The $88 shirt that made an AI surveillance camera stop seeing a person

At a Berlin surveillance camera, the trick was not sneaking around or covering a face. It was wearing a bright Hawaiian-style shirt so visually noisy that the software could not reliably recognise the person inside it. That shirt, called Digital Camouflage, was designed by German artist Simon Weckert to exploit a weakness in AI vision systems by turning their own pattern-matching habits against them.

A shirt built to disappear in full view

Digital Camouflage does not look subtle. The roughly $88 shirt is covered with intense pink, orange and green forms arranged in a chaotic pattern. To human eyes, it is the sort of clothing that instantly draws attention. To the AI systems Weckert tested, the effect was the opposite.

The project was made in response to Berlin’s August 2026 pilot rollout of automated AI surveillance at Kottbusser Tor, where software scans the behaviour of people moving through the area. Weckert’s answer was to create a shirt the system could not properly read.

He successfully tested it in June 2026 in front of an operational surveillance camera at Kottbusser Tor. The central claim behind the design is simple but strange: the wearer may still be perfectly visible to other people, while the AI camera struggles to recognise that there is a person there at all.

That contrast is what makes the shirt so arresting. It is not camouflage in the usual sense. Nothing about it blends into a background. Instead, it appears to work by overwhelming the specific visual shortcuts that machine vision relies on when it tries to identify a human body.

Why a camera can miss someone standing in front of it

Weckert’s explanation starts with a basic point about how AI image systems work. They do not understand a person in the way a human observer does. Instead, they learn statistical cues from huge sets of training images.

In the early stages, that means reading edges, contrast and texture. Deeper into the system, it means looking for familiar part-like arrangements, such as the shape formed by a head above shoulders, the proportions of limbs around a torso, and the continuous outline of a body against whatever is behind it.

The shirt is designed to attack that process at the beginning. The target is the stage where detection software first tries to place a box around a human figure, before moving on to systems such as facial recognition or gait recognition. If the software cannot confidently decide that it is looking at a person, the rest of the chain becomes much harder to start.

Its bright colours are only part of the trick. The overlapping shapes also interfere with the body’s outline, making it difficult for the software to connect head, torso and arms into one coherent figure. In Weckert’s account, what the machine reads as a person is really a bundle of recurring visual probabilities, and that bundle can be disrupted.

The pattern was taught by the same kind of technology it targets

One of the oddest parts of the project is that the shirt was created with AI as well. Weckert developed the pattern through what he described as an adversarial loop.

The process involved generating a possible design, showing it to the detector, measuring how strongly the software still identified a person, then adjusting the pattern and trying again. He repeated that cycle until the detector’s confidence dropped away. The system he used for testing was the open-source object detector YOLO.

That trial-and-error approach mattered because the blind spots of machine vision are not necessarily obvious to people looking at an image. A human designer could not simply guess what would confuse the software most effectively by eye. The detector had to be part of the design process, teaching the shirt what its own weak points were.

Weckert also said the shirt was made to hold up outside controlled demonstrations. It was designed with real fabric in mind, including folds in the cloth, changing light and different viewing angles. That practical element is important because a pattern that works only in a perfect studio setup would say much less about everyday surveillance systems.

More than a fashion oddity

Weckert framed the project as a challenge to confidence in automated surveillance, not as a tool aimed at the police. His argument was that if a simple piece of fabric can undo a system’s reading, that raises a larger question about how much trust the system deserves.

That line of thought fits with his earlier work. In 2020, he became widely known for pulling a handcart carrying 99 smartphones through empty streets in Berlin, which caused Google Maps to display false traffic jams along his route. That stunt, like Digital Camouflage, focused on exposing the gap between what a digital system reports and what is actually happening in front of it.

The shirt may look like an ugly holiday top at first glance, but its purpose is much more specific than fashion. It was built as a demonstration that the intelligence inside an AI camera is still a fragile one, dependent on patterns it has learned rather than any real understanding of the people it is supposed to see.

That is what gives the project its unusual power. The wearer is not hidden in darkness, behind a mask or outside the frame. He is standing there in a shirt loud enough for any human to notice, while the machine can be left trying to decide whether there is a person there at all.