Tesla has spent years promising that its Full Self-Driving software will transform personal transportation and launch a robotaxi network worth hundreds of billions. The company’s stock price reflects that optimism. Yet a recent investigation paints a different picture. Insiders who trained the artificial intelligence behind FSD say they have seen the system fail too often to trust it with their own lives.
Reuters spoke with nine former data labelers and one veteran self-driving engineer. Seven of the nine labelers said they would not let FSD drive them. One put it bluntly. He wouldn’t ride in a Tesla robotaxi “if you fucking paid me.” Another simply stated, “We have all seen it fail.” The engineer, who reviewed crash data for years, called the company’s safety claims “bullshit.” He added, “Definitely, don’t trust Elon on this.”
These accounts strike at the heart of Tesla’s valuation. The automaker’s $1.4 trillion market value rests in part on the belief that unsupervised autonomy sits just around the corner. Musk has repeatedly said the cars will soon operate without human monitors. In one earnings call he suggested drivers could soon text while the system runs. Yet the people closest to the training data express no such confidence.
The skepticism extends beyond anecdotes. Tesla publishes safety statistics that claim FSD is up to 10 times safer than human drivers. The company compares miles between airbag-deployment crashes in its vehicles against federal data on all crashes reported to police. That methodology produces impressive numbers. It also invites sharp criticism.
Airbag events represent only the most serious impacts. Federal statistics include far milder incidents that require a tow truck but never trigger an airbag. One analysis adjusted for that difference and found the advantage shrinks to roughly three times safer. Tesla’s fleet also skews much newer than the average American vehicle. Newer cars crash less often regardless of driver-assistance features. The company does not publish peer-reviewed studies. Ten out of 11 traffic-safety researchers consulted by Reuters described the statistics as misleading marketing.
Carnegie Mellon professor Phil Koopman captured the flaw in simple terms. “It’s like saying: ‘My jet airplane is faster than your World War II bomber.’ Yeah, so, what’s your point?”
Data labelers spent their days reviewing footage from customer vehicles. They saw FSD-equipped cars speed by 20 or 30 miles per hour over limits after the introduction of what some called “Mad Max” mode. One clip showed a car traveling 60 mph in a 25-mph zone. Others captured vehicles striking animals without braking beforehand. Near-misses with children appeared frequently enough that Tesla formed a specialized “trauma team” in Palo Alto to annotate pedestrian incidents.
The failures went beyond speed and animals. Systems failed to pull over for emergency vehicles. They gave motorcyclists too little room. On freeway off-ramps some refused to brake and slammed into concrete walls. Construction zones proved especially difficult. Cars entered work areas and nearly hit workers. School buses were sometimes ignored. Pedestrians in crosswalks went unrecognized until drivers grabbed the wheel at the last second.
But the company maintains its own narrative. Tesla’s website highlights billions of miles of real-world data and claims the software improves safety. Its latest Spring 2026 software update added AI features, a redesigned self-driving interface, and new safety tools. Owners testing version 14 releases in May reported smoother city driving and better visualization. Some enthusiasts on X described long road trips completed with minimal intervention. Positive owner experiences continue to surface daily.
Those successes, however, occur under supervision. Tesla still requires drivers to remain attentive. Regulatory bodies have taken notice of the gap between claims and performance. The National Highway Traffic Safety Administration maintains multiple open investigations into FSD and Autopilot. A $243 million verdict against the company in an Autopilot-related crash added pressure. Data turnover to investigators has reportedly been slow.
Robotaxi events have followed a similar pattern of careful preparation. When Tesla launched limited service in Austin in June 2025, the operating area had been extensively mapped. Hundreds of labelers in Utah worked overtime to annotate hazards and road details before the debut. The Cybercab unveiling in October 2024 relied on nighttime testing at a controlled studio lot. Prototypes drove mapped routes with human safety drivers. Later expansions to Dallas and Houston left some passengers waiting long periods and unable to be dropped exactly at their destinations.
Musk has projected rapid growth in unsupervised vehicles throughout 2026. He expects monitors to disappear in Texas cities first, then spread. Yet the insiders interviewed by Reuters suggest the underlying technology still struggles with basics that human drivers handle without conscious effort. One former labeler recalled clips of vehicles heading toward lakes, off bridges, or into the paths of oncoming trains. These were not everyday occurrences. They appeared often enough to erode faith.
Independent developers and competitors have taken different paths. Waymo operates driverless vehicles in multiple cities but relies on lidar, high-definition maps, and remote assistance. Its safety researchers emphasize precise language when discussing performance. John Scanlon, a Waymo safety expert, said companies must “be really careful with the language we use.” Tesla bets entirely on vision and neural networks trained on fleet data. That bet has produced rapid software gains. It has also left gaps that labelers see in raw footage every shift.
Recent coverage echoes the tension. Electrek called the statistical inflation “not subtle” and the insider perspective “devastating.” The site noted that labelers witness failures by design. Their job requires tagging errors so the system can improve. Even so, the volume and severity left many unwilling to ride. Futurism framed the story as evidence that “not even the people building Tesla’s self-driving tech trust Elon Musk’s extravagant claims.”
Tesla has not commented publicly on the specific interviews. The company continues to roll out updates and collect data. Owners who subscribe to FSD or buy the feature outright report varying degrees of satisfaction. Some describe it as the best driver-assistance system on the market. Others share videos of unexpected maneuvers that require quick human correction.
The divide matters for more than bragging rights. If FSD cannot reach reliable unsupervised operation at scale, the robotaxi business case weakens. Regulatory approval for driverless commercial service becomes harder to obtain. Insurance implications shift. And the lofty stock valuation faces fresh questions. Musk has overcome skepticism before. The gap between internal observations and public forecasts, however, has rarely been documented by so many former employees in their own words.
So the technology improves. New versions handle city streets with greater poise. Yet the people who watched terabytes of its mistakes still hesitate. Their reluctance does not prove the system will never succeed. It does suggest the remaining challenges run deeper than quarterly software notes admit. Investors, regulators, and potential robotaxi riders will weigh those voices against the optimistic roadmaps. The data labelers have seen the unfiltered record. For now, many choose to keep their hands on the wheel.
WebProNews is an iEntry Publication