In Utah, hundreds of Tesla employees review video materials collected by cars during the operation of the full self-driving (FSD) system. The footage shows instances where the cars encounter animals – cats, dogs, or deer – along with more routine crashes. Sometimes they do not brake before impact, often at high speeds. There are also near-misses with children playing on the street.

They are called “data labelers”: they train the driver-assistance artificial intelligence by labeling examples of good and bad driver behavior and passing issues to engineers who improve the system.

The chief executive Elon Musk says that FSD will soon make all Tesla vehicles fully autonomous. But surveys of nine former data labelers and a former autonomy engineer reveal that over the past few months the technology still struggles with basic maneuvers – such as avoiding emergency vehicles or stopping in front of school buses that are loading or unloading students.

Despite such risks, Musk and other executives have increasingly emphasized the safety of FSD, while promoting public demonstrations of full autonomy that the company has promised investors for years. Such demonstrations included a pilot robotaxi in Austin, Texas, launched last summer with several internal safety monitors in the car and remotely operating specialists.

Inside Tesla, according to former employees, in the run-up to such events staff worked long hours, mapped routes, and trained the system on specific hazards to make it seem more “operational” in the eyes of the public. They noted that such labor-intensive safeguards cannot be rolled out at a broad scale.

Such efforts contradict Musk’s assertion that Tesla’s autonomous systems will soon be able to operate anywhere without the need for detailed local path mapping and hazard detection. Musk said the company favors a simple approach – relying on cameras and artificial intelligence – that would allow rapid scaling of the robotaxi and offer full autonomy to Tesla owners via software updates.

Musk and other executives have heightened the emphasis on safety – they cite their own statistics as evidence that FSD is already ten times safer than human driving.

Field investigations and interviews with insiders reveal that Tesla does not adhere to a transparent and reliable methodology for comparing its data with other sources, casting doubt on the scale of safety claims as the system scales to full autonomy. Consequently, comparing accidents of the upgraded system with overall US accident rates and using airbag data may create a certain interpretive “advantage” for advertising rather than for rigorous research.

“Any new car is safer than an old one – it’s like comparing a jet aircraft to a World War II bomber. And what’s the point?”

«This is like saying: “My jet aircraft is faster than your World War II bomber.” Yes, what’s the point?»

– Phil Koopman

Tesla’s chief financial officer, Vaibhav Taneja, first stated that FSD is ten times safer last July after the Austin robotaxi launch. Chairman of the board Robyn Denholm repeated this at the shareholders’ meeting when pay packages were approved, and Musk presented another diagram showing “85% fewer accidents,” based on updated methodology.

“We almost feel comfortable letting people write messages behind the wheel, which is the so-called “killer app,”” Musk said. “In the next month or two we will consider safety closely, but we will allow writing and driving in practice.”

Six months later, Tesla still has not given the green light to text messages while driving; warnings on the FSD website continue to remind that “active features require active driver supervision and do not make the vehicle autonomous.” The company cites such cautions in lawsuits over serious crashes.

Although FSD can operate confidently in many scenarios, full autonomy remains an unattainable dream for Tesla and other market players, as it requires flawless performance of the system in the most challenging driving situations.

Seven former data labelers said they do not trust FSD to drive them on their own. “We’ve seen it doesn’t always perform reliably,” one said. Another added that he would not ride in a Tesla robotaxi “even for money.” One veteran engineer who has analyzed Tesla crashes for years called the company’s safety conclusions “lies.”

“Absolutely,” the engineer said, “don’t trust Elon with this.”

The data-labeling workers reviewing footage gain the ability to see clear frames of how FSD fails to recognize pedestrians or fails to react to them in crosswalks, leading to near-misses or close calls. Sometimes the footage also involves school buses. The actual video is not always accessible to all divisions – sometimes fragments of the “megaped” remain locked under internal controls.

The aforementioned workers also reported that handing over footage to other teams did not always happen, and the selection of material was sometimes driven by external disturbances or the need to respond quickly to public backlash.

The analysis notes that some frames show drivers manually taking control at the last moment when FSD fails to recognize pedestrians at crosswalks. Real-time cases were also mentioned when FSD did not notice students at bus stops or did not slow down on highways, which occurred during shifts. Reuters did not view the videos themselves, but this information is based on descriptions from former employees.

Access to video materials at Tesla is tightly controlled: workers see only the frames assigned to them, so they may not see the system’s worst mistakes. There are special teams of labelers who have broader access, but they are also limited in their functions.

According to former employees, the work atmosphere was tense and fragmented: priorities shifted depending on guidance from Musk and the FSD engineers, and turnover was high due to the monotony of the work and low pay. The company routinely launched new projects in response to public disclosures or posts behind the scenes on social media.

«One of the data-labeling teams saw Teslas regularly exceed the speed limit by 20–30 miles per hour after the introduction of the ‘Mad Max’ mode for more aggressive driving.»

They also mentioned efforts to mitigate how sunlight can interfere with exterior vehicle cameras, or work on railroad crossings that surfaced after public news of FSD errors. Footage of safety violations was often considered lower priority than other tasks.

After the Austin rollout the company announced plans to expand the robotaxi to Dallas and Houston, but observers noted long wait times and limited availability in new areas. Delays and detours demonstrated that scaling the system requires substantial effort and time.

Overall, the investigation points to the need for Tesla to improve its approach to hazard assessment and data transparency to truly back up safety claims and the pace of moving toward autonomy on a global scale.

In short, Musk’s ambitions for full autonomy collide with reality: for now, FSD requires active oversight and adherence to strict safety requirements, and the path to large-scale deployment remains a lengthy process that demands a far more transparent and reliable methodology.