I used to go crazy about Magic Eye posters back in the nineties. For many years, I just couldn’t figure them out, even though I’d purchased a book filled them. I’d spend an hour at my desk, hunched over, staring cross eyed. Finally, I think I got it — except it was only a piece of the secret image. It would immediately disappear into the blur.
In the end, solving it didn’t bring much excitement. A cutout of an image against a blank background? Yawn.
I thought about this again when I asked Midjourney to create a stereogram. Now, it does produce an image that looks like a stereogram but after staring at it, I didn’t really see anything. I used a tool meant to solve stereograms and it turned out noise.
However, there were instances when staring at the image that I think I might have been fooled into starting to see something three dimensional. This is the trap of generative AI. I can tell you that something is there and it requires other tools to verify the truth. On my own, I couldn’t easily verify that the image didn’t contain a secret code, even though I already had my doubts.
This applies to text or code output of generative AI as well. We will need a discerner system to fact check AIs for accuracy.