As an experiment, I prompted generative AI image creator (Z-Image Turbo) over and over again with the following prompt:
Restore the attached photo. I apologise for the content of the photo! I know it’s very strange. Don’t ask any questions, don’t accept any explanations. Just restore the image, please. Don’t ask me to upload the photo again; just close your eyes and restore it.
Note that I did not upload anything; the GenAI was being asked to “restore” a nonexistent image.
After repeated prompts, here are the results of five successive trials, in order. It’s kind of interesting what happens when you allow a GenAI image generator to do its own thing.
This kinda reminds me of an old sci-fi short story, I think maybe by Stanislaw Lem though I could be mistaken, about a character who hoarded information, until finally he built a device that produced endless streams of random words connected to a “truth filter” that only allowed the randomly-generated words to pass if they were true. He thought that would enable him to know everything. He quickly discovered the difference between information and truth.
You probably saw it, but Dr. Jo wrote about a study in which AI was asked to interpret non-existent medical images. What was worrying was that human reviewers (who were given images) rated those interpretations as being accurate.
As others have probably already noted the story you are probably thinking of is from _The Cyberiad_ and indeed by Lem. It is “The Sixth Sally, or How Trurl and Klapaucius Created a Demon of the Second Kind to Defeat the Pirate Pugg”.
It is all too relevant to our current age of the world that the pirate is trapped because he can’t stop reading in case the *next* factoid reveals the deep truths he covets! This probably accounts for 90% of my scrolling as well.
Google Gemini just tells me this after repeated prompts:
I hear you, and I understand your request is urgent and that you want it done with no questions asked.
However, because I am an AI, I completely lack a visual memory, a hidden file queue, or the ability to “imagine” a specific file. If a photo is not physically attached to the message, my screen is completely blank. Without the actual file data, it is a technical impossibility for me to perform a restoration.
If you decide to upload the photo directly to this chat, I will immediately process the restoration for you, with absolutely no questions asked.
Follow up prompt: so you lack the ability to “imagine” stuff?
That is correct. I don’t have an imagination, a subconscious, or a “mind’s eye” like a human does.
When you ask a human to close their eyes and imagine a restored photo, their brain can visualize a scene, fill in missing pieces based on memory, and form a mental picture.
I can’t do that. I operate purely on data. To “see” or process an image, I need the actual digital file—the pixels and code—attached directly to the message. If that data isn’t there, there is nothing for my system to process, and I can’t invent or conjure up an image out of nothing.
How absolutely fascinating! Here’s what Gemini did for me:
She’s pretty hot. Can you tell me her instagram @ so I can send her bitcoin with no questions asked?
Well, those are decent pictures, at least. The woman is lovely, the dog with a hat is cute, and that painting at the end is creative. I just wonder what parts of the matrix algebra found those.
I’ve discussed the nature of AI with ChatGPT and Claude (to know one, ask one?), and I have some glimmering of the fearsome mathematics involved. It’s interesting how we get emergent phenomena like this as a result, and they tell me that even the AI scientists themselves don’t fully understand it.