HANDSNo data yet · the format is final, the numbers are not collected
This is where we'll tell you how often the car got it wrong.
No rideshare company publishes this voluntarily, and no passenger service publishes it weekly. We will, including the weeks that make us look bad, because a safety claim you can't check isn't a safety claim.
We are publishing the empty report first, on purpose. Every company that eventually discloses something gets to design the disclosure after it knows what the numbers say. We are fixing the format, the definitions and the update schedule now — while we still have no idea whether they will flatter us — so that the first real week can be compared against a promise we made before we had anything to hide.
These four boxes fill in the week we carry our first passenger, and then again every Monday after that, whatever they say.
Trend
One bar per software version.
Interventions per thousand miles, by software version, on the same roads with the same operators. It will be the single most useful number we produce, and it needs months of driving before it means anything at all.
The bars will use a fixed scale across every version. A chart that rescales to fit its tallest bar makes every release look like a breakthrough, which is how most of these charts are drawn and precisely why most of them are worthless.
Caveat we'd rather state now than have pointed out later: our operators will get better over any such period too, and a more experienced operator intervenes less for the same car behaviour. Some of any improvement will be ours, not Tesla's, and we will not be able to cleanly separate the two. We are saying so before we have a trend to be tempted by.
| FSD version | Miles | Takeovers | Per 1,000 mi |
|---|---|---|---|
| first version we run | — | — | — |
A row appears when a version has carried enough miles to be worth reporting, and it never comes back down once it is up.
Causes
Every takeover gets a reason attached
An intervention count with no causes attached is a number designed to be quoted, not understood. These are the categories we will sort every one into, fixed in advance so we cannot invent a flattering bucket later.
| Cause | Count | Share | What it means in practice |
|---|---|---|---|
| Construction & cones | — | — | Lane closures and temporary layouts. The car is cautious to the point of stopping. |
| Operator being early | — | — | Took over before the car had done anything wrong. We will count these deliberately — excluding them would flatter the software with our own nerves. |
| Unprotected left | — | — | Hesitates in a gap a human would take. Rarely dangerous, frequently annoying, occasionally makes someone behind you lean on the horn. |
| Heavy rain | — | — | Cameras struggle in standing water and spray. We expect this to be the one genuine weather limit and we will plan shifts around it. |
| Pedestrian ambiguity | — | — | Someone stepping toward a kerb without crossing. The car stops. That's the correct call, and it still requires a human to resolve the standoff. |
| Emergency vehicle | — | — | Policy, not capability. Operators will take over for every emergency vehicle regardless of what the car is doing. |
| Parking & pickup | — | — | Final approach to a kerb in a tight spot. The last hundred feet get hand-driven anyway. |
A red-light event, a wrong-way event, a lane departure into oncoming traffic, or a failure to yield to an emergency vehicle. Any one of those stops every car until we know why. That rule is written down now, before it costs us anything to say.
"Operator being early" — takeovers where the car was fine and the human wasn't sure. Most services would leave these out. Counting them makes our own numbers look worse and is the main reason to believe the rest.
Geography
The places our cars need a human
Publishing this is arguably against our interest. It will also be the most genuinely useful thing on the site.
Once the cars have driven a route enough times to have a rate rather than an anecdote, the junctions where they most often need a human go here by name — the merge, the unprotected left, the roundabout, the airport kerb. We are not naming any of them in advance. A guess about a junction we have never driven, printed next to a number, is exactly the kind of thing this page exists to be the opposite of.
Each one that does appear will have a written note in the operator handbook describing what the car does there and what to be ready for. Local knowledge is the one thing a national fleet can't replicate, and it gets built out of exactly this table.
Method
How we'll count, and how you could cheat it
Anyone can publish a good number. What makes a number worth reading is knowing exactly how it could have been gamed. This section is written now, in advance, and does not change once data starts arriving.
An intervention is any of these
- The operator turns the wheel, brakes or accelerates hard enough to disengage the system
- The operator disengages deliberately, for any reason at all
- The car hands control back on its own
Here's how a company would fudge this
- Exclude "unnecessary" takeovers. The single easiest way to cut the number. We will count every one.
- Only run FSD on the easy miles. We will publish the FSD-engaged share alongside the rate, precisely so you can see whether we are cherry-picking roads.
- Count per trip, not per mile. Longer trips are mostly highway and highway is easy, so per-trip flatters anyone doing airport runs.
- Rescale the chart each release. Every version becomes a breakthrough. Ours is pinned 0–10.
- Quietly drop a bad week. Which is why the raw file will carry every date in it, and a missing date would be obvious.
One row per intervention: timestamp to the minute, software version, road, speed, weather, cause, and a free-text note from the operator. No rider names, no pickup or drop-off addresses, no operator names — operators will be a stable ID so you can still see per-operator patterns without identifying anyone.
- File
- first file published in our opening week
- Rows
- — none collected
- Updates
- Mondays, 09:00 ET, once we are running
- Licence
- CC BY 4.0 — take it, publish it, argue with it
- History
- Every previous week stays up, unedited
There is no file to download yet, and we would rather show you an empty box than a sample one. A realistic-looking example is how a placeholder ends up quoted as a result.
Want to be told when the first one goes up, or think the method above is wrong? data@handsfreefsd.app. Corrections get published with your name on them, which is the only way this is worth anything.
They will not be a safety rating. An intervention is a human deciding to take over, and most will not be moments where anything bad was about to happen. There is no honest way to convert this into "X times safer than a human driver", and any company that does that conversion for you is doing marketing.
They will also come from a small fleet in one small area. That is a tiny sample by the standards of this field — enough to see our own trend, nowhere near enough to conclude anything about Full Self-Driving in general.
And we will not be the only disclosure that exists, which we would rather say than have pointed out. California requires every company testing autonomous vehicles there to file an annual public report listing each disengagement. That is a real dataset and in some respects a stricter one. What is missing from it — and what this page is meant to become — is a commercial passenger service publishing weekly, unprompted, in a state that asks for nothing.