Tesla's newly unsealed crash reports contained two accidents. The surprising part: neither was a malfunction of the self-driving algorithm. The crashes happened at precisely the moment a human — a teleoperator steering the vehicle remotely — intervened. Behind the phrase "AI is driving," there was a person.
There's a reason this amounts to more than an entry in an accident log. Even in the field where automation is furthest along, the human role hasn't disappeared — it has been redeployed in a different form. And that redeployed human can still be the cause of a crash. For anyone weighing AI adoption at work, these two facts pose a question that hits closer to home than it might seem.
What the Unsealed Files Revealed
Tesla launched its robotaxi service in Austin, Texas, earlier this year. The service was billed as fully autonomous, but the actual operating structure was different. In situations the vehicle struggles to judge on its own — an unfamiliar intersection, a sudden obstacle, unusual road conditions — a remote human driver steps in. That person is the teleoperator.
The recently disclosed crash reports recorded two accidents. According to TechCrunch's analysis of the documents, which were filed with the National Highway Traffic Safety Administration (NHTSA), the first crash occurred when a teleoperator collided with another vehicle during remote operation; the second happened during a parking attempt. Neither was an autonomous-driving judgment error — both occurred while a human held the controls.
Tesla is reported to have notified regulators and taken corrective action, and the company maintains that the service is continuously improving. But the question these two crashes raise isn't "Is self-driving safe?" It's "Where does human responsibility sit in autonomous driving?"
The Human Role "Full Automation" Has Been Hiding
Teleoperators are something of an open secret in the autonomous-vehicle industry. Competitors like Waymo and Cruise also employ remote supervision staff. Because today's technology can't yet handle every situation independently, humans remain tethered to the system as designated exception handlers.
This structure carries a built-in vulnerability. When the automated system is working normally, there's nothing for the human to do. But at the exact moment the system gives up on a judgment — in the hardest, most exceptional situations — the human is abruptly handed control. Reaction time is short; situational awareness is incomplete. That is precisely the structure in which the Tesla crashes occurred.
The 2023 Cruise pedestrian accident left a similar question behind. The controversy then centered on the vehicle dragging a pedestrian some distance after the collision, and one cause cited was a teleoperator issuing a move command without fully grasping the situation. A direct comparison with the Tesla crashes is difficult, but the same structural question remains: when we delegate the hardest judgment calls to a human, does that human have sufficient context and preparation?
The Counterargument: Automation Still Beats Humans
The fact that teleoperators caused crashes is separate from any claim that autonomous driving as a whole is unsafe. Waymo, citing millions of miles of driving data, argues that its robotaxis crash less often than the average human driver. If automation's overall safety contribution is large enough, the logic goes, teleoperator-related accidents in a handful of edge cases don't undermine the system's overall record — and that logic is persuasive.
There's also the view that early accidents are inevitable as a technology matures. Disclosing them and reporting them through formal channels is itself responsible conduct, and this kind of transparency arguably builds safer systems over the long run. If rejecting imperfect automation would produce more accidents, not fewer, one might even conclude that the current structure is not second-best but optimal.
Even so, the core question stands. When we call an automated system "complete," how much of the human role embedded inside it do we actually acknowledge? And what training, authority, and responsibility are we giving that human?
Your Workflow Has a Teleoperator, Too
A Tesla robotaxi may sound like a story from somewhere far away. But the question of how responsibility is structured in automation has already arrived in workplaces everywhere — Korea, where this column is written, included.
There's a pattern that repeats in teams that have adopted AI tools. When the tool works well, nobody pays attention. When it errs, or when a call gets hard, responsibility has to land somewhere. "The AI did it" is not where responsibility ends. The person who chose that AI, the person who operated it, the person who reviewed the exceptions, the person who made the final call — every one of those seats is a human seat.
Do you know under what conditions the AI tools you use fail or turn low-confidence? The point where the tool gives up on a judgment is exactly your intervention point. A teleoperator is summoned when the vehicle can't decide for itself. So are you. And in that moment, how prepared are you?
It's also worth examining whether the human who intervenes has enough context. The teleoperators in the Tesla crashes had to judge unfamiliar road conditions remotely, within seconds. In your team's AI workflow, what information does the final reviewer have to process, and how fast? Confirming that this person can actually make the call comes before adopting the tool.
The regulatory environment is shifting, too. Led by the EU AI Act, rules in many countries are moving toward clearly designating who is accountable for AI systems. Mapping out the points of responsibility in the automation pipeline you're building now will leave you in a far better position later.
The view of AI as a tool that extends human capability is currently the most persuasive frame. But extension is not replacement. As automation grows, the quality of the role humans play inside it has to rise with it. The capability left to humans in the AI era is the ability to catch the hardest moments — the ones the system gives up on. That ability doesn't appear on its own. It shows up only in someone trained and prepared.
In the end, this is the question Tesla's two crashes leave behind: in the automated system you run, who occupies the human seat at the hardest moment — and how have they been prepared?
