New Tesla FSD safety data
Tesla's marketing claims are so full of holes we can conclude nothing useful
Tesla has just published FSD(supervised) safety data with much fanfare, including claims of “7x fewer major collisions”, “7x fewer minor collisions” and “5x fewer off-highway collisions”. However, by the time you read the very extensive fine print and consider the threats to validity, none of those claims stand up to scrutiny. In short, Tesla has released a document full of marketing puffery, and not a serious safety analysis.1
For this discussion I focus on Tesla’s claims regarding “Miles Driven Before a Major Collision” in North America unless otherwise noted, with the expectation that other categories have comparable issues.
More about Tesla claims and methodology:
Tesla’s Nov. 2025 FSD safety report is here: https://www.tesla.com/fsd/safety (Nov. 14, 2025 archive). Here are some salient points about their methodology:
They acknowledge that the “US Average” is cars more than 12 years old, many of which are missing effective safety features present on new vehicles.
The no-active-safety data is not new Teslas with features turned off, but rather pre-2014 Teslas made before those features were added to their cars.
Data is presented for North America vs. worldwide, and highway vs. non-highway.
The analysis separates highway from non-highway, which is a good start at addressing prior issues with urban vs. rural crash rate differences. But the criterion for classification into “highway” is unclear, simply stating: “based on road characteristics including map data routing classes, lane count and controlled access infrastructure”
Data is reported automatically based on on-board sensors. Severe crash reporting is for the most part triggered by airbag or other one-shot active safety system deployment.
FSD crashes are recorded when FSD is active within 5 seconds of a crash, with Tesla stating: “This approach accounts for the time required for drivers to recognize potential hazards and take manual control of the vehicle.”
Crashes are reported without regard to blame, which is a good practice for a primary claim regarding safety outcomes.
Threats to validity of Tesla’s report:
Tesla’s posted report has numerous threats to validity. The comparison number is overstated compared to other comparable vehicles, and the data analysis is missing so many significant factors.
Newer Teslas vs. Old Average Fleet. Essentially ALL new cars are safer than the US average, because the “average” 12+ year old car is missing important safety features required on all new cars. So a claim that a brand new Tesla chock-full of safety technology is safer than an old used car without that technology is irrelevant to the questions of whether buying a Tesla gets you a safer car than buying some other brand at a comparable trim level, or whether turning FSD on improves safety. You might as well claim that graduates of a particular high school have superior physical attributes due to having come from that particular school because they can run faster than the average population (including people who have not exercised in 30 years and those in nursing homes) rather than comparing them to other recent high school graduates. If what you want is automated driving, the comparison is to something like GM Super Cruise, not a 12-year-old average car (or even a pre-Autopilot Tesla). This biases analysis in favor of FSD when comparing to a baseline driver.
FSD On vs. Off vs. Baseline. If the question is whether FSD is safer on than off, the baseline comparison is not US average 1.5M miles/severe collision, but Tesla’s active safety number of 5M miles. This reduces the FSD safety benefit claim to 8.9/5M = 1.8x, if the analysis were sound (which it is not). The benefit from FSD has to be compared to FSD on vs. off, rather than FSD on vs. a car with no active safety features at all. This biases analysis in favor of FSD when compared to a baseline driver.
Under-Reporting of High-Severity Crashes. Tesla has a systemic under-reporting bias for the most severe crashes, in which the telematics system has been disabled or destroyed due to the crash, preventing automated reporting into their database. I am aware of several fatal crashes in which Tesla says it never received data due to catastrophic vehicle damage. This under-counts the worst crashes, which are the ones that matter the most for safety. This biases data in favor of FSD.
