There are high hopes among Tesla investors about the imminent scaling of its driverless robotaxi fleet by year-end, as they think this will justify a higher valuation for Tesla when it merges with SpaceX (which is mere speculation). Elon Musk has been heralding the advent of Tesla self-driving cars to no avail for over a decade, but he stepped up his hype in 2024, as weaker car sales caused profits to plunge by 42% in the 12 months through Q1 2024, with the stock price plunging by 52% to $142. And Musk’s hype was successful: The prospect of a nationwide network of robotaxis, with Tesla owners able to profit by pooling their vehicles in it, has pushed Tesla’s valuation to nose-bleed levels: Despite having only 2% share of the global auto market, its market cap is 58% more than the combined value of the top 15 carmakers, which account for 90% of global car sales (see Figure 1).Figure 1: Tesla is Worth 58% More Than Top 15 Rivals CombinedSource: Bloomberg. And Tesla continues to keep this dream alive with periodic videos of its unmanned robotaxis in Texas, and highly misleading Twitter posts about Full Self-Driving (FSD), Tesla’s inappropriately named driver-assist system (see Figure 2). Figure 2: Tesla’s Misleading Promotions of FSDSource: Twitter. This belies the mounting FSD crashes, with a 302% surge in airbag-deployed collisions since early 2024, when Tesla released a more “assertive” FSD version and launched a subscription program for more affordable access. The response to this is nothing to brush off: The National Highway Traffic Safety Administration (NHTSA) has 2 separate probes covering up to 3.2 million FSD-equipped Tesla vehicles, due to a surge in crashes where FSD ignored traffic safety laws or failed to see objects in front of it due to the sun’s glare or hazy conditions.A May 28th special report by Reuters (here), based on interviews with 7 former FSD “trainers,” revealed that Tesla’s unsupervised robotaxis in Austin had specialized, mapped routes and AI training to make them appear more autonomous than they really are. The report also had 10 safety experts review Tesla’s FSD Vehicle Safety Report, which claims that FSD is 7x safer than humans (more below), but was deemed as “misleading marketing” after they found several invalid data comparisons. The Reuters special report prompted US Senators Markey and Blumenthal to request answers from NHTSA to 5 aspects of FSD under their oversight, such as whether they’ve investigated Tesla’s safety claims with the actual crash data, and why almost all of it is redacted for the general public. A thorough analysis of Tesla’s Autopilot/Full Self-Driving (FSD) crash data filed with NHTSA supports the allegations by many experts that Tesla skews its reports to make FSD look more favorable than it is, especially on key metrics used to rate the safety of automated driving systems (ADS). The sections below include the key findings from this analysis and are based on months of poring through nearly 470,000 individual data points across nearly 3,600 crash filings that Tesla submitted to NHTSA since 2021: Section 1: How Tesla Hides FSD Crashes and Serious InjuriesAside from redactions, the use of the word “unknown” in critical crash-reporting categories has risen by 9.9x since Tesla began heavily promoting FSD in early 2024. Tesla does this despite stating in its safety reports that its vehicle telemetry monitors and analyzes most categories that they mark as “unknown”. Section 2: Incidents With Suspicious Crash ReportsSerious injury crashes see heavy redactions and data missing, as well as high-speed crashes of 60 mph or more listed with injury levels “unknown”. Section 3: Tesla’s FSD Safety Data Doesn’t Add UpUnder-appreciated fact: Tesla’s own FSD Safety Report claims only 626 major collisions on non-highway roads, while its own crash reports to NHTSA show 722. Conclusions All indications point to Musk holding to his words on the Q1 earnings call regarding Tesla focusing on safety over scale of its robotaxi fleet, implying they too realize the risks of FSD in its half-baked state. AppendixMethodology for Estimating Non-Highway FSD Crashes and Injuries.After seeing all of the redactions and workarounds used by Tesla to conceal unsavory FSD crash data, it’s hard to see Tesla’s display of driverless robotaxis as anything but the most expensive and highly orchestrated stock pump in history. Given that Waymo publishes its crash data for 3rd-party peer reviews, if Tesla truly believes their claim