Special Report-Why Tesla AI trainers do not trust self-driving technology or safety statistics

In an office in Utah, hundreds of Tesla employees scrutinize the video footage collected by cars using the Full Self-Driving feature (FSD). In some clips, the cars are shown hitting deer or cats. Other clips show more common accidents. They sometimes don't stop before hitting. They speed up a lot. Sometimes, workers witness children playing on the street.

These Tesla staffers, also known as "data labels," train the AI-powered Tesla driver-assistance system. They note incidents of good driving and bad driving, and report problems to engineers who are working to improve the software.

Elon Musk, Tesla's CEO, says that FSD will "soon" make all Teslas autonomous. Interviews with nine former Tesla labelers and an ex-self-driving engineering show that in recent months, the technology has struggled to perform basic maneuvers such as avoiding emergencies or stopping to let school buses load or unload students. (View the article on.com:)

Musk and other executives continue to tout the safety of FSD, despite its dangerous flaws. They have pushed Tesla into staging public demonstrations of the fully automated capability that the CEO has "promised" investors each year for the past decade. Displays include a robotaxi in Austin, Texas that was launched in June last year with human safety monitors inside the cars as well as others who worked remotely.

Four former Tesla employees said that as the events neared, Tesla staffers spent long hours mapping routes, and training software to recognize specific hazards, in order to make self-driving cars appear more capable. Staffers claimed that these labor-intensive safety measures are not feasible to implement on a large scale. These efforts, which were not previously reported, undermine Musk’s claim that Tesla’s self-driving tech will soon work globally, and won't need the laborious mapping of local roads and hazards used by competitors. Musk said Tesla relies on AI and cameras to simplify its approach. This will allow Tesla to expand its robotaxi service with "hyperexponential speed" and give current Tesla owners complete autonomy via software updates.

Musk and other Tesla leaders have reinforced the impression of robocompetence by citing statistics about company safety that, they claim, prove FSD to be up to 10x safer than human driving.

A? A?examen of Tesla's statistical method and interviews with?company?insiders shows Tesla isn't near delivering safely self-driving cars at scale, a promise that underpins the automaker's stock-market valuation of $1.6 trillion. This included an analysis of how Tesla compares their own crash data with federal crash data, a comparison of the more rigorous methodology used by robotaxi rival Waymo and interviews with eleven traffic-safety experts who reviewed Tesla's method. The review identified several invalid data comparisons that were used to support the statistics in Tesla’s FSD safety reports. Ten researchers claimed this was misleading marketing, rather than an investigation into a crucial safety issue.

Tesla, for example, exaggerates safety of the technology by comparing a crash rate in FSD-piloted Teslas which triggered airbag deployments with a federal accident rate for all vehicles, including far less severe accidents. Tesla also compares their cars with the average U.S. car, which is older than the typical Tesla. Researchers said that this results are distorted because automakers all have launched safety features to reduce accidents.

"Any new vehicle is significantly safer than an old car," said Phil Koopman. He's a Carnegie Mellon University professor of engineering and an expert in autonomous vehicles. It's like saying, "My jet plane is faster than your World War II Bomber." Yeah, then what's the point?

Tesla did not respond to the detailed questions for this report. Tesla's CFO Vaibhav Tania first claimed that the robotaxi was 10-times safer in July last year, following Tesla's Austin launch. Robyn Denholm, Tesla Board Chair, repeated the claim at a meeting in November where shareholders approved an incentive package that would grant Musk up to $1 trillion worth of Tesla stock. Musk displayed at the same meeting a chart that claimed "85% fewer crashes" based on the recently revised Tesla method.

Musk said to shareholders that he felt "almost comfortable" allowing drivers to text while driving. In the next few months, we'll be closely examining safety statistics. But we will let you text while driving.

Tesla still hasn't approved texting while driving using FSD six months after it was first announced. The fine print of its FSD site continues to warn that "currently enabled features require active supervisory supervision and do NOT make the vehicle autonomic." Tesla has often cited this disclaimer when sued for serious accidents.

FSD has been widely recognized as a technology capable of handling a wide range of driving situations and can do so for extended periods. Full autonomy is still elusive to Tesla and other car companies because it requires flawless technology execution, even in the most complicated driving situations.

