Five years, 2 million recalls, still hitting police cars: The fatal flaw inside Tesla’s self-driving tech

Just before midnight on October 13, 2026, a Tesla drove through a four-lane freeway closure on eastbound Interstate 80 in Fairfield, California. Cones, flares, and emergency lights marked the scene, where California Highway Patrol (CHP) officers were handling an earlier fatal crash. The car struck a parked Dodge Charger patrol vehicle with two officers inside, seriously injuring both.

The 22-year-old driver reportedly told investigators he was asleep with Tesla’s Full Self-Driving (FSD) system engaged. CHP released dashcam footage showing the vehicle entering the closed-off area before impact. CHP has not released the car’s model or software version.

An important caveat applies: investigators have not independently confirmed that FSD was active at the moment of impact. The driver’s statement is the main evidence so far, and the vehicle’s data logs could show what the system was doing. CHP said that driver-assistance technology does not relieve drivers of responsibility for operating their vehicles.

Even so, the crash raises an uncomfortable question for Tesla. Five years after US regulators began investigating Teslas colliding with parked emergency vehicles, and nearly three years after a recall covering about 2.03 million vehicles, another Tesla has struck a patrol car at an emergency scene.

It doesn’t prove software caused the crash, but it shows why regulators keep scrutinising these systems. It also isn’t the first such case: in June 2024, a Tesla whose driver said he had engaged “self-drive” mode while on his phone hit a parked police car in Fullerton, California, where an officer was working a fatal crash. The officer moved aside in time, and police did not confirm whether Autopilot or FSD was active.

The complacency trap: when driver assistance feels like autonomy

Tesla sells the feature as Full Self-Driving (Supervised), but the name does not mean the car can drive without human oversight. It is a driver-assistance system. The driver must stay attentive and ready to intervene.

That distinction matters because automation changes how people behave. When a car handles routine driving smoothly, drivers can become less vigilant, even though the system can still encounter situations it cannot manage. Researchers call this automation complacency. The more comfortable people get with assistance, the harder it becomes to stay alert and react quickly when something goes wrong.

Price and convenience add to the effect. In the US, FSD (Supervised) costs $99 a month, about $3.30 a day, after Tesla halved the price from $199 in April 2024. Tesla also reportedly ended outright purchases in February 2026, which would make a subscription the only way to get FSD in the US. A cheap, cancel-anytime subscription makes the system easy to try and easy to lean on, which raises the stakes for clear communication about its limits.

The Fairfield crash illustrates the risk without yet proving its cause. If the driver’s account is accurate, it raises questions about hazard detection and about whether the driver-monitoring system noticed that the person at the wheel was asleep.

Three issues deserve particular attention.

Vision-only perception has limits. Tesla relies on cameras and software to read the road. Cameras can be affected by glare, darkness, airborne obstructions, and other conditions that reduce visibility. Flashing emergency lights, flares, and unusual road layouts can add to the challenge, though investigators have not established that any visual limitation caused the Fairfield crash.

In March 2026, the National Highway Traffic Safety Administration (NHTSA) escalated its investigation of FSD to an engineering analysis (EA26002) covering an estimated 3,203,754 vehicles. The probe examines whether Tesla’s “degradation detection” system recognises reduced visibility from glare, dust and other airborne obstructions, and whether it warns drivers appropriately. Reuters reported that the upgrade followed nine crashes, one of them fatal. NHTSA said the available data raise concerns that the system, both as first deployed and as later updated, may fail to detect poor visibility or to warn drivers adequately.

Driver monitoring is not the same as driver attention. Since 2024, Tesla’s primary monitor has been a cabin camera that tracks the driver’s face and eyes. But Tesla’s own owner’s manual says the vision-based monitoring will not activate when the camera is covered, when lighting is poor, or when the driver wears sunglasses or a hat that hides the eyes.

