← TelemetryLab Research
Research · NHTSA Defect Corpus · 2022 to 2026

The Recall Signal Gap

Every U.S. vehicle recall from 2022 to mid-2026, checked against the signals a connected car already produces. The failure was often visible in that signal before the recall was issued.

Public NHTSA ODI corpus · 977,000 rows · all vehicle recalls, hardware and software · reproducible (see methodology)

When a car is recalled, every owner of that model year gets the same letter. Most of those cars were never going to fail. The defect lived in a slice of the production run, but the recall covers all of it, and it lands after the failures have already started.

Recalls work this way because a manufacturer can identify what it built, not what is actually breaking. So they are late, oversized, and physical. Since the start of 2022, 4,030 campaigns have put 147 million vehicles into recall action. We rebuilt the entire public NHTSA record and checked each recall against the signals a connected vehicle already produces. The failure was often there first.

The thesis

A recall runs on the build sheet, not the failure.

A campaign pulls in the whole production run, then waits for the failures to show. Only a fraction of those cars ever fail, and a quarter of the recalled population is never brought in at all. The failure signature, meanwhile, often sat on the car's own bus for years. Run the recall on the signal instead: catch the signature early, call in the cars already showing it, confirm the remedy took, and close software campaigns over the air where the car allows it. This paper measures how far that reaches.

Available is not the same as captured

One thing to be clear about up front. These signals exist on the vehicle today, thrown off every time the car runs. But almost none are captured, transmitted, or analyzed: they are dumped at key-off, stranded inside a supplier's module, or never logged. When this paper calls a recall reachable today, it means a fleet could start collecting that signal with no new hardware, not that anyone collects it now. Today essentially none of it is used. Reading and processing it is the product.

Part AWhat the recall record says

A1 The blunt instrument

The recalled population dwarfs the set that actually fails, and dwarfs the set that ever gets fixed.
Reported complaints are tiny beside the recalled population
Complaints per 1,000 recalled vehicles. Even the noisiest domain stays under 3% of the recalled population, a directional proxy, not a defect rate.
A quarter of recalled cars are never brought in
Recall completion, and how it falls with age. A quarter of recalled vehicles never come in for the remedy.
The US fleet now averages more than 12 years old (S&P Global Mobility), so recalls accumulate. Completion: NHTSA; Alliance for Automotive Innovation.
1 in 4
recalled vehicles never comes back for the fix, leaving tens of millions of cars on the road with an open recall.

A2 The landscape

Hardware is the volume. Software is the fast, hidden slice. Both are mostly physical to fix.
Hardware is the volume; software is the fast 15%
85% of campaigns are hardware. Software is 15% of campaigns but 30% of the vehicles, and the fastest-growing slice.
Recalls run flat near 900 campaigns a year
2022 through 2025 each run near 900 campaigns a year.
2026 is Jan to Jun 12 only (partial). 2022 through 2025 are full years.
Legacy exposure stacks up in Level 3; Tesla is the all-signal benchmark
Campaigns by access tier. L1, L2, and L3 are exclusive and sum to each maker's total: L1 capturable today with no new hardware, L2 with one integration, L3 needs new architecture. None is captured today. The L3 column is the legacy danger. Vehicle counts are campaign-exposure.
ManufacturerCampaignsVehiclesL1L2L3L1+L2 reach
Ford37643.6M707023620.0M
Stellantis18215.6M37381077.6M
Honda7714.1M713573.8M
Toyota8010.8M818544.8M
Hyundai1038.0M1613743.6M
VW Group1644.6M1254982.8M
GM1308.7M2321864.0M
Tesla6312.5M3113112.3M
Tesla's software recalls all ship over the air (30 of 30), so 98% of its recalled vehicles are L1 or L2 reachable, the benchmark for a software-defined fleet.

