A 2025 Tesla Model 3 drives off of a rural road, clips a tree, loses a tire, flips over, and comes to rest on its roof.
Luckily, the driver is alive and well, able to post about it on social media. The problem? The car was driving itself in Full Self-Driving.
The subtitle from the driver is this:
“Tesla 2025 model 3 on 13.2.8 driving off the road and crashing into a tree.”
Try to Make Sense of the Onboard Footage (don’t worry, no serious injuries)
Here’s the dash cam footage: important to emphasize here, the driver posted to this to Youtube themselves, and they suffered no major injuries, just a bit shaken up. Skip to 0:30 seconds to catch the critical error:
The driver very considerately posted all of their camera output to Youtube in a playlist: that is here. They also discuss the incident briefly in a Reddit thread, which is here.
The damage looks pretty devastating, glad they walked way alright:
Driver Had Less Than One Second to Intervene
Tesla’s legalese makes it clear before you activate Full Self-Driving (Supervised): drivers are expected to supervise the vehicle and take control immediately if anything weird happens. If your vehicle crashes, you are fully liable. But what happens when intervening “immediately” isn’t fast enough?
In the crash video, from the moment the Tesla begins its unprompted veer left (0:32) to the moment it leaves the pavement (0:33), the window of opportunity to avert disaster lasts a fraction of a second.
Let’s say that again: less than one second between safe driving and flipping your car into the woods for no good reason.
Now consider this: the average human reaction time, according to the NHTSA, is about 1.5 seconds in real-world driving conditions. That’s measured from the eye perceiving a dangerous autonomous turn is occurring, to the nerves telling the hands to turn the steering wheel back onto the road.
That’s best case scenario, too, an attentive driver with their eyes on the road and hands on the wheel.
So… even if this driver had been paying perfect attention and had hands hovering over the wheel, they still wouldn’t have been able to react in time.
An Autonomous Crash That Happened Faster Than a Human Can React
Let me be blunt: this crash was physically impossible to prevent. The car made a fatal decision faster than any human could possibly correct it. Tesla’s idea that drivers can “supervise” these systems is, at this point, nothing more than a legal loophole.
It’s like saying you should supervise your washing machine to prevent lint trap fires. Except, the washing machine is rolling down the freeway at 65 miles per hour, trying to dodge ghosts that only it can see.
So, Why Did the Car Fully Self Drive into a Tree?
Here’s my best guess: it saw a shadow.
In the crash footage, there’s a shadow across the road from a “Watch for Trucks” sign. My theory? Tesla’s neural net saw that shadow and panicked.
Instead of recognizing it as just a dark spot on the pavement, the system may have hallucinated an obstacle and took evasive action—by jerking left, directly off the road.
Insane? Yes. But this is how Tesla’s so-called “neural network” sees the world. It doesn’t use LIDAR. It doesn’t use radar. It relies on vision only, and so it can misinterpret 2-dimensional images that are processed through machine learning that’s notoriously difficult to debug or explain after the fact.
“Hey Kay, What’s an AI Hallucination?”
Glad you asked.
In machine learning, an AI hallucination is when a system confidently outputs something completely wrong. It might see a bush and call it a person. Or a shadow and call it a concrete barrier. These systems don’t “see” the way you and I do. They work off probabilities and pattern recognition, and sometimes those patterns go sideways.
Even a system that is right 99% of the time will be dead-ass wrong some of the time, and those errors can be maddeningly simple or inexplicable, like this case of a Tesla driving itself into a ditch unprovoked.
Tesla’s Full Self-Driving is powered by an end-to-end neural network that tries to interpret the road using visual input alone. And while that sounds impressive on an investor relations call, it’s also fundamentally unpredictable.
You can’t interrogate it. You can’t ask “why” it made a decision and expect a rational answer 100% of the time. And if it decides that shadow is a brick wall? Buckle up.
Coming Imminently to Austin TX: Unsupervised Tesla Taxis
Speaking of buckling up, let’s talk about Tesla’s much-hyped launch of autonomous taxis in Austin TX. T-minus 10 days, from the look of it.
Not “supervised” FSD. Not “hands on the wheel, eyes on the road” systems. They’re calling these unsupervised.
These vehicles will operate without drivers. The plan is to start with a safety operators, but we all remember how well that worked for Uber in their ill-fated autonomous attempt.
I’ve got a question for Elon and company:
If the supervised version of your software is already capable of flipping a car into the woods for no reason, how exactly is the unsupervised version supposed to be safer?
Like, let’s compared the upcoming autonomous taxis to the FSD system that left one driver upside-down in a ditch:
- Is it built on the same neural net? Yes.
- Is it still vulnerable to hallucinations that send it jumping at shadows? Yes.
- Is there any human on board able to intervene in under a second? No.
So we’re just… hoping it doesn’t crash? That sounds like the sort of gamble that folks make on a closed track with professional drivers. As it stands, I don’t see any way Tesla can deliver the bare minimum of safety for unsupervised autonomy as a product outside of lab testing.
*Credit for the cover image to Reddit OP!
