Pilot use of AI traffic cameras in Athens achieved a tiny 3% approval rate. Human officers reviewed 5,500 out of 13,000 cases and only approved 400. Besides legislative issues, AI can’t properly identify phone and seatbelt use. The pitch for letting artificial intelligence handle road safety and traffic enforcement reads well on a slide deck. Fewer cops on the roads, faster processing, citations that write themselves. Greece put that pitch to the test, and the pilot phase has already pulled the thread loose. Despite promising early signs, the handful of AI cameras switched on around Athens proved spectacularly bad at their one job, with the humans behind them validating only a sliver of what the system flagged. According to Greek media, the AI cameras generated roughly 13,000 citations across April and May, but human personnel managed to review only 5,500 of them. Of those reviewed, just 400 were judged valid. The other 5,100 were thrown out, split between 3,800 speeding violations and 1,300 covering mobile phone use, seatbelt non-compliance, and similar offenses. The two piles were rejected for completely different reasons. The 3,800 speeding citations were dropped automatically because they relied on average-speed calculations, a method Greece currently has no legislative framework to support, so that batch isn’t a software failure at all. The Software Couldn’t Tell A Vape From A Text The phone and seatbelt detection is where the technology actually struggled, accounting for the 1,300 rejected AI citations. Police officers found that the software routinely mistook dark objects for smartphones and got tripped up by unrelated driver gestures, things like taking a vape hit or simply changing gears. Photo: Ministry of Digital Governance The AI cameras struggled with lighting too, misinterpreting shadows and dark clothing, reports local newspaper Ta Nea. In several cases, the cameras flagged non-existent front-seat passengers for failing to wear a seatbelt, triggering automated notices for empty seats. In others, drivers were fined simply because their dark shirts blended into the seatbelt strap, confusing the camera’s visual sensors. More: Just Eight AI Traffic Cameras Caught 29,000 Offenders In Just a Few Weeks Then there’s context, which the cameras have none of. They log every car that crosses into an emergency or bus lane without knowing why it happened. A driver pulling aside for an ambulance reads the same as a driver breaking the rule, and so does one waving through on the hand signals of a traffic cop. One caveat carries the rest. These numbers come from leaked data, not official police sources, and Greek media have already fallen into the trap of misreading the failure rates, turning a messy story into a chaotic one. The Government Tells A Different Story Photo: Thanos Pappas The Ministry of Infrastructure and Transport is publicly spinning the program as a flawless success. According to the official broadcast from national TV channel ERT, authorities have finalized and confirmed 2,453 digital citations between late March and late May. Of those, disgruntled drivers filed 420 formal objections, but the state accepted only 52 of them, roughly 12 percent, mostly for medical emergencies. More: AI Cameras Caught $12 Million Worth Of Illegal Supercars, One Toll At A Time The government is leaning on that low acceptance rate to praise the system’s reliability. The catch is that human police officers still have to act as a secondary filter, manually scrubbing thousands of phantom violations before the real tickets reach the right drivers. The Real Number Of AI Cameras Speaking to YouTuber Vasilis Saribalidis, Kimon Logothetis clarified that the Ministry of Infrastructure and Transport currently operates only 8 AI cameras in the region. Those are the only ones that monitor speed, red light, helmet, seatbelt, and smartphone use all at once. It gets worse. The procurement competition for the supply of 1,000 AI cameras on the roads of Attica has reportedly collapsed after legal appeals from bidders, though it will likely restart down the line. Separate regional projects plan to deliver 388 standard, non-AI cameras by mid-July, limited strictly to catching red light violations. In any case, it is clear that both the humans and the algorithm must face a steep learning curve in order for the investment to make sense. Hopefully, the process will improve over time, leaving lawful citizens at peace. Photo: Thanos Pappas