Tesla's Own FSD Trainers Don't Trust the Tech They Built

In a ² investigation published May 28, nine former Tesla data labelers — the workers who spent hundreds of hours reviewing footage from Tesla's eight exterior cameras to train the Full Self‑Driving AI — were asked a simple question: would you trust FSD to drive you? Seven of the nine said no. "One put it bluntly, saying he wouldn't ride in a Tesla Robotaxi even if you paid him," Autoblog reported.

These aren't outside critics. They are the people who watched the raw, unvarnished footage of what FSD actually does on real roads — footage that, according to Reuters, is tightly guarded even within Tesla. One former self‑driving engineer who reviewed crash data for years told Reuters that Tesla's safety claims were "bullshit," adding: "Definitely, don't trust Elon on this."

The labelers — hundreds of workers based in Utah — regularly observed FSD failing at basic maneuvers that human drivers handle instinctively. According to the Reuters report and ³'s analysis, these failures included:

• Hitting animals — cats, dogs, and deer — often without braking
• Speeding dangerously — one vehicle clocked at 60 mph in a 25‑mph zone after "Mad Max" mode was introduced, with FSD exceeding speed limits by 20–30 mph
• Failing to stop for emergency vehicles and school buses
• Driving into construction zones, nearly striking workers
• Missing pedestrians in crosswalks and nearly hitting children playing
• Failing to brake on freeway off‑ramps — one Tesla hit a concrete wall

Tesla maintains a special "trauma team" in Palo Alto that reviews near‑misses with pedestrians, but these clips are rarely shared with other teams. "We have all seen it fail," one former labeler told.²

Tesla has long claimed that FSD is 7–10 times safer than the average human driver. But 10 of 11 independent traffic‑safety researchers who reviewed Tesla's methodology for ² concluded the statistics read "more like marketing than genuine safety analysis."

The core issue: Tesla compares airbag‑deployment crashes in its own vehicles to a federal database that includes all tow‑away crashes nationwide — a far less severe category where airbags often don't deploy. When University of Michigan researcher Marco Benedetti performed an apples‑to‑apples comparison using airbag‑involved crashes for both, Tesla's safety advantage dropped from roughly 10x to about 3x, Reuters reported. Even that corrected figure is unreliable because Tesla vehicles average 4.1 years old while the U.S. fleet averages 12.8 years — newer cars have better safety features regardless of automation.

"It's like saying: 'My jet airplane is faster than your World War II bomber.' Yeah, so, what's your point?" Phil Koopman, a professor at Carnegie Mellon University, told.³

Additional problems identified by researchers include: Tesla counts crashes only within 5 seconds of FSD disengagement (federal regulation uses 30 seconds), and the company's claim of saving 32,000+ lives assumes every U.S. vehicle — including trucks and motorcycles — would be replaced by FSD‑equipped Teslas that are 7x safer, a projection multiple researchers called unrealistic.

Elon Musk has repeatedly claimed Tesla's approach doesn't need the "laborious local mapping" that competitors like Waymo rely on. But internal practices tell a different story. Before the Cybercab unveiling in October 2024 and the Austin robotaxi launch in June 2025, Tesla staff worked long hours manually filming exact routes, annotating curbs, road markings, and known hazards, according to Reuters.

Tesla's Utah data‑labeling team roughly doubled to ~300 workers in the six months before the Austin launch, mostly to smooth the test zone. Nearly a year later, only about 20 unsupervised robotaxis operate in Austin's small, heavily mapped area — and some still have human monitors, Electrek reported. A former employee described the limited zone: "You can't get creative outside of that."

This directly contradicts Musk's public narrative that FSD will soon function everywhere without the kind of detailed local mapping rivals depend on. The safeguards used for staged public demos "cannot be deployed at scale," Autoblog noted.

The National Highway Traffic Safety Administration (NHTSA) has four active investigations into FSD and Autopilot, including a probe covering approximately 2.9 million vehicles that focuses on FSD running red lights, drifting into oncoming traffic, and whether a 2023 recall fix was sufficient. The investigation was upgraded after 58 complaints described FSD vehicles blowing through red lights and crossing into wrong‑way traffic.

Legal exposure is mounting. A federal judge upheld a $243 million verdict against Tesla in a fatal Autopilot wrongful death case in Florida. A January 2026 lawsuit involves a Model X crash that killed an entire family. In March 2026, a Cybertruck owner sued after the vehicle allegedly attempted to drive off a Houston overpass while FSD was engaged. Total lawsuit exposure across crash liability, false advertising, and securities fraud is estimated at up to $14.5 billion, according to Autoblog.

Meanwhile, Musk's November 2025 promise to let drivers text while using FSD remains unfulfilled six months later. Tesla's own website still warns: "Currently enabled features require active driver supervision and do not make the vehicle autonomous."

The Reuters investigation exposes a widening gap between Tesla's marketing promises and its engineering reality. The data labelers — the people with the most granular view of FSD's actual performance — don't trust it. The safety statistics that underpin Tesla's entire autonomy narrative are built on a flawed comparison that inflates results by a factor of three. The robotaxi demonstrations that generate headlines and investor enthusiasm rely on extensive manual preparation that can't scale.

For the AI industry, the implications go beyond Tesla. As more companies ship AI systems that operate in physical, safety‑critical environments — autonomous vehicles, medical devices, industrial robots — the question of how safety claims are measured and verified becomes existential. If a company with Tesla's resources and public visibility can present misleading safety statistics for years, what does that mean for smaller players with less scrutiny? The answer, according to the 11 researchers who reviewed Tesla's methodology, is that independent verification isn't optional — it's the only thing standing between marketing claims and public safety.