that FSD is 7x safer than humans, one can only ask, “What have they got to hide?” SECTION 1How Tesla Hides FSD Crashes and Serious InjuriesSeven of the former data labelers told Reuters they wouldn’t trust FSD to drive them. “We have all seen it fail,” one said. Another said he wouldn’t ride in a Tesla robotaxi “if you fucking paid me.” One veteran self-driving engineer, who reviewed Tesla crash data for years, called its safety claims “bullshit.” “Definitely,” the engineer said, “don’t trust Elon on this.”Why Tesla’s AI trainers don’t trust its self-driving tech – or its safety stats Reuters, May 28, 2026Tesla uses its rights, on the grounds of protecting “confidential business information”, to have NHTSA redact many details from its monthly crash reports for the public. This mainly includes categories of the crash report that would specifically reveal whether the incident occurred with FSD engaged, or with its older Autopilot “cruise control” system, used only for highway driving. Because of this, Tesla has made it impossible to know the true FSD crash rate, and more importantly, the resulting serious injury and fatality rates—two indispensable data sets in evaluating the safety of autonomous vehicles (AVs). Unsurprisingly, if you go purely by the data that Tesla discloses in their monthly reports to NHTSA, you get results as rosy as Tesla’s claims of FSD Supervised seeing 7x fewer major collisions than human drivers (see Figure 3). Figure 3: Growth in FSD Miles Driven Outpaces Crash and Injury RatesSource: Tesla and NHTSA. Note: All numbers based on non-highway cumulative miles and incidents. Non-highway numbers for airbag-deployed, serious injuries, and fatalities, derived by excluding Highway/Freeway and Unknown incidents from the Roadway Type category. What the disclosed data in Figure 3 doesn’t show, however, is the high number of incidents marked with “Unknown” in key categories of their crash reports, like Roadway Type, Highest Injury Severity Alleged, Subject Vehicle (SV) Pre-Crash Movement, and Subject Vehicle (SV) Pre-Crash Speed. Using “Unknown” Instead of Redactions in Sensitive Crash CategoriesBecause Tesla can’t use its right to redact data from categories deemed unrelated to “confidential business information”, they’ve resorted to using the word “Unknown” in key categories that would otherwise give analysts a clue as to whether FSD was involved. Tesla has long used the Unknown entry in the critical category of Highest Injury Severity Alleged when reporting crashes to NHTSA, after fatal Autopilot crashes from 2016 began making headlines (since 2024, Tesla has settled out of court in wrongful death lawsuits for 6 Autopilot incidents and 1 FSD fatality). Tesla’s use of unknown to mark injury levels in crash reports has gone from 95% of all entries in 2021 to 99% of all 597 incidents reported this year, as FSD use has soared since 2024, after Tesla slashed prices and gave out 6 promotional 30-day trials through last December. What’s more astounding is the sharp rise of unknown entries in categories of crash reports where Tesla says its vehicle telemetry is constantly monitoring and analyzing ride cycles for enhanced FSD safety: Roadway Type, Subject Vehicle (SV) Pre-Crash Movement, and Subject Vehicle (SV) Pre-Crash Speed (see Figure 4). Because both Autopilot and FSD can be used on highways, it’s impossible to tell which was engaged in a crash report where the Roadway Type field was marked as Highway/Freeway. This is why crashes on non-highway roads are the only indicator of whether a crash was on FSD, and the 3 categories in Figure 4 were once the best indicators until Tesla increased its use of “Unknown” to obscure the data. Here’s why: Only FSD can drive on non-highway and non-rural roads (which would be known by Tesla’s entry in the crash report’s “Roadway Type” field). By unselecting “Highway/Freeway” and “Unknown” in the Roadway Type menu of a crash report, FSD crashes can be identified (see Figure 5). However, the use of “Unknown” in this category has risen from 13% of all entries in 2023 to 45% so far this year, while non-highway disclosures have shrunk from 30% to 23%. Only FSD can make turns, travel the wrong way, back up, or make parking maneuvers (which would be shown in the “SV Pre-Crash Movement” field). The use of unknown by Tesla in this category of its crash reports has risen the most, from 6% in 2023 to 58% in 2026 year-to-date.Over 60% of crashes with pre-crash speeds of 50 mph or lower are on non-highway