How Safe Is Tesla Autopilot?

Ask ten people whether Tesla Autopilot is safe and you’ll get ten confident answers — half of them “it’s nine times safer than humans,” half of them “it’s a recall magnet under federal investigation.” The frustrating truth is that both camps are quoting real facts. Tesla publishes crash data that looks spectacular. Federal regulators have recalled Autopilot once and are actively investigating FSD right now. Neither fact cancels the other out.

We’re a pro-Tesla site, and we’ll say plainly: we think Tesla’s driver-assist tech, used correctly, is genuinely useful and probably makes attentive drivers safer on highways. But “used correctly” is doing a lot of work in that sentence, and the marketing names don’t help. So here’s the full picture — Tesla’s numbers, the caveats Tesla doesn’t print in bold, and what the regulators have actually found. We don’t test cars ourselves; everything below is attributed to its source.

What Autopilot and FSD Actually Are (Level 2, Not Self-Driving)

Start with the thing the names obscure: neither Autopilot nor Full Self-Driving makes a Tesla self-driving. Both are SAE Level 2 driver-assistance systems on the Society of Automotive Engineers’ six-level scale (0–5). At Level 2, the car can steer, accelerate, and brake simultaneously, but a human driver must supervise continuously and is legally responsible for the vehicle at every moment. Tesla’s own owner documentation says exactly this — the features “do not make the vehicle autonomous” and require “a fully attentive driver.” Tesla even renamed the consumer product FSD (Supervised) in 2024 to underline the point.

The practical difference between the two tiers:

Basic Autopilot — standard on every Tesla — bundles traffic-aware cruise control (matches the speed of traffic ahead) and Autosteer (lane centering). It’s designed primarily for divided highways.

FSD (Supervised) — a paid upgrade by subscription or one-time purchase — extends assistance to city streets: it attempts traffic lights, stop signs, turns, roundabouts, and lane changes. It is impressive technology, and it is still Level 2. The moment you stop watching the road, you’re outside the system’s design — and outside the data Tesla uses to defend it.

That framing matters for everything that follows. The honest question isn’t “can a Tesla drive itself?” (no), it’s “does a supervised human-plus-computer team crash less than a human alone?” That’s where the data fight lives.

What Tesla’s Safety Numbers Say

Here is the first thing that trips people up: Tesla no longer publishes the Autopilot number that most articles still quote. Through 2025, Tesla ran a quarterly Vehicle Safety Report built around a single Autopilot statistic. In its final published editions it reported roughly one crash per 6.36 million miles with Autopilot engaged, versus about one per 993,000 miles for Teslas driven without it, against an estimated US average near one per 702,000 miles. In 2026 Tesla retired that quarterly Autopilot report — the old address tesla.com/VehicleSafetyReport now redirects to a different, differently-built document.

What Tesla publishes today is the FSD (Supervised) Vehicle Safety Report, and it no longer gives an Autopilot miles-per-crash figure at all. Instead it compares FSD (Supervised) against Tesla’s estimate of the average US vehicle and claims roughly 7× fewer major (airbag-deployment) collisions, 7× fewer minor collisions, and 5× fewer off-highway collisions — summarized as improving US road safety “by over 80%,” across more than 12.9 billion cumulative FSD miles. Those are real, current figures, but read the next section before repeating them: they are not a continuation of the old Autopilot trendline, and they carry the same self-reported caveats.

Taken at face value, that’s roughly a 9x gap between Autopilot-engaged driving and the national average — the figure Tesla fans (and Tesla’s investor decks) repeat. It’s also worth noting that Teslas without Autopilot engaged still beat the national average in Tesla’s data, which says something about the cars’ standard active-safety hardware, the fleet’s newer average age, or its drivers — likely all three. If you’re curious how the underlying cars perform in physical crashes, see our companion piece on whether Teslas are the safest cars on the road and the latest crash-test ratings.

One honest wrinkle from the historical record: even while it was still published, that Autopilot figure was not “improving every quarter.” It drifted slightly down over its final stretch, ending below the report’s earlier highs of over 7 million miles per crash. That history matters now mainly as a caution: single self-reported safety numbers move around, and a headline built on one of them ages badly — which is precisely what happened when Tesla swapped the metric out.

Why Those Numbers Need a Big Asterisk (Highway Skew, Self-Reported)

Here’s the part a credible pro-Tesla site has to say out loud: Tesla’s miles-per-collision comparison is not apples-to-apples, and safety researchers have pointed this out for years. Three problems stack up.

