> Tesla's Q4 2023 stats should speak for themselves:

When stats speak for themselves, they’re quite likely misleading.

> > In the 4th quarter, we recorded one crash for every 5.39 million miles driven in which drivers were using Autopilot technology. For drivers who were not using Autopilot technology, we recorded one crash for every 1.00 million miles driven. By comparison, the most recent data available from NHTSA and FHWA (from 2022) shows that in the United States there was an automobile crash approximately every 670,000 miles.

Let’s start with the obvious: what about crashes that occurred within, say, 10 seconds of autopilot turning off?

And the slightly more subtle: the distributions of roads and conditions driven on autopilot is surely very different than the distribution driven manually. One could try to control for this by identifying segment of road driven on autopilot in, say 10 minute intervals. Then normalize the analysis in the autopilot and manual groups to weight each (road segment, time segment) equally in each groups.

As a concrete example of the latter issue, consider I-80 through Donner Pass. In good conditions, it’s a very nice freeway. In snow, it’s not so nice. I would expect the general accident rate to by far more than 5.39x higher in snow. But, in snow, autopilot will get rather less use. (And I bet autopilot is very likely to disengage, by itself or by driver override, if the car loses traction.)

And this is a misinformed comment from someone who blindly parrots Tesla's numbers even though we know, for a fact, that Tesla doesn't follow industry reporting practices by redacting, obscuring, hiding information.

Their numbers don't even normalize or otherwise account for different roads, road conditions, environmental conditions, cars involved and their safety systems, or drivers behind the wheel.

If they were interested in making good faith comparisons, the "US average" would read "US average of comparable vehicles" and "Tesla vehicles not using autopilot" would read "Tesla vehicles not using autopilot under equivalent weather & road conditions".

You're essentially telling us that drivers driving Teslas with active autopilot (i.e. limited to great weather conditions on high quality roads) have fewer accidents than those without active autopilot (e.g. driving in poor weather conditions, or on bad roads). That's not much of an insight.

Tesla continues to misleadingly use this comparison, and somehow people continue to fall for it.

Autopilot will refuse to engage in a wide variety of situations. It will only engage in the easiest, safest scenarios. You cannot compare those miles against all miles driven by human drivers, which include narrow winding mountain passes, two inches of snowy slush, etc.

Even the 1M non-autopilot comparison; you're comparing against a car population that includes 20 year old beaters with far fewer safety features.

> Autopilot will refuse to engage in a wide variety of situations. It will only engage in the easiest, safest scenarios.

Do you have any proof of this? I can enable it on a dirt road if there are visible enough edges of the road.

But yes, obviously, they're being a bit misleading by aggregating all miles together. In aggregate, it is overall safer on the highway, and I don't think you can argue in good faith that the safety report is wrong about that.

Will you accept Tesla's word for it?

https://www.tesla.com/ownersmanual/model3/en_us/GUID-101D1BF...

> In addition, these features may not work as intended when: The road has sharp curves or significant changes in elevation.; Road signs and signals are unclear, ambiguous, or poorly maintained.; Visibility is poor (due to heavy rain, snow, hail, etc. or poorly lit roadways at night); You are driving in a tunnel or next to a highway divider that interferes with the view of the camera(s); Bright light (such as from oncoming headlights or direct sunlight) interferes with the view of the camera(s).

https://www.tesla.com/ownersmanual/model3/en_us/GUID-E5FF5E8...

> Full Self-Driving (Beta) and its associated functions may not operate as intended and there are numerous situations in which driver intervention may be needed. Examples include (but are not limited to): Interactions with pedestrians, bicyclists, and other road users.; Unprotected turns with high-speed cross traffic.; Multi-lane turns.; Simultaneous lane changes.; Narrow roads with oncoming cars or double-parked vehicles.; Rare objects such as trailers, ramps, cargo, open doors, etc. protruding from vehicles.; Merges onto high-traffic, high-speed roads.; Debris in the road.; Construction zones.; High curvature roads, particularly at fast driving speeds.

This is basically my point at the bottom of my original post. It isn't supposed to be used in those scenarios. That doesn't change the reality that my car's Autopilot can and does work on dirt roads, when the sun is directly in the middle of its FOV, on the curvy suburban roads of Georgia, in heavy rain, etc.

Sure, but that's not the argument being made here. The data that Tesla uses is obviously massaged to make their product have the illusion of superior safety. To what extent is unclear, but it is clear that this data has an agenda. It doesn't compare like with like and can't be piecemeal interpreted to make an unbiased observation. To bring it up at all is just marketing.

Normalizing the data to make a direct comparison is a hard problem. Do you know of any research done to try to do that?

It's a lot easier to tease out the impact of individual safety features.

For example, there's quite a bit of evidence that lane departure warnings/assistance helps: https://www.consumerreports.org/car-safety/lane-departure-wa...

Less indication backup cameras do anything: https://www.latimes.com/business/la-fi-hy-back-up-cameras-20...

Tesla's non-autopilot numbers likely reflect the fact that they're exclusively newer vehicles with safety features that aren't yet widely standard. You'd probably see similarly nice numbers for other $50k+ luxury vehicles with similar bells-and-whistles.

> Tesla's non-autopilot numbers likely reflect the fact that they're exclusively newer vehicles with safety features that aren't yet widely standard. You'd probably see similarly nice numbers for other $50k+ luxury vehicles with similar bells-and-whistles.

This is probably not possible since those automakers don't do constant data collection in the same way Tesla does. Tesla's telemetry logs actual data for the distance their cars are driving and whether AP is enabled. All we have for other brands is NHTSA's approximations.

Crashes generally have multiple factors. For the moment, let's set aside the multiple driver errors (not just drunkenness) and focus on issues Tesla has control of which contributed to the death:

* FSD was wildly insufficient for this road and required constant driver input

* Despite being insufficient for the road, FSD could be enabled on the road

* Despite needing constant correction, the FSD stayed enabled

* FSD can be enabled at speeds and conditions where the driver would need to correct with very little notice

All of these are completely self-inflicted issues caused by Tesla management arrogantly insisting on having the FSD attempt much more than it is capable of. Compare this to Ford and GM where they carefully geofence their systems to highways where they are absolutely certain their driving aids will work.

There's a causation fallacy at work here. To infer that autopilot is safer according to these stats you would have to control for far too may confounding factors. For example, the majority of crashes occur in rural and urban settings. It is likely that autopilot is in use more often in situations in which crashes are less likely (such as simple, long distance highway driving).

It's almost like the term Full Self Driving is a complete lie.

Drivers of late model, $50-100k vehicles slightly safer than fleet averages. More misleading news after this word from our sponsors...

Elon can also speak for himself: https://www.youtube.com/watch?v=5cFTlx68itk&t=93s

Stop falling for Tesla's cherrypicked statistics.