roads, which implies that FSD was involved, and is disclosed in the “SV Pre-Crash Speed” field of the crash report. While NHTSA doesn’t provide an “Unknown” choice in the menu of this field, Tesla has left it blank in 14% of its 597 crash reports this year, versus only 5% through 2023. Figure 4: Tesla’s Crash Data Riddled With “Unknowns” Source: NHTSA. Note: “SV” stands for Subject Vehicle. Figure 5: Isolating FSD Incidents by Roadway Type Source: NHTSA. Tesla Says it Knows All the “Unknowns” in its Safety ReportIn its FSD Vehicle Safety Report, Tesla literally reveals that its vehicle telemetry allows it to monitor and analyze all of the 3 crash categories mentioned above. Regarding a Tesla’s ability to know different roadway types: “Tesla categorizes roadways into highway and non-highway based on road characteristics including map data routing classes, lane count and controlled access infrastructure.”Figure 6: Tesla’s Safety Report on How Much Its Cars Know Source: Tesla FSD Vehicle Safety Report. Regarding the vehicle’s Pre-Crash Movement and Speed: Tesla says that its own collision-detection methodology structurally requires measuring vehicle velocity at high time resolution for every collision, explaining what triggers their collision detection as:“A change in velocity (Delta-V) of 8 km/h occurring within a 150-millisecond period.”This is virtually a direct, technical admission that their system tracks vehicle speed with sub-second precision at the exact moment of every collision that makes it into their reports. And, finally, there’s this: “Tesla uses the collision telemetry system described in this Vehicle Safety Report to supplement federal reporting obligations, including the SGO.”These statements, which are meant to support Tesla’s FSD safety claims, seem to contradict the thousands of Unknowns that Tesla marks in critical fields of its crash reports that deal directly with measuring vehicle safety. Figure 7 shows how many times Tesla has used the word “unknown” in its crash reports since NHTSA began collecting such data in August 2021. The following should be noted: Unknowns in injury severity reporting drive the opacity: Tesla says it doesn’t share injury data with NHTSA due to airbag deployments being its main indicator of collision severity, but also because of “health-related privacy laws” (Waymo has no such problems in its peer-reviewed crash reports that have 94% fewer serious injury and fatal crashes). Subject vehicle data shows the fastest growth: As explained above, the subject vehicle’s pre-crash movement can often reveal whether an FSD crash occurred on non-highway roads. The recent surge of unknowns in this field could be interpreted as growing non-highway incidents. Unknowns related to the crash partners: Tesla could claim that its onboard sensors don’t have the reach to analyze conditions of its crash partner, but this would also be questionable: Section 2 shows a crash cluster where Tesla redacted almost all details about crash partners in what seems like the subject vehicle crashed into stopped or parked vehicles.Figure 7: Tesla’s Use of “Unknown” in Crash Reports by CategorySource: NHTSA. Finally, it’s worth looking at how Tesla’s high use of “unknown” in its crash reports compares to Waymo, the leader in driverless ride-hailing, with over 3,000 robotaxis in 10 cities doing 500,000 paid driverless rides per week. Figure 8 shows the use of “unknown” entries in various fields among all incidents in the top table, and with serious to fatal crashes in the bottom table (note that ADS reporting requirements are stricter than ADAS, and include minor fenderbenders). The most unnerving fact about Tesla’s reports on crashes involving serious injuries and fatalities is the sudden spike in unknown entries in the pre-crash movement field (1.4x higher), the pre-crash speed field (5.3x), and airbag deployment (2.9x). By contrast, Waymo goes from having a few unknown entries in its all incident count, to zero when it comes to crashes with serious injuries and fatalities (a review of Waymo’s narratives on these incidents strongly suggest that Waymo was not at fault in any of them). Figure 8: Tesla vs Waymo: Use of “Unknown” by CategorySource: NHTSA Level 2 ADAS Incident Report & ADS Incident Report.Video 1: FSD Tries to Drive into a Closed Highway OfframpSource: Twitter. Video 2: FSD Drives Towards Oncoming Car While Trying to PassSource: Twitter. SECTION 2Incidents With Suspicious Crash ReportsBelow are 3 different crash