Sampling: how cases get into the dataset
Where reports come from, which are kept, how they become cases, and what that means for what the data can support.
In brief. This is not a complete list or a random sample of incidents. It is every case found by a fixed set of news searches, in news published since January 1, 2026, that has been coded and has passed the screening rules below. 347 candidate cases are still waiting to be coded, so the published list is a small part of what has been found. It can show what kinds of incidents are reported and how they are handled. It cannot show how often they happen.
Why there are few cases, and why some are old
Collection began on November 22, 2026. Each daily run looks back 7 days. On the first day the Google News searches were also repeated one week at a time back to January 1, 2026, so the dataset draws on news published since then. Nothing earlier has been searched.
The weeks before November 15, 2026 were searched after the fact, on Google News only, and return whatever it still lists for those weeks. Expect them to be thinner than the period collected live.
Finding a report is only the first step. A case appears on the site after it has been coded and screened, and 14 of the 28 cases coded so far were set aside (see the numbers). Until the waiting cases are coded, the published list will stay much shorter than the list of what has been found.
Cases are dated by when the incident happened, not when it was reported, so an old incident appears whenever it is in current news. That happens in two ways:
- New coverage of an old event. A verdict, an appeal, a settlement or a regulator's report brings an earlier crash back into the news. A crash from 2019 can enter the dataset because of a court filing this month.
- Older articles returned by a search. One of the two news searches has no date limit and sometimes lists articles that are months or years old. They are kept if they pass the rules. Of the 629 reports collected so far, 1 was published before January 2026.
So a case dated 2023 does not mean coverage of 2023 is complete. It means one incident from 2023 was in the news we searched.
The numbers so far
| Reports kept | 629, published between May 18, 2021 and November 22, 2026 |
|---|---|
| Candidate cases | 375, after grouping reports about the same event |
| Published | 14 |
| Screened out | 14, each listed with its reason in screened_out.csv |
| Waiting to be coded | 347. These are not shown on the site until they have been coded and screened. |
These figures are recalculated every time the site is rebuilt.
Reports kept, by month of publication
| May 2021 | 1 |
|---|---|
| January 2026 | 21 |
| February 2026 | 39 |
| March 2026 | 47 |
| April 2026 | 35 |
| May 2026 | 33 |
| June 2026 | 68 |
| July 2026 | 52 |
| August 2026 | 38 |
| September 2026 | 90 |
| October 2026 | 114 |
| November 2026 | 91 |
Where reports come from
Once a day a program runs the searches below on Google News (United States, English edition) and Bing News. The Google searches are limited to the last 7 days. Bing offers no date limit, so it returns whatever it currently lists. No other source is searched: not police or court records, regulators' databases, social media or reader submissions.
Google News searches
Tesla (Autopilot OR "Full Self-Driving" OR FSD OR "self-driving") (crash OR accident OR collision)Tesla (robotaxi OR Cybercab OR driverless) (crash OR accident OR collision OR incident)Tesla (Autopilot OR FSD OR "self-driving") (police OR arrested OR charged OR asleep OR DUI)Tesla (Autopilot OR "Full Self-Driving") (lawsuit OR verdict OR trial OR investigation)Tesla ("Smart Summon" OR Summon OR Autopark) (crash OR hit OR collision)
Bing News searches
Tesla Autopilot crashTesla Full Self-Driving crashTesla robotaxi crash
Makers covered: Tesla. Each search returns only the results the service chooses to list, in an order it does not explain, and the same search can return different results on different days.
Which reports are kept
A report is kept when its headline (or, for Bing, the short description supplied with it) names an automated-driving feature and its headline points to something that happened. Headlines with a finance or product term are dropped first. The word lists are:
- Feature terms: autopilot, full self-driving, fsd, self-driving, self driving, robotaxi, cybercab, driverless, autonomous, summon, autopark, autosteer, driver-assist, driver assistance, no driver, no one behind the wheel, nobody behind the wheel, empty driver.
