June 5, 2026 - Tesla FSD Trainers Don't Trust It

This is the Telemetry Transportation Daily for June 5, 2026, and I'm Sam Abuelsamid, Vice President of Market Research for Telemetry. 

It's been nearly a decade since Elon Musk declared that all Tesla vehicles coming from the factory had the hardware to be fully Level 5 autonomous with just software updates. I've always known that to be a lie since L5 requires the vehicle to be able to drive itself in all conditions, and Tesla vehicles lacked the systems to keep all the sensors clean, which would be necessary for automated operation in poor weather. 

One year ago, Tesla launched its first robotaxi service in Austin, Texas, but it had a very limited area where it could operate, and the vehicles still had on-board safety monitors. Later, some of those staff were replaced with remote monitors, but the fleet is still very limited and has had numerous crashes. 

Now, in a new report, Reuters interviewed 10 former Tesla staff who worked on trying to get Full Self-Driving ready for the robotaxi launch. Despite claims from Musk that virtually all of the data labeling for training was being done automatically and the AI driving model was generalized so it could drive anywhere, a team based in Utah was apparently doing most of the work to refine the behavior. Musk's claims of a generalized AI contrast with the approach of other companies that spend months testing and training their systems in new cities before driverless deployments. It turns out that Tesla was doing exactly the same thing, and the system still struggles to deal with emergency vehicles, school buses, and other road hazards. 

Tesla's safety claims have always been rather opaque in their methodology and are believed to compare a small subset of Tesla crashes with airbag deployments against the population of all vehicles in the U.S. in all crashes. Since the average age of vehicles in the U.S. is now 13 years, much older than virtually the entire population of Teslas, which have more safety devices, the comparisons are not actually statistically valid. It's telling that seven of the 10 people interviewed by Reuters would not trust FSD to drive for them.  The biggest challenge is that Tesla FSD can work very well right up until it doesn't, and it fails in inconsistent ways that aren't always repeatable. The current Tesla approach is unlikely to lead to a viable automated driving system anytime soon. 

Thanks for listening.