This work looks at the current abilities of artificial intelligence, specifically in image recognition, by pushing it to the point of failure. I photographed very simple human objects and presented each image to Microsoft Copilot, asking it to identify what is in the image. Then, the images are progressively reduced in resolution until the AI is wrong in what it identifies as the subject of the photo. Each image in the exhibition is a combination of that final image where it failed, alongside its incorrect description. These failures aren’t just technical errors or uncertainty, but odd moments of AI overconfidence in giving out any answer, even if it isn't truthful. On a broader scale, I feel like this hits at the issue of misinformation through AI from people taking its confidently wrong answers as true.

Since this piece is very much a conceptual work, the outcome is unpredictable and out of my control, which leads to various results. I attached each image with the same prompt, as well as using a new chat every time so the AI can't rely on its previous guesses to inform its current guess.

The idea for the project stemmed from CAPTCHA tests, which try to tell humans apart from robots. I feel certain that the AI models of today would have no problem identifying the subjects of those images, so I wanted to find out exactly where that limit was. By stripping images down to blocks of pixels, this piece tests whether humans can actually be better at speaking a computer's native language.