On May 1, 2025, AZfamily’s investigative reporter Derek Staal did a piece about Waymo; it of course piqued my interest… What happens when a Waymo is at fault in a crash? , here’s an excerpt

Federal crash data from the National Highway Traffic Safety Administration (NHTSA) shows Waymo vehicles were involved in at least 202 crashes in Arizona between 2021 and 2024, ranging from minor scrapes and door dings to 31 crashes that resulted in injuries. But the records are heavily redacted and do not indicate who was at fault.

Arizona’s Family Investigates spent six months collecting and reviewing police reports on Waymo crashes in Phoenix, Scottsdale, Chandler, Mesa and Tempe. Of the 71 police-documented crashes involving Waymos, officers determined the self-driving car was not at fault 87% of the time. In 13% of cases, however, police blamed the Waymo vehicle.

Derek was kind enough to explain via email more details about his methodology: he make public records request to each city where Waymo operates (so: Phoenix, Tempe, Mesa, Chandler, Scottsdale) which  resulted in lists of specific incidents, which he then used to obtain Arizona Crash Reports thru normal channels

(Derek) submitted public records requests to every city in Waymo’s service area, asking for the list of all collisions involving Waymos in a given timeframe. That list has crash report numbers, which I then used to pull the underlying crash reports. Every city except Phoenix uses BuyCrash.com.

Although Derek declined to share his data, an Arizona Crash Report lists which unit in a crash the officer believes is at fault (strictly speaking “most at fault”); by marking that as unit #1, and all other units involved be they other vehicles, pedestrians or bicyclists are assigned the next ordinal number.

This is quite a paperwork undertaking/hassle, and costs several dollars apiece for each report; and it only gets more ugly as more cities get added.

Crash Database

Readers of azbikelaw will remember that ADOT maintains a database of all Crash Reports (please don’t call them accident reports). The way it works is each police jurisdiction is required make a submission of all traffic reports they investigate (so long as they meet a certain threshold, like property damage > $2,000, OR any injury). Nowadays, the vast majority (or all?) are submitted directly electronically from the jurisdiction, then ADOT’s Traffic Records Division reduces this to a pure dataset; the set excludes an PII (personal identifying information: e.g. names, phone numbers, etc), and also unfortunately excludes the narrative. Incidents not involving suspected criminal issues normally show up within a matter of a couple of weeks.

As mentioned in the news piece, autonomous vehicle companies have stringent reporting requirements, and NHTSA makes this information publicly available; but for reasons that I don’t understand they redact so much info as to be useless in terms of finding the crash report or data; for example they redact the date (listing only the month and year), and the location.

For the record, here is a copy of the NHTSA data, retrieved on May 5, 2025; just Waymo, and ordered by state, the first 305 entries are waymo, Arizona incidents. At this time, here is the page on NHTSA: Standing General Order on Crash Reporting For incidents involving ADS and Level 2 ADAS

Waymo’s own data

Happily, we don’t have to struggle with the NHTSA data because Waymo has taken it upon themselves to publish unredacted data, specifically, it contains a full date, time, and location of the incident.

It is linked from their page Safety/Impact, , and the download we’re looking for they call CSV2, Crashes with SGO identifier and group membership. (and here is a copy of that file retrieved May 6,2025) Which is said to be current as of December 2024 (most likely thru end of year, as there is an incident 12/29/2024).

Here is that file on a google sheet, listing just Arizona, and with information added in the blue columns that correlate a waymo incident to ADOT’s IncidentID number…

Waymo Crashes with SGO ID and Group Membership 202009-202412-2022benchmark AZ ONLY

The sheet is ordered with “police reported” incidents at the top: 61 rows, and the other 127 rows being not police reported. (which sums to 188, a bit fewer than the 202 that was mentioned in AZFamily story; not sure why or if there is a discrepency)

By correlating the date, time, and given location; I searched ADOT records “manually” to look at similar cross streets (the database lists cross streets, not a street address as given by waymo data) and close time. Here’s the problem: of the 61, only 28 are found in ADOT’s records. I have no idea why.

Of the 127 waymo incident that were supposedly not police reported, I found two that obviously were, because they do appear in ADOT records. (I didn’t manually look thru all 127, I got bored; ones that I searched and could not find have a ‘0’ for ADOT ID column.