Seven former data labelers said they would not trust FSD to drive them. One said, "We've all seen it fail." One man said that he would not ride in a Tesla Robotaxi "if you paid me." Another self-driving expert, who has studied Tesla crash data over the years, called Tesla's safety claims "bullshit."

The engineer replied, "I don't think you can trust Elon with this."

'TRAUMA TEAM REVIEWS NEAR MISSES

Tesla's data labels get an up-close look at FSD capabilities by reviewing footage from eight exterior cameras equipped vehicles.

Former employees said that FSD failed to perform basic tasks such as stopping for emergency vehicles or giving motorcycles enough room. They saw FSD vehicles failing to brake at freeway exit ramps. One Tesla even hit a concrete barrier. The footage, according to them, did not show if anyone was injured. Two employees claimed that clips showed FSD not avoiding construction zones. One of these incidents involved a Tesla driving into the construction zone and nearly hitting workers, according to one person.

This account is based on the descriptions provided by former staffers of the footage they saw.

Managers at Tesla carefully control access to videos. Employees only watch the clips that they are assigned. They may or may be able to see FSD's biggest failures.

Three employees reported that one data-labeling group focused on pedestrians who were close to being hit. One source claimed that the employees were known informally as "the trauma team" and worked in Palo Alto in California with special access to the footage. The trauma-team footage was closely guarded by engineers, but sometimes, some footage "slipped through" to the other teams, according to the source.

Both the person and an employee claimed to have seen clips of drivers taking control manually at the last moment when FSD failed recognize pedestrians at crosswalks. Two former employees remembered seeing videos of FSD-piloted Teslas almost hitting children last year.

Tesla has faced federal investigations, lawsuits, and other legal proceedings for many years involving fatal crashes that were blamed by drivers or regulators on the failure of FSD, or its older Autopilot system. In 2021, the U.S. National Highway Traffic Safety Administration opened an investigation on Autopilot after a number of accidents involving Teslas hitting emergency vehicles. In 2023, the investigation led to a recall where Tesla installed software updates to detect when drivers are not paying attention.

NHTSA is currently conducting four investigations on FSD and Autopilot. One of these involves dozens cases in which vehicles using FSD did not stop at red lights or turned towards traffic. A second investigation examines whether Tesla's Autopilot upgrades for 2023 were adequate to solve the safety issues. The agency is investigating nine FSD incidents, including one fatal crash where the system failed due to reduced visibility, such as in fog or sun glare. Tesla was awarded $243 million last year after an Autopiloted Tesla crash in Florida killed a woman aged 22 and seriously injured her boyfriend. Tesla has appealed. The company settled similar cases without disclosing the terms in several other similar cases.

NHTSA did not respond to a question from NHTSA about the safety of FSDs and Tesla's method when asked by NHTSA. The agency referred any questions regarding Tesla's safety claims over to the U.S. Federal Trade Commission.

The FTC refused to comment on Tesla’s safety statistics. Consumer advocacy groups and U.S. Senators have asked the FTC to look into Tesla's marketing for Autopilot and FSD.

Tesla has not been subjected to any enforcement action by the FTC.

STATS INFLATED ON TESLA'S FSD SAFETY

Tesla's CEO and board boosted their claims of safety as employees watched the videos. Tesla's leaders promoted this 10-times safer claim for much of the last year.

Tesla CFO Taneja stated in a call to discuss earnings that "a car with FSD will be 10x safer". "Even with $99 per month, you get a personal driver for less than $3.33 per day."

Tesla's method is flawed because of a comparison error. This led to a three-fold increase in the claimed level of safety. The automaker compared data from all crashes involving a tow truck with the number of Tesla crashes that had airbags deployed. This was a much less strict criterion. Tow truck crashes are often not severe enough to activate airbags.

Tesla used a flawed methodology, even though the federal data that it used included crashes in which airbags were deployed. The flawed methodology led to the conclusion that Teslas with FSD or Autopilot traveled 10 times further between crashes than an average human driver.

According to Marco Benedetti's analysis, a former NHTSA statistician and assistant research scientist from the University of Michigan Transportation Research Institute, Teslas with driver-assistance systems are three times more likely to travel between crashes in which airbags were deployed. Two other traffic safety researchers confirmed Benedetti’s calculations.

Benedetti explained that Tesla's method is flawed in several ways, and this does not mean FSD makes a driver three times more safe than an average driver.

Tesla changed its approach to only include data from vehicles that use FSD in November and excluded those using Autopilot. Tesla claimed more miles between crashes by including Autopilot because it is a less complex system that's designed for highways, where cars travel fewer miles and crash less often than in urban areas. On its website, the company continues to cite the flawed airbag crash comparison to claim that FSD is 7 times safer than an average driver. This amounts to 85% in percentage terms.

The findings of a study cast doubts on the safety of FSD.

When comparing its crash rates to those of the nation, the automaker does not take into account the age of the vehicle. Tesla compares the average age of its cars, which is just 4.1, with that of all U.S. vehicles which averages 12.8. Ten safety researchers said that this skews results because most automakers are now offering revolutionary safety features across their entire lineup, such as blind-spot detection and automatic emergency brakes.

Tesla reduces the number of crashes by counting only those that occur with FSD on or within 5 seconds after it is turned off. The U.S. Government, on the other hand, requires automakers report crashes that occur within 30 seconds after an advanced driver assistance system is deactivated.

Tesla claims FSD can save over 32,000 lives each year and prevent 1.9 million injuries. Researchers in traffic safety have called these figures meaningless, because they assume that all U.S. vehicles, including trucks and motorcycles that are crash-prone, will be replaced with Tesla cars that support FSD, and that each Tesla is at least seven-times safer than its predecessor.

The More?Rigorous Approach of WAYMO

Ten traffic safety researchers have said that the premise behind Tesla's statistics on safety is flawed as well, because FSD isn’t a true autonomous system. Tesla's executives are wrong when they claim that the company is comparing its technology with human drivers. Tesla isn't comparing its technology to human drivers, as executives claim. Instead, it compares the average driver with another human who drives a Tesla and uses FSD. Tesla fails to take into account that these drivers are able to turn FSD off and on. Research shows that motorists avoid advanced driver-assistance technologies in traffic situations with high levels of complexity, as they feel unsafe. Tesla's data indicates that FSD is mostly used on highways.

Alphabet Waymo compares its driverless robotaxis in 11 U.S. metro areas to similar human-driven cars.

Waymo adopts a more rigorous strategy than Tesla. It examines crash data from the markets in which it operates, and adjusts for the type of roads and neighborhoods that its robotaxis travel. Waymo looks at specific crash rates, such as airbag deployments and serious injuries. This is for both human-driven cars and Waymo's cars.

John Scanlon is a Waymo researcher who specializes in safety. "You have to be very specific with your research questions and conclusions."

Waymo, meanwhile, points out flaws in its data as well as collaborating with external researchers to publish its safety statistics in peer reviewed journals. Tesla, on the other hand, does not seek peer review, and only publishes top-line statistics about safety, while keeping Tesla's crash data secret.

VIDEOS OF TESLAS STRIVING DOGS, CATS AND DEAR

Data labelers at Tesla get a raw look at FSD safety. Three former Tesla employees have described videos of Teslas hitting animals at high speed without braking.

Five former employees stated that specific teams were focused on FSD’s difficulties recognizing school busses. The Dawn Project, a technology safety group, raised this concern by airing ads at the Super Bowls of 2023 and 2024 showing Teslas with FSD enabled failing to stop for school buses with flashing lights and stop signs.

Two former Tesla employees have said that they saw videos similar to those in Tesla.

Five former data labelers have described a chaotic, hurried work environment, where priorities were constantly changing based on Musk's and FSD's engineers' directives. They said that the data-labeling department was plagued by a high turnover rate due to monotonous work and low pay.

Four former employees claim that Tesla's higher-ups would often launch new projects as a reaction to news stories or social media posts that showed FSD making mistakes. One project was to fix the problem of sunlight obscuring exterior cameras on cars. A social media video showed how the light reflected off a watch of a passenger blinded a camera, shutting FSD down. A second effort was made at railroad crossings after news reports that Teslas with FSD failed to stop there.

Five of the employees also said that FSD clips regularly showed speeding. Engineers and other people higher up in the chain treated this as a problem with low priority.

After the automaker launched a "Mad Max mode" that allows for more aggressive driving, one employee reported labelers regularly seeing Teslas exceeding speed limits of 20-30 miles per hour. One labeler said he saw a vehicle driven by FSD traveling at 60 mph within a 25 mph zone.

TESLA PUBLIC ROBOTAXI DISPLAYS: BEHIND THE CURTAIN

Musk announced Tesla's self driving capabilities in October 2024, as Tesla employees struggled with FSD training. The robotaxi was unveiled at Warner Bros. Studios near Los Angeles. Musk showed off about 20 two-door prototypes of his "Cybercab" which had no pedals or steering wheel, to the invite-only audience.

He said, "The cars just pass by without any people."

Musk said Tesla's software was designed to navigate unfamiliar landscapes and work anywhere. Two former data-labeling workers claim that for the weeks leading up to the Cybercab launch, the staff collected video footage of the route the cars would take at the event. The employees claimed that labelers spent hundreds hours on video annotating road markings and curbs to avoid embarrassing incidents.

Musk has dismissed this approach as being too expensive and slow. Musk called Waymo's'very localized solutions' "quite fragile" in 2024.

Musk announced on a call for earnings in January 2025 that Tesla will launch robotaxis in Austin in June 2025, after the Warner Bros. presentation. He hailed the technology as "a generalized AI solution," which didn't need "high precision maps of a specific location."

Tesla filmed stop lights, road signage and other features for months prior to the Austin launch. Two employees who have direct knowledge of this matter say that data labelers marked up the video to make sure the software was capable of handling challenging scenarios. These included passenger pickups and responding to emergency vehicles.

Three employees claimed that the Utah data-labeling team doubled in the half-year prior to the Austin launch, from about 100 workers to 300. They said that the department worked on a variety of projects in order to ensure that the Austin test went smoothly.

Two employees reported that the software was still unreliable as Tesla data labelers were preparing for the "rollout". Some driving behaviors improved with each FSD upgrade. Some driving behaviors improved with each FSD update. Two large screens in the Utah office displayed statistics on FSD, a key safety metric for autonomous driving.

Former employees stated that the company's performance was "like the stock market", with no consistency in improvement.

Two sets of human safety monitors are available for the vehicles to hit the road: one in the front passenger's seat and another watching remotely. Labelers in Utah could watch videos to see when remote monitors were taking control of the vehicles. Former employee stated that the Austin routes were created for a small area to allow the software of the cars to be extensively trained on specific maneuvers in particular streets.

The person who said this described it as "OK, we taught a car" how to operate within a restricted area. "You can't be creative outside of this."

Four sources stated that scaling up in a safe manner could take many years. Musk said in July that the robotaxi service would be expanded to cover half of the U.S. by 2025, just a month after its launch in Austin. Musk claimed falsely in January that Tesla operated 500 “robotaxi vehicles” in Austin and San Francisco Bay Area. He added that he expected this to "double each month" along an "exponential" curve.

According to a recent presentation by officials, Tesla operates less than 50 robotaxis in Austin, despite the launch of the service nearly a full year ago. Three sources confirmed that the vehicles travel within a carefully-mapped and limited zone. According to recent observations made by a journalist, some vehicles still have human safety monitors on the front passenger's seat.

Tesla announced in April that it would be launching robotaxis throughout Dallas and Houston. It also provided maps of the areas covered. Reporters who tested the service recently in both cities reported long wait times and inconsistent availability. When a reporter was able to get a ride three times in Dallas, the robotaxi would not drop him off at his destination downtown within Tesla's service area.

Every time it took him 15 minutes to walk. (By Chris Kirkham, Los Angeles; and Rachael LEVY, Washington. Sheila Dang, Norihiko Schirouzu, and Benjamin Lesser contributed additional reporting. Brian Thevenot, David Crawshaw and Brian Thevenot edited the story.

(source: Reuters)