FSDClarity

How Safe Is Tesla FSD? What the Data Actually Shows

8 min readBy the FSDClarity team

On this page

There is no trustworthy public number for how safe FSD is. Tesla publishes figures suggesting it goes several times further between crashes than an average US driver. Those figures are real, but they are not comparable to the national statistics they're placed next to — and NHTSA currently has an open investigation sitting at the final tier before a recall decision.

This page lays out what Tesla publishes, why it doesn't answer the question, what regulators are actually looking at, and what to do with the car in the meantime.

What Tesla publishes

Tesla's safety data, released in November 2025, compares miles between collisions for cars driving on FSD against national averages:

MeasureOn FSDAll US drivers
Miles between major collisions~2.9 million~505,000
Miles between minor collisions~986,000~178,000

Two definitions matter. Tesla counts a collision as major when an airbag or pyrotechnic restraint deploys. And it attributes a crash to FSD if FSD was active within five seconds before the impact.

Taken at face value, that's roughly six times further between serious crashes. The problem is everything the comparison leaves out.

Why those numbers aren't a fair comparison

The fleet is not comparable. FSD runs on recent Teslas. The national average covers every vehicle on American roads, including cars twenty years old with no automatic emergency braking, no stability control and far weaker crash structures. Safety researcher Phil Koopman has made this point directly: comparing new cars against the whole fleet inflates the gap before FSD does anything at all.

The roads are not comparable. Driver-assistance miles concentrate on highways and well-mapped routes, which have substantially lower crash rates per mile than urban streets. The national figure includes all of it.

A human is in the loop. This is the one that gets overlooked. Every FSD mile in that statistic had an attentive driver supervising and taking over when needed. The number therefore measures FSD plus a human, not FSD. It tells you very little about how the software would do alone — which is exactly the question people are asking.

Fatalities aren't broken out, and Tesla has said its injury data isn't reliable enough to publish, so the most serious outcomes are the least visible.

None of it is auditable. Tesla doesn't publish the underlying mileage in a form anyone outside the company can check, so the per-mile rates can't be independently verified. Electrek has reported, citing a Reuters investigation, that some of Tesla's own AI trainers consider the safety report misleading.

To be fair to Tesla: publishing anything is better than the silence from most of the industry, and the figures are probably directionally meaningful for highway driving. They just can't carry the weight of "FSD is six times safer."

What NHTSA is investigating

This is the more concrete information, because it describes specific failure modes rather than aggregate claims.

InvestigationScopeStatus
EA26002 — reduced visibility~3.2 million vehiclesEngineering analysis since 18 March 2026
PE25012 — traffic violations58 reported incidentsPreliminary evaluation
Cybercab deployment audit (Austin)Commercial robotaxi serviceSpecial Order issued

The reduced-visibility probe is the serious one. NHTSA upgraded it from a preliminary evaluation to an engineering analysis in March 2026 — the last investigative stage before the agency either demands a recall or closes the matter. The pattern it describes is FSD losing track of, or never detecting, a vehicle ahead in conditions like sun glare, fog, rain or dust. Nine crashes were documented, one of which killed a pedestrian.

The agency has also indicated it suspects crashes are under-reported, because of limitations in how Tesla records and labels the data.

The second probe concerns FSD-engaged cars running red lights or crossing into oncoming lanes — 58 incidents, still at the earlier evaluation stage.

As of the most recent filings I could verify, no recall has been issued from any of these. An engineering analysis can end in a software recall, a hardware remedy, or nothing.

The crash counts, and why they're hard to read

Tesla reported 207 driver-assistance crashes in May 2026, its highest monthly total on record. That sounds alarming and it mostly isn't informative: raw counts rise as the fleet and its mileage grow, so the meaningful figure is crashes per mile — which, as above, nobody outside Tesla can calculate. The crash narratives Tesla files with NHTSA are also heavily redacted, so the circumstances of almost all of them are not public.

In October 2026 Tesla confirmed that Autopilot or FSD was engaged when a car crossed a median and killed a man. Individual cases don't establish a rate, but they do show the failure modes are not hypothetical.

What the data can and can't tell you

It can't tell you whether FSD is safer or less safe than you, per mile. Anyone claiming either with confidence is going beyond the evidence.

It can tell you where the system is known to struggle, and that's genuinely useful. The regulatory record points at reduced visibility, failing to register a lead vehicle, and intersection behavior including red lights. Those are the conditions to be most alert in — see common FSD problems for how they show up day to day.

It also tells you the system is SAE Level 2. Legally and practically, you are the driver. FSD (Supervised) is named that way for a reason, and unsupervised operation remains limited to Tesla's own robotaxi fleet.

Which hardware do you have?

Check if your Tesla has HW3, HW4, or AI5 and see which FSD features are supported.

Check Hardware →

How to supervise it properly

Based on the documented failure modes rather than general advice:

  1. Take over in poor visibility. Heavy rain, fog, dust, snow and low sun are the exact conditions in the engineering analysis. This is the single highest-value habit.
  2. Be ready at intersections. Cover the brake approaching lights and unprotected left turns, given the red-light and opposing-lane incidents under investigation.
  3. Watch for the lead vehicle. The documented pattern is FSD not registering a car ahead. If traffic ahead is slowing and the car isn't, intervene immediately rather than waiting to see.
  4. Hands on the wheel, eyes on the road. Not because of the nag, but because the stated safety figures only hold with a supervising driver.
  5. Don't use the screen while it's driving. It's both a strike risk and the most common reason people miss a developing situation.
  6. Take over in school zones and car parks, where conditional limits and pedestrians defeat a vision-only reading of the scene.
  7. Report what goes wrong. Say "Bug report" within seconds of an event, and answer the intervention prompt when it asks why you took over.

Does this affect insurance?

Being the legally responsible driver means a crash on FSD is your claim. Some insurers do price FSD miles favourably on newer hardware — see the FSD insurance discount — but no insurer treats FSD as removing your liability.

Frequently asked questions

Is Tesla FSD safer than a human driver?

Unknown from public data. Tesla's figures suggest roughly six times further between major crashes, but they compare new Teslas on mostly highway miles against every vehicle on US roads, and every FSD mile had a human supervising. The honest answer is that nobody outside Tesla can calculate a comparable rate.

How many FSD crashes have there been?

Tesla reported 207 driver-assistance crashes to NHTSA in May 2026, the highest single month on record. Counts rise with fleet size, and most crash narratives are redacted, so the number is hard to interpret.

Is FSD under federal investigation?

Yes. NHTSA has an engineering analysis into FSD in reduced visibility covering about 3.2 million vehicles, a preliminary evaluation into traffic-violation incidents, and an audit of the Austin robotaxi deployment. No recall has resulted so far.

What is an engineering analysis?

NHTSA's second and final investigation stage. It follows a preliminary evaluation and ends with either a recall demand or closure. Reaching it signals the agency found enough in the first stage to continue.

When is FSD most likely to fail?

The regulatory record points to reduced visibility — sun glare, fog, rain, dust — failing to detect a vehicle ahead, and intersection behavior including red lights and crossing into opposing lanes.

Am I liable for a crash on FSD?

Yes. FSD (Supervised) is SAE Level 2, which means you are the driver and remain responsible, whatever the car was doing.

The bottom line

Treat both the marketing and the outrage with suspicion. Tesla's numbers are better than nothing but can't support the claim made of them; the monthly crash counts look worse than they are; and the genuinely useful public information is the list of conditions where FSD has been documented to fail.

Supervise it hardest in poor visibility, at intersections, and when traffic ahead is slowing. That's not generic caution — it's where the evidence actually points.

Related guides