A windshield camera stares straight into a low afternoon sun, and for a few seconds the lens sees nothing but white.
Inside the cabin, nothing on the dashboard changes.
No chime, no warning, no dimmed indicator tells the driver the car’s eyes just went dark.
That silence is now the subject of one of the most serious federal probes facing any driver assist system on American roads.
Investigators want to know how a car can fail to notice it can no longer see.
How a camera only car is supposed to notice it has gone blind
The system relies entirely on cameras to steer, brake and judge distance, a setup Tesla built after it began dropping radar sensors in favor of a camera only approach a few years ago. When Tesla began transitioning away from using both cameras and radars to an exclusively camera-based approach, known as Tesla Vision, in mid-2021, it developed and implemented a degradation detection system that it deployed by a software update to existing and new Tesla vehicles.
That safeguard, the software designed to alert drivers when camera performance is compromised, has a job that is narrow but critical. It has to watch the live camera feed, recognize when glare, fog or airborne dust has degraded what the car can actually see, and warn the human before the car’s judgment becomes unreliable.
The problem, according to federal regulators, is that this watchdog does not reliably do its job. NHTSA found that Tesla’s degradation detection system fails to detect and/or warn the driver appropriately under degraded visibility conditions such as glare and airborne obscurants, and in the crashes reviewed, the system did not detect common roadway conditions that impaired camera visibility and/or provide alerts when camera performance had deteriorated until immediately before the crash occurred.
So the car keeps driving as if nothing changed, right up until the moment it cannot avoid what it never saw coming. The system still requires an attentive driver with hands on or near the wheel at all times, a safeguard that only matters if the alert arrives before, not after, the danger appears.
A highway scene that keeps repeating in the case file
The pattern investigators describe is strikingly consistent. A car running Tesla’s Full Self-Driving mode is traveling along a highway, sun low on the horizon or dust kicked up from a passing vehicle, and the forward camera’s view starts to wash out.
In crash after crash, the same failure unfolds. In many of the crashes reviewed, FSD also lost track of or never detected a lead vehicle in its path.
That detail matters because it is not a vague software glitch. It is the single scenario, a car losing sight of the vehicle directly ahead, that a basic radar unit was built to catch decades ago, long before any automaker tried to replace it with cameras alone.
Owners who lean on this mode for long highway stretches, the kind of convenience bundled into some sticker prices, are the ones most exposed to this exact blind spot.
What the federal case file actually shows
NHTSA first opened a preliminary evaluation into Tesla’s FSD software in October 2024 covering 2.4 million vehicles, then escalated it to an engineering analysis, a required step before it can seek a recall. That engineering analysis covers an estimated 3.2 million vehicles.
The case file lists nine documented incidents so far. NHTSA said it had reports of nine incidents that may be tied to the issue including one fatal crash and two injury crashes, and was investigating whether six other crashes may be related.
The automaker’s own internal review conceded real limits in its fix. Tesla’s post-incident analysis indicated that the update to the degradation detection system, had it been installed on the vehicles at the time, may have affected 3 of the 9 incidents identified by ODI. The resume gives no result for the other six.
Investigators have gone further, demanding an internal Tesla document submitted in response to the earlier preliminary investigation that they believe bears directly on what the company already understood about its camera only approach before it shipped the system nationwide. Full details of the investigation are posted on the federal safety database.
Where the camera only bet still runs into real limits
None of this means the system fails constantly, or that every foggy commute ends badly. The vast majority of trips under degraded light never produce an incident at all, which is precisely why nine crashes out of millions of miles driven still took years to surface as a pattern.
But the complication investigators keep returning to is a structural one, not a one time bug. NHTSA also flagged concerns about Tesla potentially under-reporting its crash rates, noting that Tesla cited internal “data and labelling limitations” that may have caused it to miss crashes during portions of the defined investigation period, which raises the separate question of whether the true crash count is even higher than nine.
That gap between what a company can measure and what actually happens on the road is familiar territory. It echoes concerns raised elsewhere in the used and certified vehicle market, including reports that claim clean histories without complete data behind them.
NHTSA typically completes an Engineering Analysis within 12 to 18 months of opening, and it can end two ways, a recall order or regulators simply closing the file once they see enough changes.
What happens next, and why the outcome still matters to drivers
For now, the cars stay on the road and the software keeps shipping updates while the review continues. Millions of owners are, in effect, part of a live test of whether a camera only safety net can be patched into something reliable, or whether it needs a harder redesign instead.
So the stakes reach past one company. Every automaker weighing a similar camera only bet is watching how this case resolves, since a recall here would ripple into how sensor choices get made industry wide for years to come.
Even so, the fix itself may be straightforward compared to the uncertainty around it. A reliable alert, one that tells a driver plainly that the car can no longer see well enough to trust itself, is not a new technology.
It is mostly a matter of building in honesty about the system’s own limits, and acting on it before the next sunset drive turns into the next case file entry.
