Tesla’s Full Self-Driving system has been involved in remarkably few reported crashes during its first months on Belgian roads. But a new test suggests the technology still struggles with some of the country’s most important traffic rules, repeatedly speeding in 30 km/h zones and attempting to overtake cyclists in bicycle streets. The contradiction exposes the central problem with Tesla’s ‘self-driving’ technology: the car may drive smoothly, but the human still has to watch for its mistakes.
Between 11 June and 31 August, Tesla vehicles using Full Self-Driving (FSD) travelled around 24.6 million kilometres in Belgium and were involved in five reported incidents, according to figures supplied by Tesla to Flemish Mobility Minister Annick De Ridder (N-VA), Gazet van Antwerpen reports. Conventionally driven Teslas covered 147 million kilometres during the same period and were involved in 150 reportable collisions.
On those figures, the accident risk with FSD is almost ten times lower on roads away from motorways and around 30% lower on motorways. Yet a new investigation by Johanna vzw, an organisation campaigning for safer pedestrian traffic, has found a very different weakness: when confronted with Belgian traffic rules designed specifically to protect cyclists and pedestrians, the system can repeatedly get the basics wrong, De Morgen reports.
The organisation tested a Tesla Model 3 fitted with a recent version of FSD, using a GoPro camera to record the journeys. Across 29 sections of 30 km/h zones, the Tesla drove too fast on 16. In six tested bicycle-street sections, it failed to display the correct 30 km/h limit in every case and actually exceeded the limit twice. In five of 12 tested residential areas, it also exceeded the 20 km/h limit.
And in the Flemish Brabant village of Winksele, video showed the Tesla repeatedly attempting to overtake a cyclist in a bicycle street.
A car that drives beautifully, but sometimes breaks the rules
That is what makes the current debate about Tesla’s technology so complicated.
I experienced FSD myself for the first time in August, as a passenger in a Tesla taxi travelling from Turnhout to Antwerp. The driver asked whether I was comfortable with him activating the system. I was curious, and we spent part of the journey talking about how it works and what the driver still has to do.
The answer was: quite a lot.
The driver had to remain alert and ready to intervene. But at the same time, he admitted that the Tesla drove better and more smoothly than he did.
It was a late-night journey, and the system was remarkably composed. The car stayed in the middle of its lane, anticipated acceleration and braking smoothly and maintained safe distances from vehicles ahead. There were none of the little fluctuations and corrections that are familiar from human driving.
As a passenger, it was reassuring.The car simply drove.And that is precisely why the new findings are important. A system can be smooth, predictable and apparently safe for hundreds of kilometres while still making a fundamental mistake when it encounters a specific road rule.
The 30 km/h problem
The Johanna test focused deliberately on situations in which Belgian traffic rules differ from those encountered on more conventional roads.
The researchers drove 29 segments in 30 km/h zones, some of them more than once, producing 59 journeys in total.
The Tesla exceeded the speed limit on 16 of the 29 segments. Its average speed on those sections was around 44 km/h.
According to Geert Van Waeg, chairman of Johanna vzw, the striking part was that the car often appeared to recognise the 30 km/h sign but nevertheless continued to behave as if the speed limit were 50 km/h.
Imec AI researcher Steven Latré suggested that this could be a software bug. Tesla vehicles do not determine their speed solely from what they see on traffic signs; they also use virtual speed maps. In urban environments, where signs and other information can conflict, the car may apparently choose the map data over the physical sign.
That would be a particularly awkward failure for an automated driving system.
A human driver may fail to notice a 30 km/h sign. A machine can actually see it and still choose the wrong speed.
Bicycle streets are an even bigger problem
The test produced similar results in bicycle streets. None of the six tested sections showed the Tesla displaying the correct 30 km/h speed limit. In two cases, the car actually drove faster than allowed. More worrying was its behaviour around cyclists. In Winksele, the Tesla was filmed repeatedly attempting to overtake a cyclist in a bicycle street.
That is not simply a question of speed. It suggests that the system may not always correctly interpret the purpose of the road environment.
Bicycle streets are specifically designed around the priority of cyclists. The challenge for an automated system is therefore not merely recognising a cyclist but understanding the traffic rule governing the entire street.
The same problem appeared in residential areas. The Tesla exceeded the permitted 20 km/h speed on five of the 12 tested sections.
For pedestrians and cyclists, these are precisely the environments in which automated driving needs to behave most conservatively.
Belgium’s roads were already a concern
The findings also echo concerns raised before Belgium approved FSD.
The Flemish Department of Mobility and Public Works had warned that the Belgian testing programme was relatively limited and that the safety of the system could not be demonstrated in every relevant scenario.
The department’s assessment identified repeated difficulties interpreting traffic signs, supplementary signs, road markings and certain infrastructure elements, including railway crossings. It also warned about situations in which drivers might fail to react quickly enough because their attention had become relaxed.
Belgian mobility expert Dirk Lauwers of the University of Antwerp has similarly pointed out that the Belgian road network is less standardised than the Dutch one. Road layouts, markings and signs can be less predictable, making Belgium a more demanding environment for automated driving.
That matters because the technology was extensively tested in the Netherlands before its Belgian approval. Dutch roads provide a useful European test environment, but they do not perfectly reproduce Belgian conditions.