No Fatality Numbers. Fatalities are not broken out as a separate category. Tesla claims that injury data is unreliable because whether someone dies is not recorded in the auto-generated computer reports. However, you can bet that Tesla has a good idea of how many people have died in crashes associated with FSD by, if no other way, counting up the incoming lawsuits. Tesla’s argument amounts to saying an airbag sensor does not know if someone has been injured (which is true as far as it goes), so their hands are tied and they can say nothing about injuries, even though they are required to report injuries and fatalities to NHTSA. They could check their own Tesla insurance data to get a sample of injury and fatality rates. Moreover, IIHS has extensive data that shows Tesla loss rates are nothing special.2 It is difficult to say how much bias this introduces, but with the resources available to Tesla one would think if they could make a credible claim of a lower fatality rate they would spend the resources to do so.
Unreported Crashes. Tesla seems to have had issues with under-counting severe crash reports. NHTSA is investigating Tesla for this,3 and Tesla was caught claiming they did not have a severe crash report that in fact had been sent to them in the Benavides case.4 At least until the NHTSA investigation is completed, any claim by Tesla based on crash rate data must be considered suspect. This potentially biases data in favor of FSD.
Towed But No Airbag Crashes. Tesla excludes crashes in which a vehicle might be towed but no airbag (or other one-shot active safety) device is deployed. However, the human driver baseline includes vehicles towed with no airbag deployments. This is referenced in the Tesla report, but there appears to be no compensation for this effect. The result is likely to be a systematic under-count of Tesla crashes compared to the human driver baseline. This is a potentially significant difference that biases data in favor of FSD.
Definition of Highway Driving. The highway classification criteria is crucial to understanding the results, but unclear. It is common for the most dangerous roads to be state highways and the safest roads to be interstate highways, with something like a factor of 5 or more difference in fatality rates.5 That means that comparing FSD miles on Interstate highways to non-FSD miles on rural, non-limited-access state routes is a grossly unfair comparison.
Tesla states baseline major collisions at 1.5M miles for highway and 505K miles for non-highway, giving a crash rate of 2.97 times higher for non-highway driving. However, NHTSA states that fatality rates are 1.92 times higher for rural driving6 — a factor of 5.7 difference between NHTSA’s data and Tesla’s baseline. This might mean that Tesla is excluding some high-risk rural state routes when they consider highway driving, but it is not possible to know with the presented information, and the potential difference swamps any claimed safety advantage for FSD.Different Drivers, Vehicles, and Locations. Tesla claims 6.4 billion miles on FSD, presumably worldwide. A NHTSA investigation into FSD safety lists approximately 2.9 million US vehicles equipped with the feature.7 That works out to about 2200 miles on FSD per vehicle (less if you consider that some FSD operation takes place outside the US, further noting that this seems to be lifetime miles rather than per year). This likely reflects a combination of partial uptake of the feature to only some cars as well as only some miles driven on FSD even for vehicles with that feature enabled. Owners who pay for FSD might have different driver safety profiles than those who do not. Moreover, FSD owners might well be concentrated in more affluent areas with better roads, and be more mature drivers with better statistical safety outcomes. The with/without FSD comparison needs to account for these issues.
Easy Miles. FSD is more likely to be used on road miles that are FSD-friendly, and in general “easier” miles. For example, there are reports that FSD turns off due to sun glare, which is also difficult for human drivers to handle and clearly presents higher risk than ordinary high-visibility driving situations. Thus, it is likely that FSD is taking credit for safety on “easier” miles than non-FSD driving biasing safety outcomes in favor of FSD.
Five Second Reporting Window. Tesla only considers crashes in which FSD was active within 5 seconds before a crash, even though the NHTSA SGO reporting window is 30 seconds before a crash. This is an intentional choice which removes some crashes from the calculation. It is well-understood that drivers can take much longer than 5 seconds of perception-response time when presented with an unexpected high-risk situation. It is likely that a longer timeframe (such as the 30 seconds required by NHTSA SGO reporting) would reveal more crashes with potential FSD contribution. This biases data in favor of FSD.