- Event terms: crash, collision, collide, accident, wreck, hit, struck, strike, slam, rear-end, kill, fatal, death, dead, die, injure, injury, hurt, asleep, sleeping, dui, dwi, owi, drunk, intoxicated, arrest, charged, cited, ticket, pulled over, police, lawsuit, sue, sued, verdict, jury, trial, settle, settlement, wrong way, red light, near miss, swerve, railroad, train, school bus, pedestrian, cyclist, motorcyclist, fire, investigation, investigate, probe, ntsb.
- Dropped if the headline has: stock, shares, earnings, price target, market crash, release notes, rolls out, accessories.
The rule is applied to headlines, not article text. An article whose headline says only “Tesla crash” is missed even if the article discusses Autopilot.
How reports become cases
The unit is the incident, not the article. Reports are joined into one case by rule, and the rules are cautious, so they are more likely to leave one event as two cases than to join two events:
- The same link, or the same publisher and headline, is counted once.
- Two reports are compared only if they were published within 45 days of each other and do not name different U.S. states.
- They are treated as the same event when their headlines are nearly identical (90% alike), or at least 75% alike and share two distinctive words, or share three distinctive words with one headline largely contained in the other. Similar wording alone is not enough: “Tesla in self-driving mode involved in crash” fits many events.
- Reports within 5 days of each other are also treated as the same event when their headlines and summary lines share at least 4 specific words (places, road numbers, charges).
- A report joins a case when it matches any report already in that case. Dates are compared report to report, so a case with months of coverage does not pull in unrelated later news.
- At coding time the model is shown similar earlier cases and may say a new one is the same incident, in which case the reports are joined.
Which cases are published
A case is published only when the coding step finds all of the following in the reports. Otherwise it is set aside and listed with the reason.
- One identifiable incident. Statistics, policy stories, explainers, product news and opinion pieces are set aside.
- Not a round-up. An article covering several separate incidents is set aside rather than split into cases, because its details usually cannot be tied to one event.
- An automated-driving feature is claimed, questioned or confirmed to be involved. A crash in which no report raises the feature is set aside.
- A covered maker. Incidents involving another maker's vehicle are set aside for now.
The coding step reads up to 5 reports per case: those with the most text, always including the newest. Where a publisher blocks automated reading, only the headline and summary are available, and the variable text_basis records which was used. A published case is never withdrawn automatically; it is recoded when new reports are attached to it.
Known gaps and biases
- News decides what is here. Deaths, arrests, lawsuits and anything on video are far more likely to be reported than minor events, so serious and unusual incidents are over-represented.
- English, and mostly the United States. The searches use English terms and a U.S. news edition.
- Search services are a filter we do not control. Their ranking, limits and removals are not documented, so the collection cannot be exactly repeated.
- Time coverage is uneven. Fullest for news published from November 15, 2026 onward, which was collected live from both search services. News from January 1, 2026 to then was searched afterwards on Google News only and is thinner. Anything earlier is accidental.
- Headline wording matters. Incidents reported without a feature term in the headline are missed; incidents where the feature is the story are favoured.
- Round-ups are lost. An incident reported only inside a round-up does not become a case.
- Duplicates and wrong joins are possible. Two cases may be one event, and rarely one case may mix two. The joining rules were tightened on November 22, 2026 after a test showed that the earlier rules let a long-running case absorb unrelated reports; nothing from that test was published.
- Coding has changed hands. Each record names its coder in the
codercolumn; the first cases were drafted by a different model from the one the daily job is set to use. None has been reviewed by a person yet. - No denominator. Nothing here measures how much these features are used, so counts cannot be turned into rates or risks.
Check it or change it
Everything on this page is read from the settings and code that run the collection, which are public:
- config.yaml: the searches, word lists and thresholds.
- collect.py and cluster.py: how reports are fetched, kept and grouped.
- coder.py: the screening and quotation checks.
- Change history: every daily refresh, with its counts.
Earlier periods are searched with python -m pipeline.run --backfill 2026-01-01, which repeats the Google searches one week at a time back to the date given. Results for past weeks are whatever Google News still lists, so a search of the past will be thinner than live collection.
See also About the data for how cases are coded and the codebook. Suggestions for other sources or search terms are welcome at consortium@crimrxiv.com.