A3 Software, the cleanest case

The root cause is software. The delivery is a dealership. ~4 in 5 times.
Software is 15% of campaigns but the sharpest illustration of the gap: even when the remedy is pure code, the architecture cannot deliver it remotely.
More than half of software recalls were pure code that still forced a dealer trip
All 612 software recalls, by how they shipped. The amber slice is the closed-loop gap: pure code, dealer visit anyway.
Readable off the OBD port today: 156 of 612 · 9.49M vehicles
What TelemetryLab reads here

We do not build the over-the-air pipe. From signals already on the bus, we catch the signature, flag the cars already showing it, and confirm the fix took. And we measure the dealer-path gap: where the remedy is code, but the car has no remote route to receive it. Tesla is the counter-proof. 30 software recalls, every one shipped over the air.

The public signal led this recall by two years
Ford fuel injector, 694K vehicles, recalled July 2025. Monthly public complaints, a lagging shadow of the standard OBD signals the car was already throwing.
A 2x-baseline detector on the public stream crosses its line 21 to 30 months before the recall (by sensitivity). The on-vehicle signal leads the public one, so the port reads it earlier still. An earlier-investigation trigger, not a root-cause backtest.

A4 Stakes and timing

The domains where software fails carry crashes and deaths in the complaint record, and the faults sit in the field a median 2.89 years before the recall.
Where software fails, the complaints include crashes and deaths
Complaint-reported and unverified, domain-level, not tied to a specific recall. Directional severity. ADAS, braking, and camera carry the cleanest software attribution.
DomainComplaintsCrashesDeaths
Powertrain / stall65,3892,290231
ADAS (FCW/AEB/lane)30,4032,806151
Braking / stability31,3652,57137
Rear camera (backover)6,76242331
Software hides in the field over a year longer than hardware
Software vs hardware. Continuous signal compresses this window.
Median build-to-recall over the 33% of campaigns with a usable build date. Continuous signal compresses this window.
Part BHow much is reachable from signal
How we read the value

We classify every recall by where its failure signature lives as a signal, then read the worth in two layers and never blend them. The first is measured today: what the record proves is already being lost while the signal goes uncaptured, the dealer-path gap, the 2.89-year blind window, the repeat-remedy pool. The second is estimated recovery: what capturing and acting on the signal is worth, every assumption stated. The headline is always the measured layer; the estimates stack on top, in their own color, never hidden inside it.

B1 Three levels of access, one platform

About half the vehicles under recall already throw off a signal that would catch the defect. A third need no new hardware to read it. Almost no one reads it today.
The access tier is about where the signal already lives and how hard it is to capture, not whether anyone captures it today (no one does). The platform is the same at every level: collection, detection, diagnostics, prognostics, and action as one closed loop. What changes is how much signal it can see.
Nearly a third is on the port today, about half with one integration
Every vehicle under recall, in exactly one tier (the three sum to the whole fleet). The tiers rank how hard the signal is to capture, not whether anyone captures it: today almost none is. Left of the dashed line needs no new architecture.
Level 1 · today

No new hardware

OBD access, or the OEM telematics layer
461 recalls · 44.3M vehicles · $2.0 to 8.1B

Signals the car already produces but no one collects: powertrain, fuel, emissions, stability, ABS, and any fault it already flags. Two paths, no rewiring: read the OBD port, or emit signals the TCU already sees.

  • Detect the failure signature earlier
  • Prioritize the vehicles already showing it
  • Verify the remedy took
  • Close software campaigns remotely where the car supports it
Level 2 · integration

One standard API

Tier-1 modules emit what they already compute
469 recalls · 27.8M vehicles · $3.8 to 17.0B

The ADAS, camera, airbag and BMS signals already exist inside Tier-1 modules; they are just not on a collectible bus. A standard API to emit them is a real integration, not full SDV architecture, and no new sensor or harness.

  • Brings the accelerating camera and ADAS wave within reach
  • Feature-state and module faults become fleet-visible
  • The largest single jump in reachable exposure
Level 3 · ceiling

Software-defined vehicle

Full signal availability + standardization
3,100 recalls · 74.8M vehicles · $14.6 to 68.3B

The mechanical and structural residual that carries no signal today. Full SDV architecture plus predictive models is where the system moves from remediation toward prevention. The long build, and the real technical ceiling.

