There was evidence of what researchers call ‘AI hallucinations’ from both tools. In one instance, the BING AI provided details that were impossible to find without EXIF data, that had already been scrubbed from the image provided. It had conjured up false camera model and exposure settings, a time the image was taken and that it had the flash off. https://www.nytimes.com/2023/03/29/technology/ai-chatbots-hallucinations.html
However the tool still failed to geolocate without significant prompting. It misplaced landmarks, placing them near parks far away and in fictitious surroundings of a fake space and science centre.
We also tested Google’s new chat bot ‘BARD’ with an identical set of images. This tool required less prompts on average, and provided Dennis Kovtun with a list of landmarks to select from in a process of elimination.
Bellingcat observed how the BING AI seemed to mimic the process of an open-source researcher in geolocating, noting key structures in the images, but failed to identify images successfully on its own.
While many AIs can generate images, most cannot analyse them and are therefore useless for the purposes of direct geolocation. BING and BARD (Google’s new AI) are different and can analyse images.
When provided with this second prompt, the BING’s AI correctly identified the location: “This is not a bridge. The image appears to have been taken on a public square, with the building with a pyramid-shaped roof in the background.”
When given an image of recognisable landmark Edmonton City Hall, the BING AI appeared to focus on the skyscrapers in the background and the water, wrongfully identifying it as a bridge in the same city.
Bellingcat put two AI bots to the test, BING’s AI tool and Google’s new BARD. We provided them with identical images and the name of the city featured and date each image was taken.