The short answer
Key facts
- 2,053 closed car and light-truck defect investigations in NHTSA's file; 672 (32.7%) ended in a recall, 1,381 (67.3%) did not.
- The rate has risen every decade: 15.9% (1980s), 28.8% (1990s), 39.1% (2000s), 48.7% (2010s), 52.7% (2020s so far). The all-history figure understates what an investigation means today.
- Median time open: 150 days. Preliminary evaluations 132 days; escalated engineering analyses 371 days.
- Escalation raises the odds, but less than the raw numbers suggest: 35.2% for probes NHTSA never escalated, 54.2% for ones it did - a ratio of 1.54, not the 2.25 the headline rates imply.
- A close with no recall is often a handover, not an all-clear: 575 preliminary evaluations closed without a recall and were followed by an engineering analysis on the same component; 285 of those produced a recall.
- No vehicle history report shows investigation status, ours included. Investigations attach to a model-year range, not a VIN.
Does an open NHTSA investigation mean a recall is coming?
It is close to a coin flip on modern data, and that is a more useful answer than either extreme. Of the 2,053 closed car and light-truck defect investigations in NHTSA's investigations file, 672 produced a recall and 1,381 did not. The all-history rate is 32.7%.
But the all-history rate is the wrong number to quote at a buyer in 2026, because it is dominated by the 1,165 probes opened in the 1980s and 1990s, when an investigation very rarely ended in a recall. The trend by decade is monotonic and steep:
| Decade opened | Closed probes | Ended in a recall | Rate |
|---|---|---|---|
| 1980s | 353 | 56 | 15.9% |
| 1990s | 812 | 234 | 28.8% |
| 2000s | 558 | 218 | 39.1% |
| 2010s | 236 | 115 | 48.7% |
| 2020s (to date) | 93 | 49 | 52.7% |
The 2010s row is the one to trust. Its probes are essentially all closed, so it is not distorted by investigations still in progress. The 2020s row looks slightly higher but should be read as a ceiling rather than a rate: 35 car and light-truck probes opened in 2024-2026 are still open, and among probes opened since 2010 the ones that close without a recall take a median 373 days against 224 for the ones that produce a recall. The slow closures that are still pending therefore skew toward no-recall outcomes, which biases every recent figure upward.
What is the difference between a preliminary evaluation and an engineering analysis?
They are two rungs of one ladder, and which rung a probe is on tells you something real - just less than it appears to. NHTSA opens a preliminary evaluation (PE) to take a first look. If the evidence warrants, it closes that and opens an engineering analysis (EA), which is the deeper stage and the one that most often precedes a recall.
The raw rates look dramatic: 23.6% of closed preliminary evaluations carry a recall campaign, against 53.2% of engineering analyses. That comparison is mostly an accounting artefact, and it is the check that most changed this article. When a preliminary evaluation is escalated, any recall that eventually follows is filed against the engineering analysis, not against the PE that started it. Credit each preliminary evaluation with its successor's recall and the PE rate rises from 23.6% to 43.7%.
What survives is a smaller, genuine effect:
| Probe type | N | Recall somewhere in the chain | Rate |
|---|---|---|---|
| Preliminary evaluation, never escalated | 787 | 277 | 35.2% |
| Preliminary evaluation, escalated to an engineering analysis | 633 | 343 | 54.2% |
So escalation multiplies the odds by about 1.54, not the 2.25 the naive figures suggest. On probes opened since 2010 the ratio narrows further, to 1.19 (54.4% against 64.7%). If you see that NHTSA has upgraded a probe on a car you are considering, that is meaningful - but it is one input, not a verdict.
If NHTSA closed an investigation without a recall, is the car in the clear?
Often not, and the file shows exactly why. "Closed, no recall" reads like an all-clear and frequently is not one: it is the paperwork of a handover to a bigger investigation.
Of the car and light-truck preliminary evaluations that closed without producing a recall of their own, 575 were followed, within a window of 30 days before to 180 days after the close, by an engineering analysis from the same Part 573 filer on the same component system. 285 of those escalated probes did produce a recall.
Honda's momentary-steering-effort case is the pattern in three rows of the frozen CSV: PE23005 opened 2023-03-17 and closed 2023-11-30 with no recall; EA23003 opened 2023-11-29 and ran 426 days; recall 24V744 came out of it. Anyone who read the first closure as an exoneration got the wrong answer fourteen months early.
The matching rule here is ours, not NHTSA's - the file does not carry an explicit "upgraded to" pointer - so the number moves with the window you choose: 39.4% of the no-recall closures match at 30 days, 42.7% at 90, 53.0% at the 180-day window used above, and 59.3% at a year. The successor_ea column in the dataset records which probe was matched to which, so the rule can be checked or rejected.
How long does a defect investigation take?
A median of 150 days, but the tail is long enough to outlast most people's car search. Preliminary evaluations close in a median 132 days; engineering analyses take 371. One in ten closed probes stayed open longer than 547 days, and 173 of them (8.4%) ran past 600.