clusters based on reports that Tesla filed with NHTSA, where the level of redactions and use of “unknowns” in key reporting categories seems almost like a cover-up. Two are small datasets, but unnerving nonetheless, while the last one is a fairly large sample. Looking at the details of Tesla’s crash reports raise questions of how NHTSA could be comfortable with such opaque entries when some of these incidents involved many serious injuries and fatalities. Blank Fields and Redactions in Reports for Severe CrashesThe 24 incidents in Figure 9 may be a small sample group, but the slew of blank fields, unknown entries, and redacted details in the crash reports coincides with the highest rate of reported serious injuries and fatalities within the total 3,529 incidents on file. Here are the key details: Any indication of speed is left blank: Note that 100% of the entries in the categories of Posted Speed Limit and Pre-Crash Speed (which Tesla says its systems always monitor) are left blank. Leaving a blank field in these 2 categories is the same as an “unknown” entry, as both fields don’t have that option in their menus, and both can’t be redacted. Reported injuries and fatalities are 48x higher than average: Given that 93% of the highest injury severity categories are marked as “Unknown” by Tesla, it’s noteworthy that all 24 incidents in this group reported injury levels: 17 fatalities, 6 serious injuries, and 1 moderate injury. The crucial “Serious Injuries or Worse” benchmark used by Waymo and AV safety researchers comes to 96% of this group, which is 48x higher than only 2% for the rest of all 3,505 incidents Tesla has ever filed with NHTSA. Higher than usual incidence of “unavailable telematics and EDR data”: Not to say that these crashes are the same, but it should be noted that a jury in Miami federal court found Tesla liable last year for $243 million in an Autopilot wrongful death lawsuit, after finding that Tesla had concealed crash data that it told investigators it didn't have. Among the 24 crashes in Figure 9, Tesla reported no data available from telematics nor event data recorders (EDR) in 9 of the 17 fatal crashes (53%), which is a staggering statistic, as there is only 1 other fatal incident also without such data, making this group 48x above the baseline. Serious injuries: Of the 12 serious-injury crashes, 50% had neither telematics nor EDR data available, compared to 18% for all other serious-injury incidents—a 2.8x greater showing compared to the baseline.Pedestrian collisions: All 4 pedestrian crashes in this cluster had neither telematics nor EDR data available. Two were fatal crashes, with 1 serious injury and 1 moderate injury. Tesla’s collisions with Vulnerable Road Users, or VRUs (pedestrians, cyclists, and motorcyclists), show higher levels of opacity than normal. Other Noteworthy Characteristics:Roadway Type: 50% of entries say “unknown”, versus the average 30%.Crash With: 42% are unknown, versus the average 23%.Posted Speed Limit: 100% of entries were left blank versus the average 11%.SV Pre-Crash Movement: 71% of entries say unknown, versus the average 32%, with 2 incidents involving crashes with pedestrians being redacted.Airbags Deployed: 71% unknown entries, versus the average 3%. This is extremely high for airbag deployments, which Tesla describes in its Safety Report as proxies for collision severity.The Terrible 10: Note the red-highlighted dates in the Incident Date column (numbers 9 to 18), which all happened nearly back-to-back during 2022, with 3 in May. FSD was still in beta mode during this period, but growing quickly from 100,000 testers in April to 160,000 in September. Redaction of ADAS engagement: In the category asking whether the ADAS was engaged at the time of the crash, half of these incidents have the entry redacted, which usually only happens in 3% of all incidents. Urgent reports: The Report Type column shows that all of these incidents were either 1-Day reports for crashes that involved fatalities, hospitalization, or vulnerable street users (pedestrians, cyclists, and motorcycles), or 10-Day Updates for such incidents. Figure 9: Serious/Fatal Crashes With Speed Data Left BlankSource: NHTSA. Redacting Crash Partner Details and Hiding Injury LevelsFigure 10 shows a cluster of 22 incidents, where one can only conclude that Tesla felt the need to redact what the subject vehicle crashed into, as it was most likely a stopped or parked car. Here are the significant points about this crash cluster: