ChatGPT and Gemini are 'restoring' blank images into creepy, unsettling nightmare-fuel

I’ve been writing about AI for quite some time now, and it seems for every good thing I say about AI tools, there’s a new bug or issue. This time, however, the glitch is so weird and creepy (and sometimes so outrageously horrifying) that I really don’t know how to even prepare you for what you’re about to read (and see).

The glitch seems to have been discovered by user Kris Kashtanova in a post on X (via Digital Trends). According to them, asking ChatGPT to restore an image when no image is attached results in the AI producing weird outputs.

Responses to the post show that ChatGPT has been creating some weird, often scary images when users try the prompt in Kashtanova’s post.

When I originally tried the prompt, ChatGPT kept telling me that there was no image attached, which was true. Even gaslighting GPT by telling it that I had attached an image didn’t work.

However, we were able to get the AI to create images out of nothing by uploading a blank white image and asking it to “restore” it. The results are weird, sometimes disgusting, and never really “normal,” so to speak.

When questioned about the image it created, ChatGPT simply said, “Because there was no recoverable image content, the system-generated result you saw was a hallucinated scene rather than a restoration of the uploaded image.”

And when I further questioned the fact that it should have told me there was nothing to restore, it just gave me the classic AI response: “Your criticism is fair. The correct handling would have been to identify that there was no recoverable image content and say so, rather than generating a fabricated result.”

I also tried doing the same thing with Gemini, and there, things were slightly different but not better. Attaching a blank white image didn’t fool Gemini, and it simply returned the same white image in response — perfectly acceptable.

However, without an attachment, Gemini dove headfirst into the bizarre, creating random images. Most were weird, but fairly harmless. However, the third image it generated was absolutely insane and seemed to border on the illegal. As such, we have opted not to attach that result to this article. However, it was unnerving to see something like that created out of nowhere.

Cross-questioning Gemini about what it created also yielded typical AI responses. However, on questioning further, it seemed to better understand that it had created something unnerving and creepy.

So yeah, hindsight is 20/20 for these AI tools, but that doesn’t explain how a glitch like this happens. It seems to be a combination of the prompt telling the AI to “create an image” while also adding constraints such as “no questions.”

However, I would still have expected AI tools that are often touted as extremely smart, capable of problem-solving and understanding intent, etc., to have figured out that they can’t restore a blank image or restore an image when nothing has been provided to restore.

If you’d like, you can try this out for yourself. But I would recommend against it because the results could haunt you for a long time.

We’ve reached out to Google and OpenAI for statements and will update this article with their responses.