ChatGPT's latest image bug is raising eyebrows: Check details here

Artificial intelligence models are supposed to understand context, identify mistakes and increasingly behave like digital assistants. But a strange image restoration glitch affecting both ChatGPT and Google's Gemini suggests today's AI systems still struggle with a much simpler task: recognising when there is nothing to work with.

Users discovered that asking the chatbots to "restore" an image without providing one could trigger the creation of entirely new pictures. The outputs ranged from bizarre to deeply unsettling, despite the absence of any original image to reconstruct.

ChatGPT and Gemini create images from nonexistent photos

According to Android Authority, the problem initially came into focus when Kris Kashtanova, an artist and researcher in AI technology, pointed out the phenomenon on the X platform. Other users provided evidence of both AI models creating weird scenarios when asked to retrieve lost and damaged images.

During tests conducted by me, ChatGPT sometimes failed to comply with my request when there was no image attached. Nevertheless, even the image consisting solely of blank white space caused the AI to create completely irrelevant images.

Gemini acted slightly differently. When an image was attached to the request, it returned a blank image, but when there was none, it generated random images.

However, perhaps the most illuminating aspect was what took place after the images were produced.

On being asked, ChatGPT confessed to producing a "hallucinated scene" rather than conducting a restoration. It further confessed that it should have told the user that there was no image data to begin with.

Gemini provided a comparable explanation, having realised that its output might turn out to be disturbing in nature.

All of this underscored one major shortcoming of generative artificial intelligence models: they usually explain their errors only post hoc.