There is a shirt going around this week that people keep describing as an invisibility cloak, and for once the description is close to accurate. Point a standard AI object detection camera at a crowd and it will draw a neat green box around every person in the frame. Put Simon Weckert’s Digital Camouflage shirt on one of those people and the box stops appearing. Not a smaller box, not a lower confidence score. Nothing. As far as the software is concerned, nobody is standing there.
It also happens to be one of the loudest garments you will ever see. Green, pink and orange, arranged in a pattern that looks like a Hawaiian shirt fed through a broken printer. Every human in a 50 foot radius will notice you. That contrast is the entire joke, and it is also the entire point.
The short version
- What it is: Digital Camouflage, a wearable art project by Berlin artist Simon Weckert
- What it beats: YOLO, the “You Only Look Once” family of object detectors that powers a huge share of real time camera analytics
- How: the pattern grabs the model’s attention and then collapses its confidence that a person is present
- Price: around $75 for the original, about $100 for the newer special edition cuts
- The disclaimer Weckert prints himself: it is “a work of art and a statement, not a certified anonymity tool”
What the shirt is actually attacking
Almost every camera system that claims to do things in real time is running some version of YOLO or a close relative. It is fast, open source, and good enough that it became the default. When a security camera tells an operator “three people in the loading bay,” a model in that family is usually the thing doing the counting.
These detectors do not recognize you. That distinction gets lost constantly. A person detector answers one question, which is whether a human shaped thing exists at these coordinates, and it attaches a confidence score to the answer. Face recognition, license plate reading and gait analysis are all separate systems that usually run downstream of detection. If nothing gets detected, the downstream systems never get handed anything to work with.
That is the crack Digital Camouflage wedges itself into. It does not disguise who you are. It argues, in a language the model finds persuasive, that you are not there.
Why the pattern looks the way it does
The obvious question is why nobody has designed a subtle version. The answer is that the pattern was not designed by a person in the usual sense. It was found by running an adversarial loop against the detector itself: generate a candidate texture, show it to the model, measure how confidently the model still finds a human, nudge the texture, repeat. Thousands of times.
Weckert has been clear that this is the only method available. A machine’s blind spots are not intuitive to us. You cannot look at a pattern and predict whether a neural network will choke on it, so the model has to tell you what it fails to see. The result is what the optimizer converged on, and the optimizer had no opinion about whether you would want to wear it to dinner.
There is a real research lineage here. Earlier attempts at this used a single printed patch, which worked beautifully when you stood square to the camera and fell apart the moment you turned. Academic work published at CVPR in 2022 solved that with what the authors called Adversarial Texture, generated by a method named TC-EGA, which produces a tileable pattern that keeps working across an entire garment and across viewing angles. Digital Camouflage is that idea taken out of a lab and put on a Cuban collar.
| Version | Cut | Colors | Price |
|---|---|---|---|
| v1 | Cuban collar short sleeve | Green, pink, orange | About $75 |
| v2.1 | Shorter polo style collar | Blue, black, light pink | About $100 |
| v2.2 | Crew neck | Blue, black, light pink | About $100 |
The list of things it does not do is longer
Weckert deserves credit for saying this himself rather than letting the marketing run. The shirt is a statement, not a certified anonymity tool, and the gap between those two is where most of the breathless coverage goes wrong.
A modern surveillance stack is not one system. It is a pile of them, and the shirt touches exactly one layer of the pile. Your phone is still broadcasting. Your card is still being swiped. Your car’s plate is still being read by an entirely different pipeline that never looks for a person shape at all. Automatic license plate readers were the subject of an ugly story earlier this month when a batch of Flock cameras turned out to be streaming to the open internet, and no shirt on earth changes what those cameras are pointed at.
Even inside the one layer it does reach, the guarantee is soft. Adversarial patterns are tuned against particular models. A camera running a different detector, a newer version, or a system that has been retrained on examples like this may not fall for it at all. Lighting, distance, motion blur and the angle of the fabric across your body all matter. This is a demonstration that works, not a product that always works.
Berlin, and then everywhere else
The project did not come out of nowhere. It was made in response to a pilot program bringing AI video surveillance to Kottbusser Tor, a busy square in Berlin where the debate over cameras has been running for years. Weckert has form here. He is the artist who dragged 99 phones around in a little red wagon and made Google Maps report a traffic jam on an empty street.
What made this one travel internationally is that Americans recognized the problem instantly, because the American version is further along. Automatic camera networks have spread through thousands of towns with very little public argument about it, and the arguments that do happen tend to arrive after something has gone wrong. In one Florida case, an officer ran his wife’s plate 717 times and the system asked him why, accepted his answer, and logged it.
Weckert also sells a parody of Palantir’s blue chore coat with a “NO” where the logo goes, which tells you how much of this is a fashion project and how much is an argument.
So should you buy one
If you want the practical answer: this will not make you meaningfully harder to track, and anyone selling that idea is overselling. Your phone gives away more in ten seconds than the shirt hides in a day. If your threat model involves a specific agency taking a specific interest in you, a loud shirt is the last thing you want, because being the only unreadable object in the frame is its own kind of signal.
If you want the honest answer: that is not really what it is for. It is a physical demonstration that machine perception has seams, that those seams can be found and worn in public, and that a system confidently reporting “no one there” can be wrong in a way nobody in the control room would ever notice. That is a genuinely useful thing to be able to point at, and it is a lot easier to explain while wearing it than in a policy document.
The counterpoint worth sitting with is that these seams close. Every published attack becomes training data eventually. Digital Camouflage works today against the detector it was tuned on, and the next generation of models will be a little more resistant because this one exists. That is the ordinary rhythm of it, and it also means the shirt is best understood as a snapshot of a moment rather than a solution. Tracking has been quietly moving toward signals that clothing cannot touch anyway, which is roughly the lesson from the discovery that AliExpress was playing silent tones through people’s speakers to work out who they were.
