The long road to autonomous vehicle safety

[mostlylong]

The technology journey for autonomous vehicles (AVs) to achieve human-level or better safety is a long, bumpy and winding road.

Waymo (the only US company with the 10s of millions of autonomous miles required to statistically validate safety) has been working on AVs since 2009 and Tesla since about 2016.

People who drive Tesla vehicles with its FSD (Full Self Driving) driver assist software (which requires full human driver supervision at all times) regularly say things like:

“It works 99% of the time, therefore FSD is really close to true autonomy” (meaning human or better safety).

This reasoning is unlikely to be true because 99% safe is far away from human-level safety.

Let’s say human-level safety is “six 9s” (a realistic assumption), meaning:

• 99.9999% safe
• 1 incident in 1,000,000 miles
• 999,999 safe miles per 1 million miles

Suppose FSD safety is four 9s:

• 99.99% safe
• 1 incident in 10,000 miles
• 9,999 safe miles per 10,000 miles

What is the safety gap?

• FSD can do 9,999 safe miles before an incident.
• Its safety target is 999,999 safe miles before an incident.
• The safety gap is 990,000 miles = 999,999 - 9,999

So FSD has solved 9,999 of the 999,999 safe miles target, or

FSD is 0.0099 or 1% of the way to human safety.

(If FSD is 99% safe - as many people like to say - then FSD is 0.01% of the way to human safety.)

@mostlylong

The march of the 9s is the progression in the number of 9s in a safety rate.

[captainccs]

@mostlylong

Let’s say human-level safety is “six 9s”

Let’s say you use a different hat to pull numbers from, then how many nines do we get?

If EVs are recording what they see on the road, how is it different for different car brands?

The Captain

[intercst]

I’m not worried about the “9’s”, but these Waymo flash mobs in suburban neighborhoods have to stop.

Empty Waymos invade Atlanta neighborhood, circle cul-de-sac for hours with no passengers

intercst

[mostlylong]

@captainccs

@mostlylong

say human-level safety is “six 9s”

Let’s say you use a different hat to pull numbers from, then how many nines do we get?

Not following, exactly.

Six 9s is a fair ballpark, broad average safety level for human drivers.

Do you have different numbers to use?

Tesla said this on their Q2 earnings call, just passed:

It’s really just about going through what we call the march of nines of reliability, where you need, how many nines of reliability do you need to scale? Ideally, you want 99.999999% reliable. It’s just the march of nines of reliability is, I think, the only thing really constraining our growth in Robotaxi.

Tesla says eight 9s - we could use that number from the CEO’s hat and then they are even farther away from their safety target than my example above.

Tesla also straight out says their need to improve safety is constraining their growth.

@intercst

not worried about the “9’s”

If I am sharing the road with robotaxis as a passenger, fellow driver, or pedestrian, then I would certainly pay attention to the 9s.

Similarly, I would have the 9s in mind if I am expecting Tesla to generate any revenue anytime soon from true autonomy.

[albaby1]

@mostlylong

Not following, exactly.

Six 9s is a fair ballpark, broad average safety level for human drivers.

Or to put it another way, the typical human driver in the U.S. gets into an accident (major or minor) every ________ miles. There are lots of different estimates of that number. Tesla says 700K, I’ve seen figures closer to 500K, and as low as 100K. Either way, as an order of magnitude it’s going to be between 10^5 and 10^6 miles. Or roughly six 9’s, the way you’ve defined it (percentage chance per miles).

Put still another way an AI driver should be able to go at least 100,000 miles between requiring a human intervention to avoid an accident. For a typical American driver, that’s between six and seven years of driving.

[intercst]

@mostlylong

Similarly, I would have the 9s in mind if I am expecting Tesla to generate any revenue anytime soon from true autonomy.

Human capacity for driving skills continues to deteriorate. It’s just a question of when they cross the current state of FSD performance.

As I was backing out of my garage last night, a DoorDash driver stopped at the curb. My Tesla started chiming alarms and slammed on the brakes. I felt contact. Upon exiting the vehicle I noticed a black streak on the rear bumper of my white Tesla which I feared was paint from the dark-colored Door Dash vehicle. Upon inspection, I found it was a thin streak of road film, and neither vehicle had any paint damage. It was that close.

intercst

[albaby1]

@intercst

Human capacity for driving skills continues to deteriorate.

Probably not.

The below chart takes the number of police-reported crashes each year and the total number of VMT (vehicle miles travelled) each year, and calculates the number of miles between such crashes. You can see that over time such crashes have become less frequent, not more (moving towards the top of the graph means more miles between crashes, and thus them becoming less common).

These are only police-reported crashes, so we don’t know what the rate of very very minor collisions might be (which is why there are different figures than 560K miles per crash) - but directionally, there’s no evidence that people are getting worse at driving.

[mostlylong]

@albaby1

Or to put it another way, the typical human driver in the U.S. gets into an accident (major or minor) every ________ miles. There are lots of different estimates of that number.

The accident rate depends on the driving scenario: interstate, city, specific routes, time of day, driver demographics, vehicle type and age, etc.

When Waymo tries to benchmark the safety of their AV, they attempt to statistically control for these variables in their specific markets (Phoenix, Bay Area, Austin, etc).

Their recent study:

@mostlylong

One statistic: Waymo says their injury rate is 82% lower than a human driving under comparable scenarios.

The data are based on 220 million autonomous miles through March 2026. Their December report was based on 170 million miles, so looks like they are accruing 50 million miles in 3 months, or 200 million miles per year.

waymo safety report

[intercst]

@albaby1

Probably not.

The below chart takes the number of police-reported crashes each year and the total number of VMT (vehicle miles travelled) each year, and calculates the number of miles between such crashes.