  • New sensing on the failure modes that are blind today
  • Prediction ahead of the defect
  • Requires Tier-1 and OEM standardization
What each tier of reach is worth
Exposed remediation burden per tier (the tiers are exclusive and sum to all recalls). The highlighted row is capturable today plus one standard integration.
Access tierVehiclesReachExposed burden
L1 today, no new hardware44.3M30%$2.0 to 8.1B
L2 + standard Tier-1 API27.8M+19%$3.8 to 17.0B
L1 + L2 reachable72.0M49%$5.8 to 25.1B
L3 full SDV, new sensing74.8M51%$14.6 to 68.3B
All recalls146.8M100%$20.5 to 93.4B

B2 What turns reach into dollars

What flying blind costs today, and what capturing the signal would recover. We never blend measured and estimated.
Measured todayThe cost of flying blind
Counted straight off the record: what the industry loses now, with the signal uncaptured.
Software dealer-path gap~329 recalls

Of 612 software recalls, about 329 were pure code that still forced a dealer trip, at roughly $40 to $120 a vehicle in reflash labor and logistics. Over-the-air would collapse most of that wherever the architecture allows a remote route.

The blind window2.89 yr

Software sits a median 2.89 years in the field before its recall. Continuous signal compresses that window, and a smaller window means a smaller eventual recall.

Repeat-remedy pool26 recalls · 4.68M

26 software recalls (4.68M vehicles) cite a prior failed fix. Per-vehicle verification keeps the same owners from being called back twice.

Estimated recoveryWhat capturing it is worth
Illustrative estimates, each assumption stated. What the measured loss becomes once the signal is captured and acted on.
Shrink the dealer billup to ~$2B

Dealer-path software recalls carry roughly $0.7 to 2.2B in reflash-and-service cost (this paper's own cost band). The pure-code share could ship over the air, recovering much of it wherever the architecture allows a remote route.

Lift completion by targeting~3.6M cars

A quarter of recalled vehicles never return, leaving the hazard on the road. Aiming outreach and service capacity at the cars that actually show the signature lifts completion: a 5-point gain on the 72M reachable vehicles is about 3.6M more cars fixed.

Catch it earlier, shrink scopeyears sooner

Software sits a median 2.89 years in the field before recall. Continuous signal compresses that window, so fewer vehicles are built with the defect before it is caught, and the eventual recall population shrinks.

How the loop would handle a real recall

This is a worked illustration, not a deployment. The recall is real: Ford fuel-injector fire, 694K vehicles, July 2025, and its public signal really did lead the recall by nearly two years. Here is how the loop would have run it:

SIGNALMisfire, fuel trim, rail pressure. Standard OBD-II.
TIERL1, readable off the port today.
DETECT21,306 public complaints, leading the recall by 21 to 30 months. The on-vehicle signal leads even that.
REMEDYHardware, so a dealer visit, but only the affected build.
VERIFYConfirm per vehicle the fix took.
Methodology and limits

Built from the full NHTSA ODI corpus compiled into a 977,000-row database, all vehicle recalls 2022-01 to 2026-06. Vehicle counts are campaign-exposure (potentially-affected units, overlap-inclusive), not unique VINs. Cost is an illustrative band from per-remedy assumptions, cross-checked against the AlixPartners ~$500 per vehicle industry blend (~$73B).

Access tiers are a keyword-and-component classification over NHTSA's own text, conservative by design, so L1+L2 at 49% of vehicles is a floor, not a ceiling. The complaint-based failure proxy and the targeting lever are directional, labeled modeled, never folded into the headline. A safety recall's legal population is the regulator's to set; available signal makes that population smaller when caught early, the remedy faster and verifiable, and for software remote. Each recall's tier follows from NHTSA's own published text, so every assignment is reproducible.

Source: NHTSA ODI recalls, complaints, and investigations (static.nhtsa.gov/odi/ffdd), pulled 2026-06-12, compiled into a 977,000-row database. External anchors: recall completion rates (NHTSA; Alliance for Automotive Innovation); ~$500 per vehicle blended recall cost (AlixPartners); software-dealer reflash cost $40 to $120 per vehicle (this paper's cost assumption). Prepared by TelemetryLab. Every figure is reproducible from the public NHTSA record.