and use them as a model. The problem is that it may not recognize that the picture is in fact of one of the celebrities and politicians that it’s supposed to avoid.

How about a “photograph” of the murderer somewhere completely different at the time of the murder? It's possible for OpenAI to embed an invisible code in its images that it could subsequently read and use to provide data about when the image was created and by what version of the software. OpenAI doesn’t say whether it does that or not. But, in any case, the comments in the previous paragraph apply.

If these public-domain images don't convince you, take a look at the ones the Washington Post whipped up in their article particularly the demonstration outside the capitol.

Welcome to the world of seeing is not believing…