Fooling AI cameras at Kottbuser Tor? Artist Simon Weckert makes himself invisible with a special camouflage shirt.
Berlin-based artist Simon Weckert has designed a "Camouflage Shirt" in response to a pilot project for AI-supported video surveillance at Kottbusser Tor subway station. Behind the garment lies a profound critique of modern surveillance algorithms.
AI-Generated Patterns to Fight AI
Even if it sounds paradoxical: the artist used artificial intelligence itself to design this special shirt. He developed a pattern featuring saturated color gradients and overlapping shapes in a so-called "Adversarial Loop" (a generative testing process). Specifically, he had psychedelic patterns generated and tested them against "YOLO" (You Only Look Once), an open-source real-time object detection system. The system indicated the statistical probability with which it detected a human silhouette. Based on these values, Weckert adjusted the pattern until the final iteration completely confused the AI.
The Blind Spot of Algorithms
The underlying algorithms of image recognition operate exclusively on probabilities. AIs do not understand the world the way humans do; instead, they take educated guesses, based on mathematical probabilities. For surveillance cameras, this means they do not know what a human actually is - they merely match visual features against millions of training images.
Weckert exploits this functionality. His pattern distracts the AI with targeted edges, contrasts, and texture gradients before the algorithm can even look for the typical silhouettes of heads and shoulders. As a result, the detector is no longer able to assemble the individual body parts into a human figure.
Whether the shirt delivers on its promise in everyday life remains to be seen. While the "YOLO" system can be tricked, independent tests under real-world conditions, such as with the cameras at Kottbusser Tor, are still pending.
Even though it comes in a functional guise, the "Camouflage Shirt" is far more than just a tool for AI distraction. The project impressively demonstrates how error-prone algorithms can be, while simultaneously raising an uncomfortable question: How much power should be wielded by systems that can be blinded by nothing more than a colorful pattern?
What do you think? Should we trust systems that can be tricked so easily - or does the added security outweigh the risks for you? Join the discussion in the comments!
