The Highway Bias: Why Tesla's Safety Claims Need Context

Every quarter, Tesla publishes a Vehicle Safety Report that makes a bold claim: driving with Autopilot engaged is dramatically safer than driving without it. In Q3 2025, Tesla reported one crash for every 6.36 million miles driven with Autopilot, compared to one crash every 702,000 miles without Autopilot. The implied message is clear — Autopilot makes you 9 times safer.

But there's a fundamental problem with this comparison that Tesla never acknowledges, and that most media coverage uncritically repeats: Autopilot is overwhelmingly used on highways and freeways, while the baseline "all driving" figure includes every road type, weather condition, and time of day.

The Road Type Problem

According to NHTSA's Fatality Analysis Reporting System (FARS) and FHWA data, fatality rates vary enormously by road type. Interstate highways have a fatality rate of 0.64 per 100 million vehicle miles traveled (VMT). Urban streets are roughly double at 1.30 per 100M VMT. Rural roads — narrow, unlit, with oncoming traffic — are the most dangerous at 2.42 per 100M VMT.

The national average of 1.26 per 100M VMT is a blend of all these environments. When Tesla compares Autopilot's crash rate against this blended number, they're comparing a highway-optimized system against a benchmark that includes the most dangerous roads in the country — roads where Autopilot is rarely or never used.

What the Numbers Actually Show

If we take Tesla's reported Autopilot crash rate and compare it to the highway-only baseline instead of the all-driving baseline, the safety factor drops from 9x to approximately 3-4x. That's still better than human highway driving — and that matters. But it's a far cry from the headline-grabbing "9 times safer" that Tesla promotes and media repeats.

The shrinkage becomes even more dramatic when you consider additional confounders. Autopilot is typically used in good weather (Tesla recommends disengaging in rain, snow, and fog). It's used during daytime hours more than night. It's used in free-flowing traffic on well-maintained roads. Each of these factors independently correlates with lower crash rates.

The Trend Is Real — But So Is the Spin

Looking at Tesla's quarterly reports over time, the trend is genuinely positive. The miles-per-crash figure for Autopilot has improved from 4.31 million in Q3 2023 to 6.36 million in Q3 2025 — a 48% improvement. The system is getting safer. That's undeniable and worth acknowledging.

But the improvement trend exists within a framework of misleading comparison. Getting better at highway driving and then comparing yourself to all-driving averages is like a swimmer improving their pool time and then claiming they're faster than runners, cyclists, and swimmers combined.

Why It Matters

This isn't just an academic quibble. Tesla's safety claims influence consumer behavior, regulatory decisions, and public policy. When a driver hears "9 times safer," they may over-trust Autopilot in situations where it performs less well — urban streets, construction zones, adverse weather. The inflated safety factor creates a false sense of security that can itself become dangerous.

Regulators evaluating whether to expand Autopilot's operating domain rely on these safety claims. If the true highway-adjusted improvement is 3-4x rather than 9x, that changes the risk calculus for allowing Autopilot on more road types.

What Would Honest Reporting Look Like?

A truly transparent safety report would compare Autopilot's crash rate to the crash rate on the same road types, at the same times of day, in the same weather conditions where Autopilot actually operates. Tesla has this data — every mile driven with Autopilot is logged with GPS coordinates, road type, weather, and time. They could publish an apples-to-apples comparison. They choose not to.

Until they do, the "9x safer" claim should come with a giant asterisk. Autopilot may well be safer than human driving in its operating domain. But by how much? That's the question Tesla doesn't want you to ask.