Tesla AI trainers doubt safety of Full Self-Driving, Reuters finds

Tesla’s internal “data labelers” – the workers who annotate video for its Full Self‑Driving (FSD) system – say they would not ride in a Tesla equipped with the feature, and Reuters has uncovered methodological flaws that inflate the company’s safety claims [1]. The investigation, based on interviews with former labelers, a former self‑driving engineer, and traffic‑safety researchers, challenges Elon Musk’s repeated assertions that FSD is up to ten times safer than human drivers.

Key takeaways

• Tesla’s claim that FSD is “10 times safer” rests on a comparison that mixes airbag‑deployment crashes with a broader federal crash dataset, inflating the figure by about three‑fold [2].
• Former labelers regularly observed FSD failures, including missed stops for emergency vehicles, near‑misses with children, and collisions with animals or construction zones [1].
• Seven of nine ex‑labelers said they would not trust FSD to drive them, and a veteran self‑driving engineer called the safety statistics “bullshit” [1].

Flawed methodology behind the safety numbers

Tesla’s CFO Vaibhav Taneja first touted the “10 times safer” claim in July, and the figure was repeated by board chair Robyn Denholm at a November shareholders’ meeting where Musk displayed a chart showing “85 % fewer crashes” [1]. Reuters’ analysis shows the comparison is misleading: Tesla counted only crashes that triggered airbag deployment in its fleet, then compared that number to a federal dataset that includes all crashes requiring a tow truck – many of which do not involve airbags at all [2]. When researcher Marco Benedetti applied an apples‑to‑apples comparison (airbag crashes versus airbag crashes), the safety advantage dropped to roughly three times, and even that figure is skewed by the age gap between Tesla’s average 4.1‑year‑old fleet and the 12.8‑year‑old U.S. fleet [2].

Ten of the eleven traffic‑safety experts who reviewed Tesla’s methodology concluded the statistics amount to “misleading marketing” rather than a rigorous safety investigation [1]. Carnegie Mellon professor Phil Koopman likened the comparison to “my jet airplane is faster than your World War II bomber,” underscoring the distortion created by mixing disparate vehicle populations [1].

Labelers see failures on the ground

In Tesla’s Utah office, hundreds of labelers review video from vehicles running FSD, annotating both successful and problematic maneuvers. Former employees described frequent failures: the system often did not brake for emergency vehicles, missed school buses loading children, and sometimes collided with animals or concrete walls [1]. A special “trauma team” in Palo Alto was tasked with reviewing near‑misses involving pedestrians, and they reported instances where drivers had to intervene at the last second to avoid hitting children in crosswalks [1].

Seven of nine former labelers said they would not trust FSD to drive them, with one stating he would not ride in a robotaxi “if you fucking paid me” [1]. A veteran self‑driving engineer who had reviewed Tesla crash data for years described the safety claims as “bullshit” and warned against trusting Musk’s statements [1].

Why it matters

The discrepancy between Tesla’s public safety narrative and the internal reality highlighted by Reuters raises questions about the readiness of FSD for broader deployment and the adequacy of regulatory oversight. As Tesla continues to market FSD as a near‑autonomous feature while requiring driver supervision, investors and regulators must weigh the inflated safety metrics against documented on‑road failures. The company has not responded to Reuters’ inquiries, leaving the gap between marketing claims and operational performance unresolved.