How it works, and how well

1. The data

Every vehicle complaint and recall NHTSA publishes: 1,609,665 complaints received from Jan 1995 to Oct 2026, and 26,445 recall campaigns. A complaint is counted once even when it names several components. Names, phone numbers, addresses and similar details are masked in the complaint text before anything else reads it.

2. Grouping complaints into problems

NHTSA files complaints under broad components such as "engine" or "electrical system", which mix dozens of unrelated failures. BrakeNews groups complaints by what the owner describes instead. Each narrative is turned into a vector of numbers that captures its meaning (sentence-transformers/all-MiniLM-L6-v2), and similar narratives are clustered together.

That gives 128 groups, holding 87% of the 1,605,624 narratives; the rest fit no group well. An AI (claude-haiku-5-5) named each group from samples, set aside the ones that are not about a defect (for example, complaints that recall parts are unavailable) and merged duplicates, leaving 93 problems that cover 78% of narratives.

The groups are approximate. Similar-sounding complaints can land together: complaints about the glare from other drivers' headlights share a group with complaints about dim headlights. The checks in step 6 exist for this reason.

Recalls are matched to problems the same way. A recall's defect summary shortlists the problems it resembles, and the AI picks the ones it actually covers. It checked 21,649 recalls; 62% cover at least one problem. A recall covers a vehicle's problem when it names that make, model and model year. Matching misses recalls filed under another spelling of the model (such as "Wrangler 4xe" for "Wrangler").

3. Flagging a vehicle

Two tests run every month for every vehicle and problem.

Compared with other vehicles

Over the last 24 months: what share of this vehicle's complaints are about the problem, and what share of all other vehicles' complaints are? The ratio of the two is the proportional reporting ratio (PRR), the method drug-safety agencies use on adverse-event reports. Comparing shares rather than counts cancels what inflates counts without a defect: a popular model, an older vehicle, a year with more complaints overall. A vehicle is flagged when the problem has at least 3 complaints, the ratio is at least 2, and a chi-square test says the difference is unlikely to be chance.

Compared with its own past

A surge test asks whether the vehicle's complaints in the last three months have shifted toward the problem beyond what its own history predicts (a binomial test, with the false discovery rate controlled across all vehicles each month).

4. Track record

The tests were run month by month over the whole history, using only the complaints received up to each month, and compared with the recalls that came later.

Recalled problems flagged first

73%

Of 2,986 vehicle problems later recalled (with 3 or more complaints before the recall)

How early

18 months

Median time from the first flag to the recall

Flags followed by a recall

3.2%

Within 24 months, of 38,512 flags old enough to judge

A flag made a recall about 5.1 times as likely as for problems with as many complaints that never stood out; a surge about 8.3 times. Most flags are still never followed by a recall. Some are not defects, and many real defects are fixed without a recall. A flag is a reason to look closer, not a prediction.

5. The recall estimate

A model (gradient-boosted trees, LightGBM) estimates for each vehicle and problem the chance that a recall covering it follows within 24 months. It sees only what was known at the time: complaint counts and trends, both tests above, severity (crashes, fires, injuries), the vehicle's age, and context such as how often this problem has been recalled on other vehicles, complaints about the problem in the last 24 months, the vehicle's age, signals for the same problem on other model years of the model.

It was trained on data from Jan 1995 to Oct 2013, tuned on data from Jan 2016 to Oct 2018, and tested on data from Jan 2021 to Oct 2024, with gaps between the periods so no recall in one period leaks into another. On the test period, at the same number of alerts as the PRR flag (55,777):

Ranking quality on the test period: average precision 0.041 (the PRR ratio alone: 0.009), ROC AUC 0.798. The estimates run slightly low: they average 0.7% against 1.0% observed.

6. AI checks and investigation reports

Because the problem groups are approximate, an AI (claude-haiku-5-5) reads the recent complaints behind each of the 4,254 highest-ranked alerts and decides, one complaint at a time, whether it is about that problem on that vehicle. 2,478 alerts were confirmed, 103 were mixed and 1,671 were not confirmed. Vehicle pages flag a problem only when its alert was confirmed and no recall covers it yet.

For the top confirmed alerts, an AI investigator (claude-opus-5-5) reads the complaints, compares other model years and vehicles, looks up recalls, and writes a report that cites the NHTSA complaint numbers behind each finding. A second AI (claude-opus-5-5) fact-checks every finding against the full text of the complaints it cites, and the investigator revises if needed. Only reports the reviewer approves are published. No person reviews them.

7. Limits

• Complaints are written by owners and are not verified by NHTSA or by BrakeNews.
• A recall is only one outcome of a real defect, and many recalls follow a manufacturer's own findings rather than complaints. "Followed by a recall" undercounts real problems.
• Problem groups, recall matches and checks are made by AI and can be wrong.
• Vehicles with fewer than 25 complaints have no page; their numbers are too small to compare.
• Nothing here is safety advice. Check your vehicle's open recalls with your VIN at nhtsa.gov/recalls.

Source code: github.com/Hariswar/BrakeNews.