If humans were maintaining their collective driving skills, I’d expect to see a slow decline in the number of crashes as a larger percentage of the active fleet has modern safety features like automatic braking, forward collision avoidance, blind-spot monitoring & rear cross-traffic alert, and lane keeping assist. The fact that were not seeing a decline, tells me that drivers are getting worse and the safety technology is merely maintaining the baseline crash rate.

intercst

[captainccs]

@mostlylong

Not following, exactly.

Six 9s is a fair ballpark, broad average safety level for human drivers.

Do you have different numbers to use?

My question, Where did you get the “Six 9s is a fair ballpark” opinion? I asked Google AI for help using the Tesla quote as a prompt:

The phrase “march of nines” refers to the exponential effort required to increase a system’s reliability percentage by adding sequential “9s” to the figure (e.g., moving from 99.9% to 99.99% and beyond). [1, 2]

What the “Nines” Mean

Reliability percentages translate to allowable failure or downtime rates over a given timeframe: [1]

• Three 9s (99.9%): ~8.7 hours of downtime/failure per year.
• Five 9s (99.999%): ~5 minutes of downtime/failure per year.
• Eight 9s (99.999999%): Practically zero tolerance for error, equating to fractions of a millisecond of acceptable failure time across massive operational scales. [1, 2]

Why It Matters for Robotaxis

As discussed during events like the Tesla Q2 2026 Earnings Call, scaling autonomous fleets requires pushing reliability to extreme levels. [1]

• Exponential Difficulty: Each added nine roughly doubles or exponentially increases the engineering, redundancy, and testing cost.
• Safety Constraint: Human drivers operate at roughly a “six 9s” ballpark safety equivalent, but commercializing driverless robotaxis safely across diverse real-world edge cases demands far higher certainty to prevent catastrophic failures.

The long road to autonomous vehicle safety

o o o o o o o o o o o o o o o o o o o o

Notice that the “Six 9s is a fair ballpark” is a quote by you, not some independent party.

Where did you get it?

The Captain

[albaby1]

@intercst

If humans were maintaining their collective driving skills, I’d expect to see a slow decline in the number of crashes

Sure - that’s absolutely a fair point. But isn’t that what we see? Over the long-term, that graph goes up and to the right - decline in the number of crashes over time. There was a break in that when they fiddled with the methodology, and the pandemic slashed the number of miles driven which also disrupted the trend in the graph. But after each of those things, the general “up and to the right” trend continued. Generally speaking, we have seen a long-term system decline in the number of crashes.

[albaby1]

@captainccs

Notice that the “Six 9s is a fair ballpark” is a quote by you, not some independent party.

Where did you get it?

I thought that was pretty clear.

Humans get into an accident about once every 700,000 miles. That’s per Tesla. There’s a lot of different figures that vary depending on what accidents you count, etc. But we can kind of bracket it. A human driver in the United States will typically go more than 100,000 miles, but less than 1,000,000 miles, between collisions. So a rough and ready estimate of how safe an AI driver is to be safer than a human is that it has to go 1,000,000 miles between crashes.

One million miles between crashes is a safety rate of 99.9999% per mile, or six nines.

If your AI driver only gets to five nines (measured on miles) there’s a good chance it’s not safer than the average human driver in the United States. If it gets to six nines, it probably is.

[mostlylong]

@captainccs

the “Six 9s is a fair ballpark” is a quote by you, not some independent party.

Where did you get it?

Al provided a very nice synopsis:

@albaby1

A human driver in the United States will typically go more than 100,000 miles, but less than 1,000,000 miles, between collisions. So a rough and ready estimate of how safe an AI driver is to be safer than a human is that it has to go 1,000,000 miles between crashes.

I’ll add a little more detail.

First I’ll note that the Tesla CEO’s comment from the Q2 2026 earnings call

1. mentioned a safety target of eight 9s, so even safer than six 9s
2. stated that reaching sufficient safety is the only thing keeping their robotaxi fleet from growing

It’s really just about going through what we call the march of nines of reliability, where you need, how many nines of reliability do you need to scale? Ideally, you want 99.999999% reliable. It’s just the march of nines of reliability is, I think, the only thing really constraining our growth in Robotaxi.

As far as sources for accident rates in the US,

First, the Waymo Safety Report provided upthread has a ton of relevant info and references with many details, including many technical research studies:

@mostlylong

per year.
waymo safety report

Second, Al upthread showed a chart of miles per accident:

@albaby1

below chart takes the number of police-reported crashes each year and the total number of VMT (vehicle miles travelled) each year, and calculates the number of miles between such crashes

Third, NHTSA compiles this kind of data:
NHTSA Motor Vehicle Traffic Crashes

For example, look at Tables 1 (VMT) and 2 (Total Crashes) for year 2024.

1. Total Crashes = 6,180,241
2. Vehicle Miles Traveled (VMT, millions) = 3,294,031
3. Crashes per Million Miles = 6,180,241/3,294,03 1= 1.88 = 1.88 crashes per 10^6 miles

The six 9s comes from the exponent 6 for 1 million in 10^6.

Of course, the 1.88 rate above is an overall average over a very wide variety of driving scenarios.

As already mentioned upthread:

@mostlylong

The accident rate depends on the driving scenario: interstate, city, specific routes, time of day, driver demographics, vehicle type and age, etc.

When Waymo tries to benchmark the safety of their AV, they attempt to statistically control for these variables in their specific markets (Phoenix, Bay Area, Austin, etc).

Everything stated by me and Al above about the 9s of vehicle safety is 100% consistent with what the Tesla CEO stated on the last earnings call.