1. It’s a curated comparison, not a controlled study. The classic objection to the old Autopilot number was highway skew — Autopilot ran mostly on highways, which are far safer per mile than city streets, so comparing highway-heavy Autopilot miles against an all-roads national average flattered the system. The new FSD (Supervised) report is a different comparison (FSD is used on city streets too, and Tesla now reports a highway/non-highway split), but the underlying limitation is the same: Tesla measures its own fleet with its own telemetry and compares it to a US-average baseline it assembled from federal datasets that were never designed to line up with Tesla’s crash definitions. Tesla itself concedes the baseline “may skew… higher or lower.”

2. Self-reported, not audited. Both the old and new reports are compiled by Tesla, on Tesla’s own methodology, with no independent verification. The current FSD report counts a “major” collision as an airbag or pyrotechnic-restraint deployment and a “minor” one as a change in velocity of at least 8 km/h (following 49 C.F.R. § 563.5), attributes no fault, builds its “US average” by dividing federal FHWA vehicle-miles-traveled by NHTSA CISS towed-crash counts, and excludes China. Defensible choices — but Tesla’s choices, measured a different, narrower way than the police-reported crashes behind the national baseline. Different definitions, different denominators.

3. Fleet and driver differences. Teslas are newer than the average US vehicle (average US car age is over 12 years) and skew toward demographics with lower baseline crash rates. Some of the gap would exist with no Autopilot at all — Tesla’s own no-Autopilot figure proves it.

What NHTSA Has Found (Investigations & Recalls, Factually)

The National Highway Traffic Safety Administration has been scrutinizing Tesla’s driver-assist systems since 2021. The factual record, without spin:

The December 2023 Autopilot recall. After a multi-year defect investigation that examined hundreds of crashes — including at least 13 fatal ones where NHTSA said foreseeable driver misuse played an apparent role — Tesla recalled about 2.03 million US vehicles, essentially its entire US fleet at the time. NHTSA’s finding: Autosteer’s driver-engagement safeguards were inadequate — too easy to use inattentively, and usable on roads it wasn’t designed for. The remedy was an over-the-air software update with larger alerts, stricter monitoring, and suspension policies for repeat misuse. Worth noting on both sides: it was fixed by software push rather than a service visit, and NHTSA subsequently opened a recall query (2024) into whether the fix went far enough after logging crashes in updated vehicles.

The active FSD investigation. In October 2024, NHTSA opened a defect probe (PE24031) into FSD after crashes in reduced-visibility conditions — sun glare, fog, airborne dust — including one fatal pedestrian crash. On March 18, 2026 the agency upgraded it to Engineering Analysis EA26002, covering an estimated 3,203,754 vehicles (2016–2026 Model S and X, 2017–2026 Model 3, 2020–2026 Model Y, and 2023–2026 Cybertruck equipped with FSD) — the final investigative stage before NHTSA can request a recall. The core question is whether FSD’s “degradation detection” system recognizes when its cameras are impaired and warns the driver in time; the agency is reviewing nine crashes (two injuries, one fatality) plus six more recent potentially related incidents. Separately, NHTSA has examined reports of FSD-equipped Teslas running red lights and reviewed Tesla’s crash-reporting practices.

The IIHS verdict on safeguards. In its first-ever partial-automation safeguard ratings (March 2024), the Insurance Institute for Highway Safety rated both Autopilot and FSD “poor” — though notably, so were 11 of the 14 systems tested industry-wide. IIHS’s broader position is blunt: there is little evidence that any partial automation system makes driving safer by itself; what matters is how well it keeps the driver engaged.

None of this means Autopilot is a death trap — an engineering analysis is a question, not a verdict, and recalls-by-software are now routine across the industry. But anyone telling you the regulators have no concerns isn’t reading the docket.

Driver Monitoring & AEB: The Layers Under the Headline Features

Two less-famous systems carry a lot of the real safety load.

Driver monitoring. Teslas check on you two ways: torque sensors in the steering wheel (the “nag” that demands a wheel wiggle) and, on cars built since mid-2021, a cabin camera that watches for eyes-off-road and phone use. Post-recall software made both stricter — look at your phone with FSD engaged and you’ll be warned within seconds, and repeated strikes suspend the feature for a week. Critics note the camera works poorly in darkness and with sunglasses, and IIHS dinged Tesla’s monitoring as insufficient. The trend, though, is clearly toward tighter enforcement with each software update.

Automatic emergency braking. Every Tesla ships with AEB, forward collision warning, blind-spot collision warning, and lane-departure avoidance as standard — and crucially, they work whether or not Autopilot is engaged. This layer is genuinely uncontroversial: NHTSA considers AEB so valuable it has mandated the technology on all new US vehicles by 2029, and Tesla’s standard active-safety suite is a big reason even the no-Autopilot Tesla crash figure beats the national average. It also helps explain why some insurers price Teslas the way they do — more on that in our Tesla insurance guide.

