Tesla Published Its First European FSD Safety Report. The Netherlands Data Is Compelling — and Full of Holes.

Zero collisions in 16.6 million kilometers.

That is the highway safety number from Tesla’s first jurisdiction-specific European FSD safety report, published June 9, covering Full Self-Driving Supervised operations on Dutch roads from April 10 to June 5, 2026. The dataset spans 23.6 million total kilometers driven under FSD, making it the largest regional safety disclosure Tesla has produced outside the United States. By every metric Tesla chose to report, the system outperformed human drivers: 3.5 times fewer collisions overall, 14.9 times fewer automatic emergency braking events, 8.8 times less harsh acceleration, 7.3 times less harsh braking, and 8 times fewer hard swerves.

Impressive numbers. But the context around them is where this story fractures, because the headline comparison mixes measurements that do not belong in the same sentence, and the sample has problems that no amount of favorable ratios can fix.

What the Data Actually Shows

Tesla broke its collision data into highway and non-highway segments, and the split matters more than the headline figure.

Road Type	FSD km	FSD Collisions	Manual km	Manual Collisions	FSD Safety Ratio
Highway	16.6M	0	158.7M	33	∞ (zero crashes)
Non-highway	7.0M	3	152.9M	109	1.6×
Total	23.6M	3	311.6M	142	3.5×

The 3.5× headline is heavily weighted by the highway result. Seventy percent of FSD kilometers were driven on highways, where the system recorded zero collisions against a manual baseline of 33 in nearly ten times more distance. But highway driving is the easiest thing an ADAS system does — it is what Autopilot was built for a decade ago, before FSD existed. The hard test is city streets. On non-highway roads, where cyclists weave between trams and roundabouts force continuous judgment calls, the safety advantage shrinks to 1.6 times.

That 1.6× figure is still positive, but it is built on exactly three collisions. A sample of three events over two months, against a manual comparison pool that is 22 times larger in distance, does not produce statistical reliability, and if one of those collisions had gone differently — higher speed, worse angle, a pedestrian instead of a car — the ratio would move substantially in either direction.

The Denominator Problem Nobody Ran

Here is the comparison that no other coverage of this report has made, and it illustrates why the 3.5× number is less meaningful than it appears.

Tesla reports all collisions — from fender-benders in parking lots to serious impacts. The Netherlands national road safety statistics, published by CBS (Statistics Netherlands) and the road safety research institute SWOV, report fatalities. In 2024, the Netherlands recorded 675 traffic deaths across roughly 140 billion vehicle-kilometers, yielding approximately 4.8 deaths per billion vehicle-km. In 2024, 220 of those deaths were passenger car occupants.

Tesla’s non-highway collision rate works out to 3 collisions in 7 million km, or 428 collisions per billion kilometers. Manual non-highway: 109 in 152.9 million km, or 713 per billion. Now line that up against the Netherlands national fatality rate for passenger cars, which runs roughly 1.6 deaths per billion vehicle-km based on CBS data showing 220 car occupant deaths across approximately 140 billion vehicle-kilometers in 2024.

The problem is immediate. You cannot compare 428 collisions per billion km to 1.6 fatalities per billion km — they measure fundamentally different things. Tesla counts every collision regardless of severity: the parking-lot scrape, the low-speed rear-end, the mirror clip. National statistics count only deaths, and the ratio between total collisions and fatalities in the Netherlands runs several hundred to one, which means any analysis placing these numbers side by side without acknowledging the denominator mismatch is misleading, whether the person doing it knows better or not.

This does not make Tesla’s data useless — far from it. The within-dataset comparison, FSD collisions versus manual collisions in the same fleet over the same roads in the same time period, is methodologically sound as a ratio. The 3.5× figure compares like with like. The problem starts when anyone uses Tesla’s self-reported collision count to make claims about how FSD compares to the national road safety baseline, because the national baseline does not count what Tesla is counting.

The Bigger Story: Europe Requires This Data, America Does Not

The RDW type approval that enabled this data release came after 18 months of testing: 1.6 million kilometers on EU roads, 13,000 customer ride-alongs, 4,500 closed-track tests, and verification against over 400 compliance requirements. The European version of FSD runs software v14.2.2.5, which Tesla says “differs substantially” from the US version, including stricter driver attention monitoring and a requirement that every software update be pre-approved by the RDW.

The approval was classified under UN Regulation No. 171, which governs Driver Control Assistance Systems (DCAS) — Level 2 autonomy, meaning the driver retains full legal responsibility. As a condition of that classification, Tesla must report safety data to the RDW annually.

