Tesla’s self-driving story has a new problem: some of the people who helped train the system reportedly do not trust it enough to ride in it.

Tesla has spent years telling customers and investors that Full Self-Driving is moving toward a future where cars can drive themselves at scale. The latest concern is not coming from a rival automaker or a skeptical regulator. It is coming from former workers who reviewed the system’s mistakes, labeled its training data and saw the awkward edge cases that marketing language tends to smooth over.

According to a Reuters investigation published May 28, seven of nine former Tesla data labelers interviewed said they would not trust the company’s Full Self-Driving system to drive them. Reuters also interviewed a former self-driving engineer and 11 traffic-safety researchers, focusing not only on the technology itself but also on how Tesla presents its safety record. That matters because Tesla is not selling a normal driver-assistance feature. It is selling the central idea that its cars, software and AI approach can become the foundation for a robotaxi business.

The gap between those two stories is now the issue. Tesla calls the current product Full Self-Driving (Supervised), and its own support material says the system requires active driver supervision and does not make the vehicle autonomous. Yet the broader pitch around Tesla has increasingly leaned on autonomy, with Elon Musk continuing to frame robotaxis and AI as the future of the company. When the people who trained the model describe hesitation, it becomes harder to treat this as a simple rollout problem.

The Reuters report also raised questions about Tesla’s safety comparisons. The company has suggested that FSD can be far safer than human driving, but researchers reviewing its methodology told Reuters the comparisons were not strong enough to support that claim. One concern is that Tesla’s figures compare different kinds of crash data, including more severe FSD incidents against broader federal crash statistics, which can make the system look better than a like-for-like analysis would.

This is not a small accounting dispute. Safety statistics are doing heavy work in Tesla’s story. They help persuade customers to pay for the software, encourage investors to value Tesla as an AI and robotics company, and give regulators a reason to believe wider deployment can be managed responsibly. If those numbers are seen as marketing rather than measurement, the whole argument becomes less sturdy.

The timing is awkward for Tesla because regulators are already looking closely at the system. In March, the National Highway Traffic Safety Administration escalated its investigation into Full Self-Driving in reduced-visibility conditions to an engineering analysis. AP reported at the time that the probe was examining nine crashes and could involve about 3.2 million Tesla vehicles. The question in that review is whether the system can recognize when conditions have degraded and give the driver enough time to take back control.

That is where supervised autonomy becomes a hard business problem. A system that fails constantly is annoying, but it keeps the driver alert. A system that works most of the time can invite trust before it has earned it. The danger is not only that the car makes a mistake. It is that the human is asked to supervise something designed to make supervision feel unnecessary.

Tesla’s approach is different from rivals such as Waymo, which use more sensors and tightly mapped operating areas. Musk has argued that Tesla’s camera-based system can scale faster because it learns from a large fleet of customer vehicles and does not need the same kind of local mapping. If it works, the advantage is enormous. If it does not, the same simplicity becomes the weakness.

The former Tesla workers cited by Reuters described the system struggling with basic situations in recent months, including emergency vehicles and school buses. Those are not obscure laboratory puzzles. They are the exact moments when a self-driving system needs to be boringly reliable. For ordinary drivers, one bad intervention can erase months of confidence. For regulators, the bar is even higher because one company’s software mistake becomes a public road problem.

There is also a financial reason this story matters beyond auto safety. Tesla’s valuation has depended heavily on the belief that it is more than an electric vehicle manufacturer. The company’s future multiple rests on software margins, autonomy and robotaxi economics. If FSD adoption slows, if regulators demand more evidence, or if customers begin to treat the product as advanced cruise control rather than a step toward autonomy, that premium becomes harder to defend.

This does not mean Tesla cannot improve the system. AI products often get better through exposure, feedback and iteration, and Tesla has more real-world driving data than almost anyone. But cars are not chatbots. A bad answer in a spreadsheet wastes time. A bad decision at an intersection can injure someone. That difference is why confidence has to come from independent evidence, not only from product demos and executive claims.

The next thing to watch is whether Tesla gives outsiders more detailed safety data as it pushes further into robotaxis and international approvals. If the company can prove that FSD is safer under comparable conditions, with clear definitions and transparent reporting, the argument changes. Until then, the most important signal may be the simplest one: the people closest to the training process are not all willing to be passengers.

Also read: Erin Brockovich puts AI data centers on notice. • OpenAI is hiring its way back into robotics • Wix cuts 1,000 jobs as AI economics start to bite