How to Use Autopilot Safely

Used as designed, Level 2 assistance reduces fatigue and adds a second set of (camera) eyes. Used as a chauffeur, it’s a hazard. The owner’s playbook:

Stay the driver. Hands on the wheel, eyes on the road, every mile. Autopilot is supervising-assistance, not a nap pod — treat the monitoring nags as a feature, not an insult.

Match the road to the tool. Basic Autopilot belongs on divided highways. With FSD on city streets, hover-readiness matters most exactly where the NHTSA probe points: intersections, glare, and weather.

Take over early in bad visibility. Sun glare at dawn and dusk, heavy rain, fog, snow-dusted lane lines, construction zones — the documented weak spots. Don’t wait for the car to ask.

Never defeat the safeguards. Wheel weights and camera covers exist; they’re how Level 2 misuse becomes a fatality statistic, and post-recall software increasingly detects them anyway.

Keep software current and AEB on. Tesla ships safety improvements over the air — the 2023 recall remedy was one. Updating promptly is the cheapest safety upgrade you’ll ever get.

The Numbers Side by Side

Two different Tesla reports, side by side. The first three rows are Tesla’s final published Autopilot figures (before Tesla retired the report) — a metric Tesla has since retired and no longer updates. The last row is what Tesla publishes now. They are not directly comparable, which is exactly the point:

The Bottom Line

Our honest take as Tesla fans: Autopilot used as directed on the highway is very likely a net safety positive — the crash-rate signal in Tesla’s data is too large to be entirely an artifact, AEB and the standard safety suite are excellent, and driver monitoring keeps tightening. The underlying cars are also among the best crash-test performers ever measured, which we cover in our crash-test ratings breakdown.

But the 9x marketing number deserves the asterisk we’ve given it, the names “Autopilot” and “Full Self-Driving” promise more than Level 2 delivers, and a 3.2-million-vehicle federal engineering analysis is open right now. Buy the tech, enjoy the tech — and stay the driver. That’s not a disclaimer; it’s the entire safety case.

Frequently Asked Questions

Is Tesla Autopilot self-driving?

No. Autopilot and FSD (Supervised) are both SAE Level 2 driver-assistance systems. The driver is legally responsible at all times, must watch the road, and must be ready to take over instantly. Tesla’s own documentation says the features do not make the vehicle autonomous.

How safe is Autopilot according to Tesla’s data?

That depends which report — and the one most people quote no longer exists. Tesla’s final Autopilot Vehicle Safety Report claimed roughly one crash per 6.36 million miles with Autopilot engaged, versus ~993,000 miles without it and an estimated ~702,000-mile US average. Tesla retired that quarterly Autopilot report in 2026 and now publishes an FSD (Supervised) report claiming 7×/7×/5× fewer major/minor/off-highway collisions than the US average (“over 80% safer”). All of it is self-reported and shaped by which roads and vehicles are compared, so none of it is apples-to-apples.

Why do experts question Tesla’s safety numbers?

Three reasons: the data is self-reported and not independently audited; Autopilot runs mostly on highways, which have far lower crash rates per mile than the mixed roads in the national average; and Tesla counts roughly airbag-deployment-level events while the national baseline counts police-reported crashes.

Has Tesla Autopilot ever been recalled?

Yes. In December 2023, Tesla recalled about 2.03 million US vehicles — nearly its entire US fleet — after NHTSA found Autopilot’s driver-engagement safeguards inadequate. The remedy was an over-the-air update with stronger warnings and stricter monitoring; NHTSA later opened a query into whether that update was sufficient.

Is NHTSA currently investigating FSD?

Yes. NHTSA opened defect probe PE24031 into FSD in October 2024 over crashes in low-visibility conditions, and on March 18, 2026 upgraded it to Engineering Analysis EA26002, covering an estimated 3.2 million Teslas equipped with FSD — the last step before a recall can be requested. Separate inquiries cover FSD traffic-signal behavior and Tesla’s crash reporting.

What’s the difference between Autopilot and FSD (Supervised)?

Basic Autopilot (standard on every Tesla) bundles traffic-aware cruise control and lane centering, mainly for highways. FSD (Supervised) is a paid upgrade that also attempts city streets — traffic lights, stop signs, turns, lane changes. Both are Level 2 and require a fully attentive driver.

Does every Tesla have automatic emergency braking?

Yes. AEB, forward collision warning, and related active-safety features are standard on every Tesla and work whether or not Autopilot is engaged. NHTSA will require AEB on all new US vehicles by 2029.

How do I use Autopilot safely?

Treat it as a fatigue reducer, not a chauffeur: hands on the wheel, eyes up, use it primarily on divided highways, take over early in glare, rain, fog, or construction, never defeat the driver monitoring, and keep your software and safety settings current.

Related: Are Teslas Among the Safest Cars on the Road? · Tesla Crash-Test Ratings Explained · Tesla Insurance in 2026