The extremes are worth seeing, because they are the reason "NHTSA is looking into it" cannot be treated as a short wait:
| Investigation | Filer | Subject | Days open | Outcome |
|---|---|---|---|---|
| PE16014 | Chrysler (FCA US) | Vehicle rollaway | 2,596 | closed, no recall |
| EA14002 | Kia | Panoramic sunroof breakage | 2,440 | closed, no recall |
| PE18007 | Ford | Loss of motive power | 2,356 | closed, no recall |
| PE24024 | Chrysler (FCA US) | Underhood fires | 748 | recall 26V363000 |
| EA25003 | Honda | Rear-view camera failure | 584 | closed, no recall |
What do this week's two closures actually show?
One of each outcome, closed six days apart, and together they are the whole mechanism.
The one that ended in a recall. PE24024 opened on 2024-09-06 into underhood fires and closed on 2026-09-24 after 748 days, producing recall 26V363000: 1,076,999 model-year 2021-2025 Jeep Wrangler, Wrangler 4xe and Gladiator vehicles. NHTSA records it under a hydraulic power-steering pump component label with a vehicle fire as the consequence, and its remedy text tells owners to park outside and away from structures until the repair is done. The manufacturer, component label and unit count all match NHTSA's recalls API exactly.
The one that did not. EA25003, into rear-view camera failure, closed on 2026-09-18 after 584 days with no recall - but it was the third NHTSA action on the same defect. PE24004 opened 2024-02-13, closed 2024-08-05, and produced recall 24V321000 (187,290 model-year 2020-2024 Honda Ridgeline vehicles). A recall query followed. Then the escalated engineering analysis closed with nothing further. One defect, three consecutive federal actions, two years and seven months, and one recall sitting in the middle of the chain rather than at the end of it.
That is the practical lesson for a used-car buyer: the status of a single investigation on the day you look is a snapshot of a process, not a verdict on the car.
Which manufacturers' investigations most often end in a recall?
We ran this ranking, and our own check says not to lean on it. The raw per-company table spans 57.1% (Volvo Cars) down to 25.8% (General Motors). Two things make it untrustworthy as a league table:
- Volume. The Spearman correlation between a filer's number of closed probes and its recall rate is -0.568. Filers with the most investigations have the lowest rates.
- Era, which is the actual driver. Ford and General Motors hold 362 and 323 pre-2000 probes respectively, from the decades when recalls rarely followed. Restrict to probes opened since 2000 and the order changes materially: Ford rises to 40.2% and sits above Toyota at 35.0%, while Honda goes from 12.5% pre-2000 to 56.9% after it.
| Company (consolidated filer) | Closed probes | Ended in a recall | All-history rate | Rate, opened 2000+ |
|---|---|---|---|---|
| Hyundai | 55 | 29 | 52.7% | 68.8% |
| Honda | 77 | 34 | 44.2% | 56.9% |
| Nissan | 77 | 34 | 44.2% | 50.0% |
| BMW | 46 | 22 | 47.8% | 50.0% |
| Stellantis (Chrysler/FCA) | 339 | 121 | 35.7% | 45.2% |
| Ford | 532 | 147 | 27.6% | 40.2% |
| General Motors | 504 | 130 | 25.8% | 36.7% |
| Toyota | 82 | 28 | 34.1% | 35.0% |
| Volkswagen Group | 81 | 29 | 35.8% | 34.8% |
Read that as descriptive, not as a measure of how readily a company recalls. One further caution that applies to any per-company cut of this file: a single investigation can list more than one Part 573 filer, so the per-company counts sum above the 2,053-probe universe, and NHTSA's filer field spells one company several ways over time (Mazda, Jaguar Land Rover and Mercedes-Benz each appear under multiple entity strings and were consolidated by hand here).
What a VIN check can and cannot tell you here
This is a story about the one thing a vehicle history report genuinely cannot see, so it is worth being exact.
What a Zilocar VIN check can show: that a recall exists on the vehicle, and its count; accident and damage records including airbag deployment; odometer readings and rollback indications; theft records; junk and salvage auction records; ownership history; sales-listing history; specifications; NHTSA and IIHS ratings; and a valuation.
What it cannot show, and what you should go elsewhere for:
| Question | In a VIN report? | Where it actually lives |
|---|---|---|
| Is there an open recall on this vehicle? | Recall presence, yes | NHTSA VIN lookup for per-VIN status |
| Was the recall repair actually done? | No | NHTSA's VIN lookup or a franchised dealer |
| Is this model under NHTSA investigation? | No | NHTSA's investigations pages |
| Will an investigation become a recall? | No | Nobody knows; this page gives you the base rate |
| Is the title legally branded salvage or rebuilt? | No - we show the auction record | NMVTIS and the state DMV |
An investigation does not attach to a VIN. It attaches to a manufacturer, a component and a model-year range, which is why no report keyed on a 17-character VIN can return one. If a car you are considering sits inside a model-year range NHTSA is investigating, the honest reading is the base rate on this page plus the knowledge that you may be waiting a year or more to find out.