The Flemish tests, which began only shortly before the June approval, were intended partly to establish how the system coped with those differences.
The safety figures tell another story
Yet there is another side to the Belgian experience. According to figures presented by Minister De Ridder, FSD-equipped Teslas travelled almost 25 million kilometres between June and the end of August and were involved in only five incidents.
One was a reportable collision involving minor damage. The driver had switched off FSD shortly before the crash. The other four incidents were described as very minor, including cases in which another road user apparently touched the Tesla.
By comparison, conventionally driven Teslas covered 147 million kilometres and were involved in 150 reportable collisions during the same period.
The government therefore argues that the technology is already demonstrating a safety benefit.
But the figures come from Tesla and are not the result of a controlled comparison between otherwise identical drivers, vehicles and journeys. Different types of roads, driving patterns and drivers can influence accident rates.
The numbers are consequently an important early indication, rather than proof that FSD is safer in every circumstance.
The technology remains Level 2
Tesla’s response to the new findings underlines the crucial point.
The company says it is difficult to comment on individual incidents but stresses that its system is approved as a Level 2 driver-assistance system. That means the driver remains responsible for the vehicle, must stay alert and must be ready to intervene.
In other words, Tesla does not claim that the car can take responsibility for driving.
The terminology is confusing because Tesla calls the system Full Self-Driving. The company’s own qualification – Supervised – is therefore essential. You cannot activate FSD and go to sleep. You cannot move to the back seat. You cannot read a newspaper and assume the computer will deal with whatever happens. And you cannot assume that because the car has driven perfectly for the previous 100 kilometres it will necessarily understand the next 30 km/h zone.
The automation paradox
That creates one of the strangest aspects of the technology. The better the Tesla drives, the less the driver has to do.
During my own journey from Turnhout to Antwerp, there was little to complain about. The car kept a steady position in its lane, maintained distance and handled acceleration and braking smoothly. As a passenger, it was easy to understand why someone might quickly become comfortable with the technology.
But comfort is precisely what the system cannot afford to encourage too much. VAB motoring journalist Gert Verhoeven has described the situation as inherently strange: the driver is sitting in a sophisticated automated vehicle, doing very little, but remains completely responsible for spotting problems and intervening.
That can create automation complacency. If the car has behaved correctly for hour after hour, the human brain has little incentive to maintain the same level of vigilance.
Tesla therefore monitors driver attention and issues warnings when it believes the driver is not sufficiently engaged. Repeated failures to respond can cause FSD to disengage or suspend access.
The technology is designed around the assumption that the human will catch the machine when it gets something wrong.
The railway crossing problem
The new Johanna findings are particularly relevant because they reinforce concerns already raised about infrastructure.
The Flemish assessment identified railway crossings among the situations that could cause difficulties.
Lauwers has warned that failure to recognise a railway crossing correctly would be extremely dangerous. That illustrates the difference between testing an automated vehicle on a road where everything behaves predictably and testing it in the real world.
The real world contains faded markings, contradictory signs, temporary roadworks, unusual intersections, cyclists behaving unpredictably, pedestrians stepping into the road and infrastructure that does not look quite like the examples in the training data.
The machine has to cope with all of them.
Other manufacturers are waiting in the wings
The Belgian experience could also have consequences beyond Tesla.
De Ridder has said that Belgium should be open to other manufacturers developing comparable automated-driving systems. Mercedes, BMW and Chinese manufacturers such as BYD are among the companies developing increasingly advanced driver-assistance technologies.
The question is therefore no longer simply whether Tesla should be allowed to use FSD in Belgium.
It is how Belgium should assess any automated driving system when the technology becomes capable of handling most ordinary traffic but can still fail spectacularly on specific local rules.
A failed driving test?
Johanna vzw is calling for the Belgian authorisation of Tesla’s FSD to be suspended.
The Brussels-Capital Region Mobility Minister Elke Van den Brandt (Groen) has also criticised the technology following the findings, arguing that Tesla’s system would currently fail a driving test and pointing to other automated-driving systems that she says have produced stronger results.
The Flemish Government, however, has no intention of suspending the authorisation. De Ridder’s cabinet continues to stress that FSD is a supervised system and that drivers must intervene whenever the vehicle makes a mistake or fails to apply a traffic rule correctly.
That leaves Belgium with an unusual situation. The technology has produced encouraging accident figures while simultaneously demonstrating that it can break basic traffic rules in precisely the environments where vulnerable road users need the greatest protection.
Both things can be true. A Tesla may be statistically involved in fewer crashes while still making individual decisions that a human driver should immediately correct.
So, is Tesla’s FSD safe?
The early Belgian experience does not provide a simple yes or no. The accident data suggest that automated driving may have significant safety potential. My own experience as a passenger also showed how smoothly and predictably the system can drive over a long journey.
But the Johanna tests demonstrate the other side of the technology. A car can drive beautifully and still misunderstand a 30 km/h zone. It can see a cyclist and still try to overtake them in a bicycle street. It can recognise a traffic sign but apparently give greater weight to another source of information.
That is why the word Supervised matters so much. For now, Tesla’s FSD is not a system that replaces the driver. It is a system that takes over much of the driving while asking the driver to remain ready for the moments when it cannot.
And those moments may be precisely the ones that matter most.
