BS Report
Polymarket's AI-invisibility shirt post
JUST IN: German technologist develops a "digital camouflage" shirt designed to make wearers effectively invisible to AI surveillance cameras. · @Polymarket, X, 2026-08-29 · 3,106,352 views · 9,575 likes · 719 reposts · 322 replies
2026-09-04 · bullshit-detector 0.14.0
the shirt is real and the wording is the artist's own; the missing sentence is that it was only ever shown to one open-source model
Tally: 6 claims extracted, 6 individually source-checked — 3 confirmed, 1 plausible, 2 misleading.
Ambiguous: 0 claims dropped before verification. Nineteen words, every one of them resolving to a single reading. The post's problem is not that it could mean several things; it is that one word, "cameras", is a category where the evidence supports a single open-source model.
What it says (neutral summary)
A 19-word post, no link and no image, reporting that a German technologist has developed a "digital camouflage" shirt intended to make the wearer effectively invisible to AI surveillance cameras. It carries a breaking-news marker. It names no person, no city, no camera system and no test, and links to nothing a reader could check.
Load-bearing claims
The ones the thesis dies without. Verify all of them.
| # | Claim | Type | Verdict | Evidence |
|---|---|---|---|---|
| 1 | A garment called "digital camouflage" exists as a real, purchasable shirt [X post, 2026-08-29] — "develops a “digital camouflage” shirt" | factual | ✅ confirmed | [4 URLs → 2 origins, judged] Origins: Weckert's own project page and 404 Media's first-hand visit. The artefact is real and the name is exact: the project is titled Digital Camouflage, listed on Weckert's site as "Installation / Wearable, 2025-ongoing" [tier 4: the maker on his own work], with a shop at digitalcamouflage.org selling it at about $88 / GBP 65 [tier 3]. The strongest evidence is not the maker's: 404 Media's Jason Koebler visited in person on 2026-09-01 and wrote that he stood in front of a live detector, was boxed as "PERSON", held the shirt up, and "The box and the word 'PERSON' disappears" [tier 2: a reporter's own observation, not a press release]. A button-down shirt named Digital Camouflage exists and does something. (source) |
| 2 | The shirt is designed with the intent of making wearers effectively invisible to AI camera systems [X post, 2026-08-29] — "designed to make wearers effectively invisible to AI surveillance cameras" | factual | ✅ confirmed | [2 URLs → 1 origin, judged] 2 URLs -> 1 origin: the maker himself Give the post this one squarely. "Effectively invisible" is not the aggregator's embellishment - it is a near-transcription of Weckert's own sentence: "The wearer becomes, in effect, invisible to the machine - present to every human eye, absent to the model" [tier 4: the maker on his own work]. His YouTube upload of 2026-08-19 is captioned "A garment that makes you invisible to AI surveillance." As a statement about design intent this is accurate to the source, and Polymarket kept the hedge "designed to" that Dezeen's own headline dropped. (source) |
| 3 | The shirt makes wearers effectively invisible to AI surveillance cameras as a class, i.e. to the automated camera systems actually deployed in public space [X post, 2026-08-29] | factual | 🟠 misleading | [5 URLs → 4 origins, judged] Origins: Weckert's own disclaimer, the arXiv paper he cites, the InvisDefense result, and netzpolitik on the Berlin system. One system was tested, and it was not a surveillance camera. Every account agrees the only detector involved is YOLO, an open-source object detector: "Demonstrations use a generic open-source object-detection system (YOLO). The work makes no claim about any specific government system" [tier 4, and the maker is arguing against his own product here]. The same page states flatly: "No pattern can promise to defeat every system under every condition" and "it is not a promise of anonymity." No test data, no success rate and no independent evaluation has been published for this shirt at all. The nearest measured comparison runs the other way: InvisDefense, the Wuhan University coat sold for about $71 in December 2022 with an AAAI 2023 paper behind it, cut campus pedestrian-detection accuracy by 57% - a large effect, and nothing like invisibility [tier 2]. Even the state of the art Weckert cites degrades under defence: Zhang et al., arXiv:2510.17322 (2025-10-20) report 96.06% attack success against an undefended Faster R-CNN but only "over 64.84%" against nine defended models [tier 1], i.e. roughly a third of frames still detect you once the operator hardens the detector. And Berlin's actual system is not YOLO: netzpolitik identified the vendor as Adesso SE building a behaviour scanner for the police [tier 2], which also retires Weckert's line that the manufacturer is secret. Rated misleading rather than false because the effect is real, demonstrated on camera and witnessed by a reporter - what is missing is that it was shown against one open-source model, and that the artist says so himself. (source) |
| 4 | This is breaking news, i.e. the shirt and the capability behind it are new as of [2026-08-29] [X post] — "JUST IN: German technologist develops" | factual | 🟠 misleading | [5 URLs → 4 origins, judged] Origins: Dezeen's dated original, Weckert's own site and video, the 2019 KU Leuven paper, and Dezeen's 2023 Cap_able coverage. Nothing broke on 2026-08-29. Dezeen published the interview this post relays on 2026-08-25, four days earlier; Weckert's demo video went up 2026-08-19; his own site dates the work "2025-ongoing." The capability is older still. Thys, Van Ranst and Goedeme, "Fooling Automated Surveillance Cameras: Adversarial Patches to Attack Person Detection," CVPRW 2019 [tier 1], is the founding result, and its authors closed by saying they believed the technique could "design a T-shirt print that could make a person virtually invisible for automatic surveillance cameras" - the tweet's claim, written seven years earlier. Wearable versions have been on sale since: InvisDefense (Wuhan University, ~$71, December 2022) and Cap_able's Manifesto collection (Italy, January 2023, ~$400 knitwear, also tested against YOLO, covered by Dezeen itself at the time) [tier 3]. Weckert's three cited papers are all other people's: Hu et al. 2022 (arXiv:2203.03373), Zhang et al. October 2025, Zhou et al. November 2025. What is new here is a German artwork and a product launch, not a technique. (source) |
Incidental claims
Supporting detail. Wrong here is embarrassing, not fatal.
| # | Claim | Type | Verdict | Evidence |
|---|---|---|---|---|
| 5 | The developer of the shirt is German [X post, 2026-08-29] — "German technologist develops" | factual | ✅ confirmed | [2 URLs → 2 origins, judged] Origins: Weckert's own CV page and PetaPixel's independent description. Simon Weckert was born in Karl-Marx-Stadt, Germany, studied New Media Art at the Berlin University of the Arts and works in Berlin [tier 4: his own CV]. PetaPixel independently describes him as "a digital artist from Berlin, Germany" [tier 3]. The one word in the post nobody can argue with. (source) |
| 6 | The developer is a technologist [X post, 2026-08-29] — "German technologist develops" | factual | 🟡 plausible | [3 URLs → 3 origins, judged] Origins: his own about page, Dezeen and Dexerto. Not one source uses the word. Dezeen calls him a "designer", PetaPixel "a digital artist", Dexerto "German artist", and his own page offers "a modern-day digital wizard and mischievous tinkerer" whose work is "part tech, part art, and part cheeky social critique" [tier 4]. He does have real technical standing - member of the Chaos Computer Club since 2020, AI Researcher for Digital Curation Technologies at Art+Com on QURATOR.AI from 2020 to 2022, guest professor at HFG Karlsruhe - so the label is not invented. It is a promotion: the project is filed on his own site under "Installation / Wearable", and calling an artwork's author a technologist moves it from gallery to laboratory, which is exactly the move the rest of the post depends on. (source) |
Tally: 6 claims extracted, 6 individually source-checked — 3 confirmed, 1 plausible, 2 misleading.
Ambiguous: 0 claims dropped before verification. Nineteen words, every one of them resolving to a single reading. The post's problem is not that it could mean several things; it is that one word, "cameras", is a category where the evidence supports a single open-source model.
Unreachable: 3 sources - 1 blocked (Fast Company, 403 then a JS wall, for claim 3), 1 paywalled (404 Media beyond the opening section, for claim 1), 1 empty (the digitalcamouflage.org shop, whose price is JS-loaded and never rendered, for claim 1). Named in the rows that needed them. Dezeen refused a direct fetch and was read through a text proxy instead, so its body text is quoted at one remove.
The chain of citation, nailed down. Simon Weckert uploaded his demo video on 2026-08-19. Dezeen's Tom Ravenscroft published the interview every later piece quotes on 2026-08-25. PetaPixel rewrote it on 2026-08-27, crediting Dezeen in the body. Polymarket posted on 2026-08-29 at 23:04 UTC. Dexerto followed on 2026-08-30, Oddity Central and Famous Campaigns on 2026-08-31, and 404 Media published the only piece involving an independent visit on 2026-09-01. So: one artist, one Dezeen interview, one YOLO demo, a dozen URLs, and exactly one journalist who has seen the thing work in person - after the post went up.
Hype signals observed
- A breaking-news marker on a four-day-old design story. "JUST IN" for a Dezeen interview published 2026-08-25 and a video posted 2026-08-19.
- Occupation upgrade. "German technologist develops" for a new-media artist whose own site files the work under Installation / Wearable and describes him as a digital wizard and tinkerer. Every outlet covering it says artist or designer.
- Every disclaimer dropped. The artist's page carries three sentences the post does not: that no pattern can defeat every system under every condition, that it is not a promise of anonymity, and that the demonstrations use a generic open-source detector and make no claim about any government system. The post has room for none of them and does not link the page that has them.
- The most consequential word is the vaguest. "AI surveillance cameras" is a category; what was tested is one open-source object detector. The post never narrows it, so the reader supplies the widest reading.
- Zero citation surface. 19 words, no link, no image, no name. A reader who wants to check any of this has to find the artist themselves.
Not observed, and worth saying: no prompt injection, no hidden text, no fabricated institutions, no invented numbers. The post contains no figure at all. The failure is compression, not fabrication - and unusually for this genre, the loudest phrase in it was written by the person being reported on.
Incentive analysis
Polymarket is a prediction market that earns from trading volume and new accounts. Its X account runs a high-cadence "JUST IN" wire that functions as top-of-funnel: this post links to no tradeable market and does not need to. What is being sold is the habit of treating the account as a live news source. That incentive does not care whether the shirt works; it cares whether the sentence gets reposted, and it selects for the reading with the widest implication. Note what compression did here: the post kept the artist's hedge "designed to" while dropping the three disclaimers sitting on the same page, which is the exact edit that maximises reach at minimum liability.
Weckert's incentive is more direct and more sympathetic, and it should be named anyway. He sells the shirt for about $88 through his own shop, with a stated share of proceeds going to digital civil rights organisations. The framing that makes it a protest object - a system that cannot be publicly audited, undone by a piece of fabric - is also the framing that sells it. He is unusually honest about the limits for someone in that position, which is why the disclaimers exist at all. They are just not on X.
Bottom line
Most of this post is true, and the part people assume is marketing spin turns out to be a quotation. The shirt exists, costs about $88, is named Digital Camouflage, and was made by a German artist in Berlin. "Effectively invisible" is not Polymarket's flourish: Weckert's own page says the wearer "becomes, in effect, invisible to the machine", and his video is captioned "A garment that makes you invisible to AI surveillance." 404 Media's Jason Koebler stood in front of a running detector on 2026-09-01, held the shirt up, and watched the PERSON box vanish and reappear. That is a real effect, independently witnessed.
What the post removes is the scope. The only system anyone has shown this working against is YOLO, an open-source object detector Weckert ran in a loop until its confidence collapsed. No success rate has been published for the shirt, no independent evaluation exists, and no test against a deployed system has been done or claimed - Weckert says so himself, on the page the post is relaying: the demonstrations use a generic open-source detector, and "No pattern can promise to defeat every system under every condition." Berlin's actual behaviour scanner at Kottbusser Tor is built by Adesso SE and has never been tested against this shirt by anyone.
The measured evidence that does exist points at a real but partial effect. InvisDefense, the Wuhan University coat from December 2022 with an AAAI 2023 paper behind it, cut pedestrian detection accuracy by 57% in campus tests - large, and nothing like invisibility. The 2025 paper Weckert cites reports 96% attack success against an undefended detector and only about 65% against nine defended ones, meaning a third of frames still find you as soon as the operator hardens the model. Adversarial clothing is a probability reduction, not a cloak.
And none of it is new. Thys, Van Ranst and Goedeme published the founding result at CVPR workshops in 2019 and ended it by predicting a T-shirt print that would make a person "virtually invisible for automatic surveillance cameras" - the tweet's headline, seven years early. Wearable products have been on sale since InvisDefense in 2022 and Cap_able's Manifesto collection in January 2023, which was also tested against YOLO and also covered by Dezeen. Weckert's own page credits three papers, all by other researchers.
The honest version of the post is one clause longer: a German artist has made a shirt that defeats an open-source person detector, and is selling it as a protest against surveillance nobody outside the police can test. That version is still a good story. The version that went to 3.1 million people invites a reader to believe they can buy their way out of being watched, which the artist explicitly says they cannot.
What a hostile reader would hit first
- The artist's own website refutes the headline. One click reaches "No pattern can promise to defeat every system under every condition" and "it is not a promise of anonymity", on the same page as the sentence the post borrowed. Being contradicted by your own source is the worst position in this list.
- One model, and it is not a surveillance system. YOLO is a free object detector anyone can pip-install. Nothing here has been shown against a deployed camera system, and Berlin's is built by a named vendor, Adesso SE, that has never seen this shirt.
- "JUST IN" on a story Dezeen ran four days earlier. Cheap to check, and the demo video predates the post by ten days. Fixable by deleting two words.
- The technique is from 2019 and the products are from 2022. Thys et al. predicted exactly this T-shirt at CVPR 2019; InvisDefense sold one for $71 in 2022; Cap_able sold knitwear tested against the same YOLO in 2023. "Develops" implies a first.
- "Technologist" is nobody's word. Dezeen says designer, PetaPixel says digital artist, Dexerto says artist, and Weckert says digital wizard and tinkerer. It lands slowest but reframes the whole post, because an artwork and an engineering result carry different burdens of proof.
run: 7m15s, searches 8, tools 33, coverage 0, per claim 72s, model claude-opus-5, effort high