AI traffic cameras introduced in Greece still have a 90% error rate

Greece's pilot AI traffic enforcement cameras have been found to still produce errors in nearly 90% of cases. Not only do they struggle with nighttime recognition, but they also cannot yet distinguish between e-cigarettes and smartphones.

AI is now being rapidly integrated into public infrastructure, including enforcement cameras. South Korea has already implemented this system, where AI is used to analyse footage from existing cameras at a data level, and its recognition rate is reportedly quite high. However, Greece went a step further by piloting so-called AI cameras that directly identify violations in real-time. But in conclusion, this pilot project in Greece has essentially failed, as the error rate reached as high as 90 to 95%.

Over the past two months, the AI traffic cameras installed in Athens issued nearly 13,000 traffic violation tickets. The scope of enforcement was quite broad, ranging from speeding to not wearing a seatbelt and using a smartphone while driving. The system was first introduced last December, with around eight AI cameras installed in central Athens. Even in the early stages of implementation, about 29,000 tickets were issued over roughly two months. To put it into perspective, it took only four days to detect 1,000 traffic violations—and that was with just eight cameras.

Initially, the sheer volume of tickets issued made it seem as though the AI traffic cameras could completely replace traffic police. In fact, they caught the world's attention by flagging a number of violations that traffic officers could never have spotted. For example, one camera detected over 8,000 instances of mobile phone use during the pilot period, along with 1,028 speeding violations. Elsewhere, 13,700 red-light violations were recorded, alongside countless cases of not wearing a seatbelt. So, were there really that many lawbreakers in central Athens? As it turns out, no.

Only 3 to 5% of all detected violations actually resulted in a fine being issued. The rest were errors that didn't even warrant a review by the responsible officers, and among those that were reviewed, only a few hundred were deemed valid. The most common errors involved smartphone use while driving and not wearing a seatbelt.

In the case of smartphone use, the AI cameras classified drivers as violating the law even when they were simply vaping or even changing gears. The same issue occurred with seatbelts: if a driver was wearing a black shirt, the camera mistook it for the seatbelt not being worn, as the colours matched. There were even instances where empty seats with no passengers were flagged for not wearing a seatbelt.

In some cases, an ambulance driving in a bus lane was flagged for illegal parking, and vehicles moving according to a traffic officer's hand signals were cited for running a red light. Ultimately, this only proved that the system's accuracy was woefully inadequate. But the bigger problem is that these issues came from just eight cameras. In other words, analysts had to manually check tens of thousands of violations from only eight cameras, and only 3 to 5% of them were actual offences. This means that rather than reducing the workload, AI actually increased it.

What's even more surprising is that the Greek government plans to introduce additional AI cameras in the future—around 1,000 of them. If the current error rate isn't improved, this will require even more manpower. Of course, they claim that errors will gradually be reduced through machine learning and algorithm adjustments, but for the time being, this will inevitably demand a huge amount of human resources.

Recently, South Korea also piloted a similar AI-based unmanned enforcement system. While the exact recognition rate hasn't been disclosed, they tested how accurately the AI could analyse complex situations, such as blocking intersections. According to reports, the recognition rate was quite accurate, and there have been announcements about expanding the system nationwide from 2027.

In any case, AI is now being introduced into traffic cameras worldwide. Although there are still errors, one thing is clear: even the slightest moment of carelessness or a small mistake on my part will be caught without mercy. And that's how it should be. But when it comes to things like wearing a seatbelt while driving and not using a smartphone—rules that are essential for the safety of both myself and others—they must be followed without exception.