The Reporting Gap: How Many AV Crashes Go Unreported?

The 30-Second Rule: A Built-In Blind Spot

NHTSA's Standing General Order (SGO 2021-01) requires manufacturers to report crashes where the automated driving system was engaged within 30 seconds before the crash. This sounds reasonable — but it creates a significant loophole.

Consider this scenario: A Tesla driver is using Autopilot on a highway. The system encounters a situation it can't handle and disengages, handing control back to the driver with a warning chime. The driver, who may have been inattentive (relying on Autopilot), takes 31 seconds to process the situation and crash. Under the SGO rules, this crash doesn't need to be reported as an AV-related incident. The automation was involved — it created the conditions for the crash — but the 30-second clock ran out.

How often does this happen? We don't know. Tesla doesn't publish disengagement data the way California requires of ADS operators. The 30-second threshold was chosen somewhat arbitrarily, and there's no public analysis of how many crashes fall just outside that window. Even a 35-second or 45-second threshold would capture significantly more incidents where automation played a causal role.

Self-Reporting: The Fox Guarding the Henhouse

Under the SGO, manufacturers report their own incidents. There is no independent audit, no third-party verification, and no mechanism for NHTSA to independently discover unreported crashes. If a manufacturer doesn't report a crash — whether through oversight, interpretation of the rules, or deliberate omission — NHTSA may never know about it.

The reporting requirements specify that fatal or serious-injury crashes must be reported within 1 business day, and other crashes within 10 business days. Updated reports are required when new information becomes available. But "new information" is subjective, and the timeliness of reporting varies widely. Some manufacturers file detailed reports within hours. Others submit bare-minimum reports weeks later.

For ADAS manufacturers like Tesla, the challenge is compounded: Tesla may not learn about a crash involving its vehicle until the owner reports it, a police report is filed, or media covers the incident. Unlike Waymo, which has real-time telemetry from every vehicle in its fleet, Tesla relies on a more fragmented information pipeline.

The TeslaDeaths.com Discrepancy

TeslaDeaths.com, maintained by independent researchers, has documented at least 65 deaths involving Tesla's Autopilot or FSD features. The NHTSA SGO database shows 56 Tesla fatalities. That's a gap of at least 9 deaths — a 16% discrepancy.

Where do the missing deaths come from? Several possibilities:

• Pre-SGO deaths: Some Autopilot fatalities occurred before the SGO took effect in June 2021. The NHTSA database only captures incidents reported under the SGO framework.
• 30-second rule exclusions: Crashes where Autopilot disengaged more than 30 seconds before impact may not appear in the SGO data.
• Definitional disagreements: Tesla may determine that certain crashes didn't involve Autopilot engagement based on their telemetry data, while independent investigators disagree based on witness accounts or vehicle evidence.
• Delayed reporting: Some incidents may still be working through Tesla's internal review and haven't been submitted to NHTSA yet.

What Falls Through the Cracks

Beyond the 30-second rule and self-reporting issues, several categories of incidents systematically escape the SGO database:

Near-misses and dangerous behaviors that don't result in contact. If Autopilot swerves into an oncoming lane and the driver corrects just in time, no crash occurred — and nothing gets reported. These near-miss events could reveal systemic issues but are invisible to NHTSA.

Crashes where the driver disables the system before impact.If a Tesla driver sees an impending crash and quickly turns off Autopilot by grabbing the wheel and braking, the system wasn't "engaged at the time of the crash" — even though it created the dangerous situation.

Low-speed property damage that goes unreported. A Tesla backing into a garage wall while using Smart Summon might cause $2,000 in damage. The owner may fix it out of pocket and never report it. Under the SGO, a crash with only property damage must still be reported if automation was engaged — but enforcement depends entirely on the manufacturer knowing about it.

International incidents.The SGO only covers U.S. crashes. Tesla operates Autopilot globally, and foreign incidents — including fatalities — don't appear in NHTSA's data.

How Complete Is the ADS Picture?

Interestingly, the reporting gap may be smaller for fully autonomous (ADS) operators like Waymo. Because Waymo operates a fleet of company-owned vehicles with comprehensive telemetry, it knows about every incident in real-time. Waymo also operates under California DMV reporting requirements that are separate from (and often more stringent than) the federal SGO. The incentive and ability to report are both higher for fleet operators than for consumer vehicle manufacturers.

Cruise, before pausing operations, similarly had complete fleet visibility. The Cruise pedestrian-dragging incident in San Francisco was actually a case where under-reportingby the company led to regulatory consequences — the California DMV revoked Cruise's permit partly because Cruise initially failed to disclose the full severity of the incident.

What Would Fix the Gap

Several reforms could narrow the reporting gap. Extending the engagement window from 30 seconds to 5 minutes (or eliminating it entirely for fatal crashes) would capture more causally-related incidents. Requiring manufacturers to submit telemetry data alongside crash reports would enable independent verification. Mandating near-miss reporting for known dangerous system behaviors would reveal patterns before they produce fatalities. And creating a confidential reporting channel for vehicle owners — separate from the manufacturer's own reporting pipeline — would reduce dependence on self-policing.

Until these reforms happen, the 6,215 incidents in the NHTSA database represent a lower boundon AV/ADAS crash reality. The true number is higher. How much higher? That's the gap we can't yet measure — and that uncertainty itself is a safety concern.