Tesla Called It a Robotaxi. Its Humans Were Driving It Remotely.
Tesla Called It a Robotaxi. Its Humans Were Driving It Remotely.
For more than a year, Tesla kept the details of its robotaxi crashes hidden behind a claim that regulators had never heard from any other autonomous vehicle company: releasing the data would cause the automaker financial harm, according to Electrek's reporting on Tesla's filings with NHTSA.
The National Highway Traffic Safety Administration just released the unredacted versions. What they show is not a fleet learning to drive itself. They show a system that, in at least two separate incidents, required a remote human operator to physically take control and drive the vehicle into an object, as Electrek reported when it published the full crash narratives.
The first happened in July 2025. A Tesla robotaxi had stopped on a street in Austin and would not proceed. The safety monitor in the passenger seat called for help. A teleoperator remotely drove the car left, up a curb, and into a metal fence at eight miles per hour. The monitor was injured and did not require hospitalization. The second happened in January 2026. Another safety monitor needed navigational assistance. Another teleoperator took over and drove the vehicle straight into a temporary construction barricade at nine miles per hour. No injuries were reported, according to the NHTSA crash narratives Electrek published.
Both crashes occurred while a human safety monitor was in the vehicle. Neither involved a passenger. Both are now in a federal database that every autonomous vehicle operator in the country is required to use.
Tesla was the only company that bothered to redact everything, Electrek noted in its analysis of the data release. Every other robotaxi operator, including Waymo and Zoox, filed crash narratives with NHTSA that included detailed descriptions of what happened. Tesla filed the same forms with the same sections blank, stamped with the same phrase: confidential business information. The company told NHTSA that releasing the data would cause it financial harm.
The crash narratives that Tesla fought to keep hidden now span 17 incidents from July 2025 through March 2026, according to Electrek's compilation of the NHTSA data. Most were minor. Several were rear-endings where the Tesla was stopped and a human driver failed to stop in time. That pattern is not unique to Tesla; Waymo reports the same thing. But the teleoperator incidents are a different category of problem. When the automated system stalled, a human had to drive the vehicle out. That human then crashed it.
The broader dataset does not paint Tesla's system as catastrophically unsafe. Of 17 incidents, 13 resulted in property damage only, two involved no injuries, one involved a minor injury without hospitalization, and one involved a minor injury requiring hospitalization, per Electrek's breakdown of the NHTSA narratives. But the incidents collectively raise a structural question about what Tesla is actually selling.
The company calls it a robotaxi. Elon Musk has described the underlying technology as Full Self-Driving, a name that implies human oversight is unnecessary. Tesla's own safety literature describes the system as capable of navigating urban environments without driver input. The company has fewer than 100 vehicles operating across Austin, Dallas, and Houston, WIRED reported. Waymo, by contrast, operates nearly 4,000 vehicles. In Austin alone, Tesla has roughly 50 cars where Waymo has more than 250, according to Reuters' reporting on a slideshow presentation by Austin officials.
The two companies use fundamentally different approaches to the human in the loop. Waymo's remote workers provide advice to the autonomous system, suggestions it can accept or reject. Waymo says its software remains in control at all times and its workers do not drive the vehicles, Reuters reported in February. Tesla's teleoperators physically operate the car when the system cannot resolve a situation on its own.
This distinction matters for the economics that underpin the entire robotaxi thesis. The argument for autonomous ride-hailing is that vehicles can operate without a driver, eliminating the largest cost in traditional ride-hailing. If remote human operators must actively drive vehicles to avoid collisions, the labor cost does not disappear. It moves to a different job classification.
Tesla's crash rate, as reported to NHTSA, works out to roughly one incident every 57,000 miles, Electrek calculated from Tesla's own vehicle data. The average human driver in the United States reports one crash every 229,000 miles. By that measure, Tesla's system is operating at roughly four times the human crash rate. The caveat is that Tesla's fleet is small and new. Waymo has orders of magnitude more miles of data, and early-stage fleets with limited operational design domains will always look worse on a per-mile basis until the software matures.
But the teleoperator data cuts through that caveat. The system did not simply crash more than humans. In two documented cases, the system stopped and a human driver took over and then crashed. That is not a software maturity problem. That is a human in the loop problem.
Tesla has argued it was protecting competitive information when it redacted the crash narratives. The competitive information it was protecting included the fact that its robotaxis were being driven by people in another location, in real time, in response to system failures. That is not competitive advantage. That is a different product.
The data is public now. The question is what anyone plans to do with it.