Baseline Inconsistency for Non-Highway Data. Tesla argues that pre-2014 vehicle data is a good proxy for the US average fleet since they are about the same age. This seems plausible for their highway data, but is clearly incorrect for non-highway data. Tesla shows a much lower severe crash rate for pre-2014 Teslas than US vehicles for non-highway data (776K vs 505K miles per major collision). It might be that Tesla drivers are superior to the US population, or that they drive in less risky environments, or that Tesla has a systematic under-reporting of crashes for non-highway crashes. This discrepancy undermines Tesla’s baseline crash argument validity.
Better Methodology Available. Tesla admits the comparison to a baseline human driver is difficult. However, they do not acknowledge that there is progress in doing better than this, published in January 2025: https://www.tandfonline.com/doi/full/10.1080/15389588.2024.2435620
No Independent Review Possibility. We cannot cross-check data analysis because Tesla redacts relevant data from public view that they provide to NHTSA via the SGO reporting process. This level of redaction is something no other company does.
FSD Safety Recalls. There is more to safety than net crash rates. For example, even a vehicle with acceptable net crash rates can be subject to NHTSA safety recalls stemming from unreasonable risk to safety due to specific malfunctioning behaviors. FSD has already been subject to such a recall for dangerous driving behavior,8 and is being further investigated.
FSD Practical Misuse. An additional concern is ample anecdotal evidence that FSD is being abused by drunk drivers to drive while intoxicated,9 and has driver monitoring that is permissive enough that it stays engaged even when drivers are inattentive. Blaming drivers will not stop the next crash, but better driver monitoring technology might go a long way to help in this area.
In light of these threats to validity, an apples-to-apples comparison of FSD turned on vs. turned off in comparable conditions could well turn out that FSD is more dangerous. Or not.
The threats to validity are so pervasive and so fundamental that we can conclude nothing useful about the practical safety of FSD from this report.
We do know, however, that the Insurance Institute for Highway Safety (IIHS) has concluded that features like FSD are convenience features, and not safety features.10 There is nothing in Tesla’s FSD safety report to suggest IIHS is wrong.
Other potential comparisons:
SuperCruise claims zero at-fault crashes with 200,000 users on an installed base of 500,000 vehicles, and 700 million hands-free miles.11
Waymo claims an improvement in airbag deployment crashes of 79%, which is a factor of 4.8x compared to a much more nuanced human driver benchmark.12 The difference between Waymo and FSD safety is complicated by the fact that in a Tesla there is a human driver who is supposed to be supervising the automation, and Waymo operates driverless vehicles.
Given those comparisons, Tesla’s claim of 5.9x safer compared to a human driver baseline, even if it were a fully accurate claim, does not seem like it sets them apart from the pack of other companies in the chase to deploy robotaxis. For sure this is not reasonable proof that they are the safest car on the road as one often hears claimed.
See also, some insights from Noah Goodall on a Bluesky thread:
(h/t Matt Wansley)
Perhaps some other, more serious, analysis would show some aspects of safety advantage to FSD. Or not. But Tesla’s latest report, like all their previous reports, lacks rigor needed to be taken as a serious analysis of safety outcomes.
Tesla vehicles have lackluster insurance records for injury and medical claims. See: https://www.iihs.org/research-areas/auto-insurance/insurance-losses-by-make-and-model
One Pennsylvania data set shows non-limited-access state highways having a 6.5 times higher fatality rate per VMT than the PA Turnpike. See: https://www.pa.gov/content/dam/copapwp-pagov/en/penndot/documents/travelinpa/safety/documents/2019_cfb_linked.pdf
The number equipped is not the same as the number with FSD activated. See: https://static.nhtsa.gov/odi/inv/2025/INOA-PE25012-19171.pdf
At-fault crashes are not the same as crashes, but this is not an analysis of SuperCruise crash rates. Source: https://insideevs.com/news/776467/gm-super-cruise-users-zero-crashes/
I’ll let Waymo make their own comparison to Tesla. See: https://waymo.com/safety/impact
Excellent post, Phil. For a similar analysis, see this Bluesky thread by Noah Goodall: https://bsky.app/profile/did:plc:tfondz7k6jhvs6hv7t7dhjti/post/3m5mrttz4hs2b