Picture the citation landing in your mailbox: a seatbelt violation, photographed, timestamped, officially issued, for the empty passenger seat next to you. No passenger. No unbuckled human. Just a chair. That’s a real category of ticket that came out of Greece’s experiment in letting artificial intelligence run traffic enforcement around Athens, and it’s the single cleanest illustration of why the pilot went sideways. A machine that can’t reliably tell whether a person exists is not ready to decide whether that person broke the law.
Why a Camera Accuses a Seat of a Crime
Start with how a camera manages to accuse a seat of a crime, because the failure is more instructive than funny. These systems don’t “see” a passenger and check for a belt the way a person would. They pattern-match pixels, and the pattern-matching turned out to be shockingly literal. Empty front seats got flagged for riding beltless. Drivers wearing dark shirts got cited because the fabric blended into the strap and the software lost track of where the belt ended and the person began. Reach for the gearshift or take a pull off a vape and the system could read the motion as a phone against your ear. Shadows and dark clothing threw it off, so did lighting. Greek newspaper Ta Nea documented the pattern: the camera wasn’t catching violations so much as inventing them wherever contrast confused it.
10 Car Accessories You’ll Actually Use
Then there’s the problem the cameras can’t solve at any resolution, because it isn’t about image quality, it’s about meaning. A camera logs a car crossing into a bus or emergency lane and has no idea why. A driver easing aside to let an ambulance pass looks identical to a driver cutting the lane illegally, at least to a system with no concept of context for why something wrong actually happened.
The Scoreboard Is Brutal
Now the numbers, which is where the “AI will modernize enforcement” pitch really falls apart. Across April and May, the Athens cameras generated roughly 13,000 citations. Officers only got through 5,500 of them, and of that reviewed batch, just 400 held up. The other 5,100 were thrown out. Worth pausing on the ratio: for every ticket that survived scrutiny, roughly a dozen did not. A tool sold on the promise of doing more with fewer officers produced a mountain of work that officers then had to dig back out of by hand.
The rejected tickets aren’t all the same kind of wrong, though, and the distinction matters. About 3,800 of them were speeding citations, and those weren’t really a software malfunction at all. They relied on average-speed calculations that Greece currently has no legal framework to enforce, so they were dead on arrival for a legislative reason, not a technical one. Strip those out and you’re left with roughly 1,300 tickets covering phones, seatbelts, and similar offenses. That pile is the genuine machine-vision failure, the empty seats, the vape-as-phone, the shirt-as-strap. Lumping the two together makes the AI look more broken than it is on speed and, crucially, less broken than it is on the thing it was actually supposed to be good at: watching people inside the car.
A caveat belongs here, because it changes how hard you can lean on any single figure. These review numbers come from leaked data rather than an official police release, and Greek outlets have already garbled the failure rates in a few retellings, turning a messy story into a chaotic one. Treat the exact percentages as directional, not gospel.
The Government Tells a Tidier Story
What isn’t in doubt is the daylight between the government’s version and the picture the data paints. The Ministry of Infrastructure and Transport has been selling the program as a success, and on national broadcaster ERT it put up its own tidier set of figures: 2,453 citations finalized between late March and late May, 420 formal objections from drivers, and only 52 of those objections upheld, about 12 percent, most tied to medical emergencies. Officials point at that low 12 percent as proof the system is accurate.
But a low objection rate doesn’t measure whether the machine is good, it measures whether the tickets that reached drivers had already been cleaned up first. The phantom citations mostly never make it to a mailbox, because a human catches them upstream. In other words, the number the government is bragging about is a product of the very manual labor the AI was supposed to eliminate. The system looks precise from the outside precisely because people are doing the discriminating the software can’t.
It’s Eight Cameras
And the scale of the whole thing punctures the futuristic framing. This is not a city blanketed in all-seeing machines. Kimon Logothetis, speaking to YouTuber Vasilis Saribalidis, put the real count at just eight AI cameras in the region, the only units watching speed, red lights, helmets, seatbelts, and phones all at once. The plan to expand dramatically has stalled: the procurement competition to supply 1,000 AI cameras across Attica reportedly collapsed after losing bidders filed legal appeals, and while it may restart someday, the separate rollout that’s actually happening is 388 conventional, non-AI cameras by mid-July, restricted to red-light running.
So the grand automated-enforcement future currently amounts to eight working cameras, a procurement deal in litigation, and a backlog of invented tickets that human officers have to babysit. There’s a real lesson in here for every American city eyeing AI cameras as a budget-friendly force multiplier: a system that can’t distinguish a person from an empty seat doesn’t reduce the human workload, it relocates it, from writing tickets to un-writing them. The empty seat that got a seatbelt ticket isn’t a funny outlier. It’s the whole pilot in one frame, and the burden of proof sits squarely with the people who insisted the machines were ready.
Sources: Ta Nea; ERT (national broadcaster); Kimon Logothetis via Vasilis Saribalidis. Figures on rejected citations are drawn from leaked review data and should be treated as preliminary.