In July 2026, a driver was filmed apparently asleep at about 100 km/h on British Columbia’s Trans-Canada Highway with two children in the car, wearing large sunglasses. Reports have also described owners fooling the camera with plastic doll heads. A US congressman has pressed the Department of Transportation on the issue after citing 43 videos of sleeping drivers, and asked whether the camera still works when a driver wears sunglasses or covers the lens.

In another case, a driver apparently almost died when the FSD-driven Tesla did not attempt to brake or slow down at a railway crossing – while the train was moving. The driver said FSD failed to slow down for a railroad crossing in fog, and he swerved only at the last moment.

These cases should be judged on their own evidence, but they underline a broader point. Monitoring must be judged by how well it prevents misuse, not by whether it is installed.

The name can create the wrong impression. “Full Self-Driving” sounds more capable than a system that needs continuous supervision. The word “Supervised” can get lost when the shorter name dominates headlines and social media. In December 2025, when asked whether drivers could text while using FSD, Elon Musk replied, “Depending on context of surrounding traffic, yes”, even though texting behind the wheel is illegal in nearly every US state. Statements like that blur the line between assistance and autonomy, and the driver remains responsible either way.

Europe has pushed back on the name. After Germany’s transport ministry called it misleading, Tesla offered to rename the system “Tesla Assisted Driving” in Europe. The US name has not changed.

The US regulatory approach is being tested

In August 2021, NHTSA opened an investigation into Teslas on Autopilot crashing into stopped emergency vehicles. It covered about 765,000 vehicles and began with 11 crashes at scenes marked by warning lights, flares, or cones.

In December 2023, Tesla recalled 2,031,220 US vehicles under recall 23V-838 because Autopilot’s controls were insufficient to prevent misuse. Tesla did not agree with NHTSA’s analysis but chose to carry out the recall voluntarily. The over-the-air remedy made visual alerts more prominent, simplified engaging and disengaging Autosteer, added checks, and allowed Autosteer to be suspended for drivers who repeatedly fail to stay engaged.

A recall’s size should be read correctly. It counts the vehicles covered, not the vehicles in crashes. And a high completion rate for an over-the-air update shows the fix was delivered, not that it worked. NHTSA itself opened a recall query in April 2024 to assess whether the remedy was adequate, citing crashes after the update and its own tests of remedied vehicles. NHTSA’s opening documents say preliminary analysis found crashes in vehicles that already had the remedy installed.

Scrutiny has since widened. Beyond the visibility probe, NHTSA opened a separate investigation in October 2025 into 2.88 million FSD vehicles over more than 50 reports of traffic-safety violations and a series of crashes. Federal scrutiny has grown roughly fourfold, from 765,000 vehicles in 2021 to about 3.2 million in 2026.

Autopilot and FSD are different products, and findings about one should not automatically be applied to the other. Together, though, the investigations raise a broader question: do existing safeguards adequately address the risk of drivers relying too heavily on increasingly capable automation? The US approach leans on monitoring vehicles in use, investigating incidents and requiring recalls when defects are found. That can uncover real-world problems, but it also means some risks only become visible after vehicles are already on the road.

China and Europe are taking different approaches

Rules outside the US are also evolving, and they shouldn’t be reduced to a contrast between permissive American rules and stricter foreign ones. China published GB 44721-2026 on July 30, 2026, its first mandatory national standard for Level 3 and Level 4 automated-driving systems. It takes effect on July 1, 2027. It requires manufacturers to build safety-assurance processes, sets performance requirements for hazard response, covers driver interaction, and calls for confirmatory testing by third-party organisations using closed-course, on-road and simulation methods. It also aligns with a UN global technical regulation on automated driving adopted in June 2026.

Two cautions apply. The standard covers Level 3 and Level 4 systems, while FSD is a Level 2 driver-assistance system, so it does not directly govern products like FSD. And it is not a blanket rule shifting legal liability to manufacturers whenever automation is active. It should be read within China’s wider vehicle-approval framework.

