Powerful images of the tornado that struck Windsor-Essex on Sept. 3 quickly spread across social media — but not everything posted online showed what actually happened. Fabricated tornado images were also circulating. Devin Pastorius of advertising agency Narwhal Creative said creating a convincing fake requires little technical knowledge. “It’s as easy as putting a prompt into a box,” Pastorius said. “If you can write a status update on social media, you can ask AI to do something for you.” Pastorius said artificial images can sometimes be identified by studying details that do not make physical sense, such as inconsistent damage, misplaced debris or elements that appear unusually perfect. However, those clues are becoming more difficult to spot as the technology improves. “For the longest time, it was always, ‘Pics or it didn’t happen,’” Pastorius said. “Photos and video were kind of like beacons of truth, and that’s just not the case anymore.” Adam Skinner, director and CEO of Instant Weather, said suspected AI images tend to surface during larger weather events, particularly when tornadoes are involved. In many cases, he said, someone takes a real photograph without a funnel cloud and uses an AI platform to add one. Skinner’s team compares suspicious submissions with radar information, weather conditions and reports received from its community. Images may also be reviewed using AI-detection tools. He said the surrounding context can be just as revealing as the image itself. If a photograph supposedly shows a large tornado touching down in a populated area, there would likely be several witnesses and multiple images — not just one isolated post. “We have such a large community in all these different places, and we would easily have many, many photos for that type of thing,” Skinner said. Instant Weather has also introduced a reporting feature within its free app to help authenticate submissions. Images captured through the app receive an Instant Weather watermark and include information showing where the photograph was taken and the direction the camera was facing. The submissions are then automatically screened through detection tools. “If it comes right through our app, we’re very confident that it was taken by a community member and is an actual, legitimate report,” Skinner said. He estimates the organization encounters one or two suspected AI images during some significant weather events, although many events produce none. Both experts recommend pausing before sharing a dramatic image, tracing it to its original source and checking whether credible outlets or other witnesses have documented the same event. The poster’s previous content may offer another clue. Opening the comments can also reveal whether other users have identified the image as AI or traced it back to an earlier source. Metadata and image dimensions can sometimes provide additional information, but Pastorius cautioned that metadata can be removed or altered and may be difficult for the average person to interpret. Ultimately, he said, people must adjust to an online landscape in which seeing is no longer necessarily believing. “You kind of have to question everything — and that sucks,” Pastorius said. “Even if it is in a photo, even if it is in a video.”