At Fault?

All the details are in the google doc sheet mentioned above; it includes a worksheet with the Unit details for each incident:

• • As mentioned above, AzFAmily found Waymo at fault 18% of the 71 incidents, so 13 of the 71.
  • Crash Database has Waymo at fault, coincidentally, also 18% (5 of 28 incidents)
  • Waymo’s narrative clearly reduces that number to 11% (3 of 28 incidents)

In Waymo’s narrative, in two of the incidents where police faulted waymo, their narrative describes a driver that abruptly/unsafely changed lanes immediately ahead, resulting in a rear-end collision. I tend to believe Waymo; narrative because I know they have an infinite amount of video from all angles. My surmise is this video was not available to the investigating officer at the time the Arizona Crash Report was completed; they took a statement from the other involved driver (“no, officer– that waymo just slammed into me”), and that was that; these reports never get corrected or updated.

Some observations on the data (see the Unit Details): inconsistent UnitType, probably should all be DRIVERLESS, however a bunch of them say DRIVER. BodyStyle and Color are inconsistent, all (except for the oldest one) are Jaguar I-Pace’s and all are white. Arizona’s database does not list a model, but rather a BodyStyle and were reported somewhat randomly as, Passenger 4Dsw Station Wagon 4 Dr, Passenger Sw Station Wagon, Passenger Ht Hardtop. Four of them had a color of not_reported.

Looking at the Person data (another tab in the google sheet), there were inconsistencies e.g. in driver’s age (some zero, some 255, some not reported, some 122, 123) (the 12x is typically used for hit-run drivers), and gender (a mixture of U and null). In any event consistency would be welcome.

Why is it that many incidents (Arizona Crash Reports) cannot be found in state crash data?

I can only find 28, whereas AZFamily found 71 Crash Reports. Perhaps a few can be explained by a time-frame overlap. I was looking at data current as of March 2025 (some reports lag, but typically not by much).
Perhaps a few are there but in some way mis-reported (e.g. as to location, time, date?)

Perhaps it’s because the cities are not transmitting all the Arizona Crash Reports… but why? Perhaps in some instances the police took a report but later decided it doesn’t meet the $ PDO (Property damage Only) threshold; that wouldn’t explain the four incidents in Waymo’s data that say injury reported, but cannot be found in state crash records.

Is fault getting better? In other words is Waymo “learning”?

?

Who, in Arizona is watching / evaluating the safety of the AV programs?

Going all the way back to 2015, Arizona state gov’t has been notably friendly to autonomous vehicles from a legislative/administrative standpoint. There’s a webpage on ADOT’s site that sounds good,  Arizona Self-Driving Vehicle Oversight Committee, but appears to be an orphan, last updated in 2016(!)

.

Some Queries (please ignore)

Note that both a parked car and an autonomously driven vehicle will show up as a DRIVERLESS UnitType (which is kindof annoying, couldn't a parked car be something else like e.g. PARKED?). Same with PersonType is even more confusing... stemming from the fact that a truely driverless car will always have a person (person #1) and as I mentioned above sometimes it's DRIVERLESS and sometimes DRIVER, and the demographics (age, sex) are all over the place.

SELECT * FROM unit 
WHERE ( (MakeDesc LIKE "Jagu%") OR (MakeDesc LIKE "Chry%" AND IncidentID = 3880902) )
AND IncidentID IN (3880902,4074944,4087789,4116330,4117194,4117796,4125808,4133057,4142822,4155292,4160650,4165652,4177377,4183257,4184701,4186570,4187775,4195951,4196428,4202916,4203764,4207262,4222717,4229295,4234686,4238994,4243980,4250814)
ORDER BY IncidentID;

SELECT IncidentID,UnitNumber,PersonID,ePersonType,Age,Sex, eViolation1 FROM person 
WHERE PersonNumber=1
AND IncidentID IN (3880902,4074944,4087789,4116330,4117194,4117796,4125808,4133057,4142822,4155292,4160650,4165652,4177377,4183257,4184701,4186570,4187775,4195951,4196428,4202916,4203764,4207262,4222717,4229295,4234686,4238994,4243980,4250814)
ORDER BY IncidentID;

.