Europe has taken a different route: national approvals and efforts to win wider acceptance across the EU. The Dutch regulator RDW granted provisional approval for FSD (Supervised) in April 2026, and other countries have since built on it.

In October 2026, Germany’s transport minister said he would push for EU-wide approval and agreed with Tesla that the system may exceed speed limits by up to 10%. Sweden and France raised concerns about speeding, and the EU vote has slipped to December at the earliest. Reuters also reported that safety researchers warn Tesla’s safety claims are inflated.

Other systems show that the problem is not unique to Tesla, but also that design choices matter. GM’s Super Cruise and Ford’s BlueCruise allow hands-free driving only on mapped highways and use infrared cameras to watch the driver’s face. Even so, NHTSA escalated an investigation into BlueCruise after two fatal night-time crashes into stationary vehicles, noting that Ford’s cruise control ignores stationary objects at high speed.

Waymo takes a different path altogether: its driverless vehicles combine lidar and radar with cameras, and an IIHS study in July 2026 found they were involved in 68% fewer crashes per mile than human drivers in the cities studied, though only on mapped urban roads. We would like to state that comparing these systems is difficult because they do different jobs, and no ranking should be read as a verdict on the Fairfield crash.

The legal shift: Driver negligence and manufacturer responsibility can coexist

Tesla’s owner manuals say drivers using its supervised systems must stay attentive and ready to take control. That warning matters, but it does not settle every product-liability question.

A 2019 fatal crash in Key Largo, Florida, shows how. The driver of an Autopilot-equipped Model S had dropped his phone and was looking for it as he drove through an intersection at about 62 mph. In August 2025, a federal jury found Tesla 33% responsible and awarded about $243 million: $19.5 million and $23.1 million in compensatory damages to the two plaintiffs, plus $200 million in punitive damages. It was the first federal jury verdict in a fatal Autopilot case. Tesla argued that the driver alone was to blame and that the car was not defective. In February 2026, a federal judge upheld the verdict, and Tesla was expected to appeal.

The case shows that a driver can bear responsibility while a manufacturer is also held liable. But it was an Autopilot case, not a ruling on the Fairfield crash, and it does not make Tesla liable for every collision involving its driver-assistance systems. Courts can weigh product design, warnings, driver behaviour and marketing claims together. A manufacturer’s warning to stay attentive is important, but courts may also ask whether the design and presentation convey the system’s limits.

For Fairfield, those questions remain open. Investigators will need to establish whether FSD was active, what the car detected, how it responded, and whether the driver-monitoring safeguards worked as intended.

What drivers can do

Until the facts are in, the practical lessons are plain. Treat FSD as a driver-assistance feature, not a chauffeur. Keep your eyes on the road and your hands ready, and never use a phone behind the wheel. Don’t defeat or work around the monitoring system. Drive as if a closed lane, a flare, or a flashing light could appear at any moment, and slow down and move over for emergency vehicles as US move-over laws require.

The Fairfield crash is a reminder that branding cannot replace demonstrated safety. Emergency responders working on highways depend on approaching drivers to see closures, slow down, and respond to warnings. Any driver-assistance system operating in those conditions should be judged on its real performance, not its name.

The evidence already available justifies continued scrutiny. NHTSA’s investigations have raised concerns about driver engagement and reduced visibility, and the 2023 recall shows the scale at which software safety fixes can be rolled out. But the precise role of FSD in the Fairfield crash has not been established, and it would be premature to treat this incident alone as proof of a software defect.

Responsibility is shared. Drivers must stay attentive when using supervised systems. Manufacturers must communicate limitations clearly, build effective safeguards, and show that they work in real-world conditions. Regulators must investigate incidents, test safety claims, and require corrective action when the evidence supports it.

The central question is not whether driver assistance makes driving easier. It is whether drivers, manufacturers, and regulators understand its limits well enough to stop that convenience from becoming a public-safety risk.

RELATED COVERAGE: