What if a shirt could make AI-powered cameras fail to recognise you as a person? German designer Simon Weckert is exploring that possibility through a colourful button-down shirt called Digital Camouflage, designed to interfere with AI-based video surveillance systems.
At first glance, the shirt appears to be an abstract fashion piece featuring bright green, pink and orange shapes. However, the unusual design serves a specific purpose: disrupting the visual cues AI systems use to identify humans.
In one demonstration shared by Weckert, an image-recognition system detects a person and places a green box around them with the label “PERSON”. When Weckert positions the shirt in front of his body, both the box and the label disappear. Once the shirt is moved away, the system identifies the person again.
Why use a shirt to trick AI?
In cities such as Berlin, authorities are preparing to introduce AI-powered cameras capable of recognising objects and monitoring activity in public areas. Weckert created Digital Camouflage to explore how effectively such AI surveillance systems can identify people and whether their detection capabilities can be disrupted.
How does the shirt confuse AI?
The concept is based on how AI vision systems identify people. Unlike humans, these systems do not truly understand what a person looks like. Instead, they are trained using millions of images and learn to identify recurring visual patterns.
When a camera captures a person, the AI searches for cues such as body shape, outlines, edges, colours and textures that are commonly associated with humans.
Digital Camouflage is designed to interfere with those cues. Its bright colour combinations and overlapping shapes can attract the system's attention while disrupting the body's outline. As a result, the AI may struggle to connect the different visual elements and determine that they belong to a single person.
In simple terms, the shirt overwhelms the visual signals an AI system normally relies on to detect a human, making it harder for the software to recognise the person wearing it.
Weckert did not simply rely on trial and error to develop the pattern. He used AI itself to create and refine the design through what he describes as an “adversarial loop”. In this process, AI generates a pattern, tests it against an object-detection system and measures how confidently the system can still identify a person. The design is then modified and tested repeatedly until the detection system has increasing difficulty recognising the individual.
Weckert tested Digital Camouflage against YOLO, an open-source AI system capable of detecting objects in real time and comparable to technology used in various surveillance applications.
The project forms part of a broader field of research into adversarial clothing, in which garments are designed to interfere with computer-vision systems. Other experiments have similarly used specially created patterns to cause AI systems to misidentify objects or fail to detect people altogether.
However, Weckert does not claim that his shirt can fool every CCTV system or provide guaranteed anonymity. Instead, the project is intended to demonstrate a potential weakness in AI-based vision technology and encourage discussion about the reliability of automated surveillance systems.