Method
Source file. NHTSA's Office of Defects Investigation investigations flat file, FLAT_INV.txt, inside FLAT_INV.zip. Downloaded 2026-10-09 at 07:15 UTC; the server reported Last-Modified: Thu, 08 Oct 2026 09:46:33 GMT. The file carries 154,429 rows resolving to 5,349 distinct NHTSA action numbers; its newest investigation open date is 2026-09-21 and its newest close date is 2026-09-24.
Universe and filter. The file stores one row per action-number-and-model-year combination, while the open date, close date and CAMPNO are action-level fields repeated on every row, so the file is collapsed to one record per NHTSA action number first. We keep action numbers prefixed PE (preliminary evaluation) and EA (engineering analysis): these are the two rungs of the defect-investigation ladder and they ask the same question. Excluded by design: DP (defect petitions, which are consumer-initiated and end in a recall only 1.6% of the time), RQ (recall queries, which by definition start from a recall and so would bias the rate upward at 65.0%), and the audit, equipment and other query prefixes.
What counts as ending in a recall. The CAMPNO field, non-empty. This is NHTSA's own link from an investigation to the Part 573 recall campaign it produced, and it is the field that makes this analysis possible; no consumer-facing NHTSA tool and no commercial vehicle-history aggregator exposes it.
The car and light-truck restriction. The file covers motorcycles, heavy trucks, buses, motorhomes, trailers, child seats and standalone equipment alongside cars. Because this is a used-car page, the headline basis excludes them, in three passes: an allowlist of Part 573 filers that build US passenger cars and light trucks; then probes whose every component label is CHILD SEAT, EQUIPMENT or a standalone tire label; then probes whose every model string is a motorcycle or medium/heavy chassis. The resulting per-probe call ships as a vehicle_class column in the frozen CSV so a reader can redo or reject it. The restriction takes the universe from 4,218 closed probes to 2,053 and raises the recall rate from 28.4% to 32.7% - the count roughly halves while the rate moves 4.3 points, so the two statistics do not degrade together and both bases are published.
Aggregation. Rates are distinct action numbers, not rows. Days open is close date minus open date. Per-company figures are grouped on the Part 573 filer field rather than the make name, with the multiple entity spellings of Mazda, Jaguar Land Rover and Mercedes-Benz consolidated by hand; because one investigation can list several filers, the per-company counts sum above the universe total.
Robustness checks, reported either way. Two of the four bite.
- The decade trend is the finding, and the all-history rate is misleading. 15.9% of 1980s probes ended in a recall, rising monotonically through 28.8%, 39.1%, 48.7% to 52.7% in the 2020s. The 32.7% all-history headline is dominated by the 1,165 probes opened before 2000 and understates what an investigation means now. The 2010s figure (48.7%, 236 probes, essentially all closed) is the reliable modern number.
- The preliminary-evaluation versus engineering-analysis gap is largely an accounting artefact. The naive read is 23.6% against 53.2%. Because a recall from an escalated probe is filed under the engineering analysis, crediting each preliminary evaluation with its successor's recall lifts the PE rate to 43.7%. The surviving effect is 35.2% (never escalated) against 54.2% (escalated), a ratio of 1.54 rather than 2.25, falling to 1.19 on probes opened since 2010.
- The per-company ranking is confounded by volume and era and is not load-bearing. Spearman correlation between a filer's probe count and its recall rate is -0.568; restricting to probes opened since 2000 moves Ford from last-but-one to above Toyota. Published as descriptive only, with the correlation stated.
- Censoring biases the recent rates upward, not downward. Among probes opened since 2010, those closing without a recall take a median 373 days against 224 for those producing one, so the 35 still-open probes skew toward no-recall outcomes. The 2020s rate is a ceiling.
The escalation-match rule is ours and is stated as such. The file carries no explicit "upgraded to" pointer, so a successor engineering analysis is matched on same filer, overlapping top-level component label, and an open date from 30 days before to 180 days after the preliminary evaluation's close. The match is window-sensitive - 39.4% at 30 days, 42.7% at 90, 53.0% at 180, 59.3% at a year - and every match is recorded in the successor_ea column so it can be audited.
Corroboration. The four load-bearing recall campaigns (26V363000, 24V321000, 26V621000, 26V227000) were re-queried against api.nhtsa.gov/recalls/campaignNumber and matched the investigations file exactly on manufacturer and component label, 4 of 4; unit counts and report-received dates are quoted from that API rather than from press coverage.
Known limits.
- An investigation outcome is not a safety verdict. NHTSA closes probes for many reasons, including a manufacturer acting before the agency has to, and a close with no recall does not establish that no defect existed.
CAMPNOrecords that a recall followed the investigation, not that the investigation caused it.- This file measures agency process, not road risk. It cannot support any claim about how dangerous a given model is, or how many unsafe vehicles are on the road.
- The decade comparison spans four decades of changing NHTSA practice, staffing and law. The rising rate may reflect how the agency works as much as how manufacturers behave, and nothing here separates the two.
- Open probes are excluded from every rate, which is what makes the censoring check in point 4 necessary.
