The morning sun crests a highway overpass and floods the windshield with white light.
A forward-facing camera, mounted behind the rearview mirror, takes the full blast. The driver-assist software is active, steering the car, reading the lane. Inside the cabin, no alarm sounds. No warning light blinks.
And then, in the seconds before impact, a lead vehicle that was always there finally appears on the screen.
How does a system designed to protect drivers end up failing exactly when conditions get hard?
How the self-check is supposed to work, and where it breaks down
Every camera-based driver-assist system faces a problem that radar and lidar do not: a camera cannot see what it cannot see, and it has no independent way to know it has gone blind. To solve that, engineers build a degradation detection layer into the software. This secondary process monitors the image stream for signs of impairment, washed-out pixels, low contrast, sudden drops in object detection confidence, and is supposed to trigger a driver alert the moment performance falls below a safe threshold.
The flaw federal safety investigators identified is not in the cameras themselves. When sun glare, fog, dust, or airborne particles reduce camera clarity, the detection system sometimes fails to disengage the assist feature or alert the driver in time.
The detector is watching the images, but it is not reading the right signals fast enough to give a human driver the seconds they need to take back control. In each crash investigators reviewed, the system also lost track of, or never detected, a lead vehicle in its path. That double failure, the camera going soft and the object tracker going quiet at the same moment, is what turns a sunny morning or a foggy interchange into an accident scene.
What 3.2 million cars on US roads actually carry
The system at the center of this investigation is Tesla’s Full Self-Driving (FSD), a camera-only architecture that Tesla switched to in mid-2021, removing radar from the sensor stack and calling the result Tesla Vision. Some engineers warned at the time that cameras alone would be vulnerable to environmental interference. Every vehicle in the federal investigation docket relies on cameras alone, with no radar and no lidar backup.
The probe covers Model S and X from 2016 onward, Model 3 from 2017, Model Y from 2020, and Cybertruck from 2023, sedan, SUV, and truck variants spanning the bulk of Tesla’s current lineup. The driver-assist feature is not a niche add-on sold to a handful of early adopters. It is a subscription-capable package active in millions of daily commuter cars, at dawn and dusk, the exact hours when forward-facing cameras take the hardest light.
The conditions that trigger the failure are not rare edge cases. Sun glare, fog, and airborne dust are things any driver encounters regularly, and the degradation detection mechanism is supposed to catch all of them before the system drives into something it cannot see.
Nine crashes, one fatality and two injuries, the evidence that led investigators to escalate
The federal safety regulator escalated its probe to a formal Engineering Analysis on March 18, covering roughly 3.2 million vehicles. That escalation is the step that typically precedes a recall demand, and it gives investigators authority to compel the automaker to hand over far wider data than a preliminary evaluation can reach.
The Engineering Analysis upgrades a preliminary evaluation launched in October 2024 after identifying four crashes in reduced visibility conditions, including one that fatally struck a pedestrian. The scope has since expanded to nine total incidents, one fatality, two injuries, and six additional potentially related incidents still under review.
A fatal crash occurred on November 28, 2023. Tesla submitted the required report the following June, seven months later, and began developing an update to the degradation detection system the very next day.
That update, by the automaker’s own analysis, may have affected only three of the nine crashes investigators identified. Investigators also noted that internal data limitations may have led to under-reporting of similar incidents.
Where the camera-only approach holds up and where it does not
Camera vision systems have delivered genuine progress in clear daylight, handling lane lines, traffic signals, and pedestrians with reliability that earlier radar-only systems could not match. The problem is asymmetric: strong performance in clear conditions comes paired with real vulnerability when cameras encounter environmental interference.
A human driver who hits a wall of fog slows down, turns on the hazard lights, and leans forward. A camera-dependent system has none of those instincts. It processes incoming image data and acts on what it reads. If the read is corrupted and the corruption checker misses it, the system keeps driving at speed into conditions it can no longer see.
A newer variation of the system expands into city streets, navigating intersections and stopping at red lights, pushing into an environment where light conditions are far less predictable than a straight highway. Nine documented failures on controlled roads remain unresolved while that expansion continues.
What the engineering analysis changes and what remains open
Investigators will gather further information on the updated degradation detection system, including the status of updating vehicles, the scope of compatible vehicles, and the alerts provided to the driver. That is a broader sweep than the preliminary evaluation allowed, and it means the automaker must respond with production-level data.
An engineering analysis does not guarantee a recall, and the investigation could close if the updated software proves sufficient. But the timeline already on record, a fatality, a seven-month reporting lag, and a patch covering fewer than half the documented incidents, sets a high bar for closure. The full scope of what investigators are demanding is laid out in the official NHTSA filing for Engineering Analysis EA26002. You can also see how regulators handle federal safety escalations in the recent electric cargo van case, where investigators similarly moved before injuries mounted.
The broader driver-assist market is watching closely, as the outcome will set a template for how regulators treat camera-dependent systems across new-vehicle lineups already under market pressure. The key open question is whether a degradation detection system running on the same cameras it is supposed to monitor can ever be fast enough when the light changes in a fraction of a second.
