Tesla's self-driving safety claims face internal skepticism and methodological flaws

**The Gap Between Tesla's Safety Claims and Its Own Workers' Reality**

Tesla's Full Self-Driving software has been marketed to investors and the public with a striking statistic: it is up to ten times safer than a human driver. That figure has been repeated by CFO Vaibhav Taneja on earnings calls, by board chair Robyn Denholm at a shareholder meeting where she helped approve a pay package potentially worth $1 trillion in Tesla stock to Elon Musk, and by Musk himself, who displayed a chart at the same meeting claiming FSD causes 85% fewer crashes than average drivers. At that meeting, Musk also suggested that within a month or two Tesla might allow drivers to text while FSD was active — a threshold the company has still not crossed six months later, while its own website continues to warn that the feature requires active human supervision.

The gap between those claims and the experience of people who work most closely with the technology is striking. Reuters interviewed nine former data labelers — employees based in a Utah office whose job is to annotate FSD footage and flag problems for engineers — along with a former Tesla self-driving engineer. All seven of the former labelers who were asked said they would not trust FSD to drive them. One said he would not ride in a Tesla robotaxi under any circumstances. The former engineer, who spent years reviewing Tesla crash data, described the company's safety claims in blunt, unprintable terms and advised the public not to trust Musk's representations on the subject.

What these workers reviewed on a daily basis included vehicles hitting cats, dogs, and deer without braking; failing to pull over for emergency vehicles; not stopping for school buses loading and unloading children; failing to brake on freeway off-ramps (in one case a Tesla struck a concrete wall); and driving into active construction zones and nearly striking workers. A specialized internal team, known informally among staff as the "trauma team" and based in Palo Alto with elevated data access, focused specifically on near-misses involving pedestrians. Footage occasionally leaked to other teams showing FSD failing to recognize pedestrians in crosswalks, with drivers intervening manually at the last second. Two former employees separately recalled seeing footage from last year showing FSD-piloted vehicles nearly striking children.

---

**How Tesla's Safety Statistics Were Built to Mislead**

The methodology underlying Tesla's headline safety claims contains several layered errors that, taken together, transform a modest and contested performance advantage into a dramatic one — or possibly eliminate the advantage entirely.

The most significant single error involves comparing different types of crashes. Tesla counted incidents in its own fleet where airbags deployed and compared that rate against federal data covering all crashes in which a tow truck was required to remove a vehicle. Airbag deployments are a far more restrictive criterion than tow-truck removals; many crashes serious enough to require a tow are not severe enough to trigger airbags. This asymmetry inflated Tesla's claimed safety margin by roughly a factor of three. The particular absurdity of the choice, noted by multiple researchers, is that the same federal dataset Tesla used already contained airbag-deployment data — meaning an apples-to-apples comparison was directly available and simply was not used.

Marco Benedetti, an assistant research scientist at the University of Michigan Transportation Research Institute and a former statistician at the National Highway Traffic Safety Administration, performed a corrected analysis for Reuters using the proper like-for-like comparison. His finding: Teslas using FSD or Autopilot travel roughly three times farther between airbag-deploying crashes than the average vehicle — not ten times farther. Two other traffic-safety researchers independently vetted his calculations and agreed. Even that three-times figure, Benedetti cautioned, does not mean FSD is genuinely three times safer, because additional methodological problems remain uncorrected.

Among those additional problems: Tesla compares its vehicles, which average 4.1 years of age according to S&P Global Mobility data, against the entire U.S. fleet, which averages 12.8 years old. This matters enormously because recent model-year vehicles across all manufacturers have incorporated technologies — automatic emergency braking, blind-spot monitoring, lane-keeping assistance — that substantially reduce crash rates regardless of whether any self-driving software is running. Phil Koopman, a Carnegie Mellon University engineering professor and autonomous-vehicle safety expert, compared the logic to claiming a modern jet is faster than a World War II bomber and treating the gap as meaningful. Ten of eleven traffic-safety researchers who reviewed Tesla's methodology for Reuters described the company's safety report as misleading marketing rather than rigorous safety analysis.

Tesla also narrows its own crash tally by counting only incidents that occur while FSD is active or within five seconds of being switched off. Federal reporting requirements for advanced driver-assistance systems set that window at thirty seconds — a meaningful difference given that driver behavior and vehicle dynamics can reflect prior system activity for considerably longer. In November, Tesla revised part of its methodology to exclude Autopilot data and count only FSD-active miles, partly because Autopilot, being a highway-only system, accumulated long stretches of low-crash-risk driving that artificially boosted the combined safety figure. The company's website, however, continues to publish the seven-times-safer figure derived from the older, flawed airbag comparison.

---

**Concealed Preparations and the Limits of Scalability**

Beyond the statistical questions, former employees described internal practices that directly contradict Musk's repeated public assertions about how FSD works. Musk has consistently argued that Tesla's approach — relying solely on cameras and AI rather than detailed pre-mapped routes — is fundamentally superior to competitors like Waymo and will allow the company to scale its robotaxi service at what he called "hyperexponential" speed. The premise is that because Tesla does not need to map roads in advance, it can operate anywhere immediately.

Before the Austin robotaxi launch last June, however, Tesla spent months filming the specific routes within the robotaxi zone to build targeted maps of that area's particular hazards, four former employees told Reuters. Staff worked extended hours training the software on specific conditions and edge cases unique to that geography. These preparations were labor-intensive and, critically, are not the kind of work that can be replicated quickly or cheaply across new cities or regions. The employees described the safeguards as fundamentally impossible to deploy at broad scale — precisely the scale Musk has promised.

The Austin service launched with human safety monitors both inside the vehicles and supervising remotely, a configuration that itself reflects ongoing uncertainty about autonomous reliability. Musk had predicted the service would expand to cover half the U.S. population by the end of 2025. Tesla currently faces four active federal investigations into FSD and Autopilot, including one examining dozens of cases where vehicles failed to stop for red lights or turned into oncoming traffic, and another probing at least nine incidents — including one fatality — where FSD failed in low-visibility conditions such as fog or sun glare. Last year a Florida jury returned a $243 million verdict against Tesla after an Autopilot-involved crash killed a 22-year-old woman; Tesla has appealed. Multiple similar cases have been settled confidentially. Neither the Federal Trade Commission nor NHTSA has taken enforcement action against the company's safety marketing claims, despite calls from consumer groups and U.S. senators to do so.