That annual reporting requirement is the genuinely novel element of this story — not the collision numbers, not the behavioral metrics, but the regulatory structure that forces data into public view. In the United States, NHTSA relies on Standing General Orders for crash reporting — a voluntary framework that requires manufacturers to report crashes involving automated driving systems, but imposes no pre-market approval, no annual data submission, and no software update review before deployment. Tesla has driven billions of FSD-miles on American roads without any regulatory requirement to produce the kind of jurisdiction-specific safety report it just published for a country smaller than New Jersey.

The regulatory asymmetry is stark. Europe now has a framework that forces data into public view. The Netherlands, Lithuania, Estonia, and Denmark have each approved FSD through mutual recognition of the RDW type approval. But regulators elsewhere are not on board. Reuters reporting on leaked regulatory emails shows that officials in Sweden, Finland, Denmark, and Norway have raised concerns about FSD’s tendency to exceed posted speed limits, its safety on icy roads, and Tesla’s lobbying tactics of encouraging owners to pressure regulators. An EU-wide committee vote that was on the June 30 agenda has been removed; Nordic objections have pushed it to autumn 2026 or Q1 2027 at the earliest.

The Strongest Case for the Data

The counterargument deserves its full weight, because it is strong enough to reframe the skepticism.

Self-reported data is the only kind of data that exists at this scale, and nobody else has it. No government has the infrastructure to independently measure collision rates across an active commercial fleet in real time, and complaining about self-reporting without proposing an alternative is criticism that leads nowhere. The RDW tested for 18 months before approval and will audit Tesla’s reporting annually — that is not “trust me,” it is “trust me, with a regulatory backstop and an expiration date.” If Tesla’s data turns out to be fabricated or systematically biased, the RDW audit will catch it, and the type approval can be revoked.

The behavioral metrics may actually matter more than the collision counts. Fourteen times fewer AEB events means the system is not reaching emergency-braking situations in the first place — it is avoiding the precursors to accidents, not just the accidents themselves, and these are leading indicators that aggregate thousands of micro-events rather than a handful of collisions and are consequently much harder to game through selection bias or definitional tricks.

And the sample bias cuts both ways — yes, early adopters are more attentive, but they also stress-test the system deliberately, driving FSD through Amsterdam’s canal streets and rush-hour roundabouts because they want to see it break. The Netherlands has excellent road infrastructure, but it also has the densest cycling traffic in Europe, where electric fatbikes, cargo bikes, and e-scooters create a moving obstacle course that ADAS systems historically cannot parse. If FSD handles Amsterdam for two months without incident, that counts for something — even if it does not count for everything.

Limitations

Everything in this analysis comes from Tesla. The RDW has not independently verified collision counts or behavioral metrics, and no severity breakdown exists for the three FSD non-highway collisions or the 109 manual collisions — we do not know whether any of FSD’s three crashes were fender-benders or serious impacts requiring medical attention, which makes the safety ratio impossible to interpret at the level that actually matters to a crash victim.

Disengagement data is missing entirely. Tesla does not disclose whether the system handed control back to the driver in the seconds before any of the 142 manual collisions, which would shift those events from the “manual” column to one that is harder to classify. The manual driving comparison pool of 311.6 million km includes all Tesla drivers in the Netherlands, not a cohort matched for age, driving experience, or route selection, and the 70/30 highway-to-non-highway split in FSD kilometers may not match the split in manual driving. NHTSA’s active investigation into FSD red-light-running crashes in the United States involves different software versions and is not directly comparable.

The Bottom Line

Tesla’s Netherlands data is the first time any company has published jurisdiction-specific, collision-level safety data for an ADAS system operating at scale on European roads. That alone makes it significant, and the behavioral metrics favor FSD even more convincingly than the collision counts do.

If you are a European regulator evaluating FSD for your country, the actionable takeaway is not the 3.5× headline. It is the reporting framework. The RDW built a system that produces public data as a condition of approval. Demand the same. When Tesla applies to operate in your jurisdiction, require annual safety reports with severity breakdowns, disengagement data, and a matched comparison cohort — the Netherlands proved the data can be extracted, so make the data better.

If you are a Tesla owner in Europe, the data says FSD Supervised is performing well on Dutch roads, but it does not say FSD is 3.5 times safer than you are. It says FSD collides less often than the average Tesla driver operating manually, in a two-month window, in one country, on a fleet that skews toward highways. That is encouraging, not a reason to stop paying attention behind the wheel, which remains a legal requirement under the DCAS classification that makes this system street-legal in the first place.

If you are following NHTSA from across the Atlantic, the question is no longer whether FSD data exists — it does. The question is why the agency that oversees the country where FSD has driven the most miles still has no equivalent requirement to produce it.