2026-06-18 — views

Physical AI Safety Record — Waymo 6.8x Lower Injury Crash Rate vs Human Drivers and Tesla FSD NHTSA History Benchmarked

Waymo: 6.8x lower injury crash rate vs humans across 10M-plus driverless miles; Tesla FSD had 2 NHTSA OTA recalls; safety data is the AV permit currency.

Article 133 in the Physical AI Benchmark Series — Physical AI Safety Record and Incident Database: Waymo vs Tesla Crash Rates, NHTSA Investigations, and Whether Safety Data Is the AV Industry’s Most Competitive Moat

Safety is not merely an ethical imperative for autonomous vehicle companies — it is the primary regulatory currency. The company with the best demonstrated safety record earns permits, expands to new cities, and sets the benchmarks that every competitor must meet. Three public data streams determine how fast any AV company can grow: NHTSA investigations and recalls, California DMV autonomous vehicle incident reports, and miles-per-disengagement (or equivalent driverless safety metrics). This article maps safety as a Physical AI benchmark dimension, comparing Waymo’s published driverless crash data against Tesla FSD’s supervised-mode NHTSA history to assess which company holds the decisive safety moat today.

All figures labeled “(est.)” are derived from public disclosures, industry analyst estimates, and reasonable inference rather than independently verified primary data.

Section 1 — Waymo Safety Record: The Data

Waymo has been more aggressive than any AV competitor in publishing detailed safety statistics. Its 2024 Nature Communications study — the most rigorous publicly available AV safety comparison — covers driverless commercial operation in Phoenix and provides a direct statistical comparison against human drivers in comparable urban conditions. The data below synthesizes that study and Waymo’s annual safety reports with CA DMV records.

Metric	Waymo Data	Source / Notes
Miles of driverless commercial operation	10M-plus miles fully driverless commercial (as of early 2026, est.)	Waymo discloses cumulative driverless miles periodically; exact current figure not always published
Airbag-deployment crash rate	~0.41 per million miles (driverless) vs ~2.78 for human drivers in comparable conditions	Waymo-published Nature Communications study 2024; covers Phoenix driverless operations
Injury-causing crash rate	~0.68 per million miles (driverless) vs ~4.62 for human drivers	Same study; covers airbag-deployment plus injury-producing incidents
Injury crash rate improvement vs human baseline	Approximately 6.8x lower injury-producing crash rate than human drivers in comparable urban conditions	Waymo-published; independent external replication limited
CA DMV incident reports	Hundreds of incident reports filed since 2017; vast majority are minor (rear-ended by human driver while stopped)	CA DMV Autonomous Vehicle Incident Reports — public database
NHTSA Special Order (2023)	Waymo received NHTSA Special Order WP-23-001 covering 22 incidents (primarily minor collisions); investigation closed 2024 with no recall action	NHTSA Special Order WP-23-001
At-fault crashes in driverless mode	Extremely rare in public record; most CA DMV reports show Waymo vehicles were stopped or moving slowly when struck by human drivers	CA DMV public reports; Waymo internal classification
Safety metrics used	Collision Avoidance at Fault (CAF) and Contact Avoidance at Fault (COAF) — published in Waymo 2024 Safety Report	Waymo 2024 Safety Report

Waymo’s safety reporting posture is a deliberate strategic asset. By publishing peer-reviewed data in Nature Communications rather than keeping statistics internal, Waymo creates a public, citable safety record that regulators can reference in permit decisions. This is not philanthropy — it is regulatory engineering. A company with a peer-reviewed paper showing a 6.8x lower injury crash rate than human drivers has a qualitatively different permit application than a competitor that says “trust us.”

The CA DMV incident database reveals the character of Waymo’s incidents: the overwhelming majority involve a human driver hitting a stationary or slow-moving Waymo vehicle, not a Waymo vehicle striking someone. This matters for liability, insurance pricing, and public trust in ways that aggregate accident counts do not capture. Counting incidents without fault attribution overstates Waymo’s risk profile.

Section 2 — Tesla FSD Safety: The Data

Tesla’s safety data situation is structurally different from Waymo’s in one critical respect: FSD is legally a supervised driver-assistance system, not an autonomous vehicle. This means Tesla is not required to participate in CA’s AV pilot program, does not file CA DMV disengagement reports for FSD, and human drivers bear ultimate legal responsibility for FSD-engaged crashes. The comparison to Waymo is therefore apples-to-oranges until Tesla’s driverless (no safety driver) commercial operations accumulate sufficient data — which the Austin robotaxi launch is beginning to provide.

Metric	Tesla Data	Source / Notes
Cumulative miles with FSD engaged	~5-6 billion miles (supervised FSD; est.)	Tesla AI Day disclosures plus fleet mile extrapolation; not all miles represent full FSD engagement
FSD critical disengagement rate	~0.03 per 1,000 miles (FSD v14, est.)	Tesla has not published a direct disengagement rate; CA AV pilot reports predate FSD v12
NHTSA Autopilot investigation (2021)	NHTSA opened PE21-020 covering crashes involving Tesla Autopilot at stationary emergency vehicles; Tesla issued OTA recall in 2022 covering ~300,000-plus vehicles	NHTSA PE21-020
NHTSA FSD recall (2023)	~362,000 vehicles recalled via OTA for FSD beta failing to fully stop at stop signs in rolling right turns and not correctly responding to certain traffic signals	NHTSA recall 23V-016 (February 2023)
Tesla Autopilot crash rate (NHTSA SGO data)	Tesla submitted to NHTSA: Autopilot-engaged crash rate ~0.27 per million miles vs ~2.82 per million miles without Autopilot (2022 Standing General Order data)	NHTSA Standing General Order; Tesla submission
FSD vs Autopilot distinction	NHTSA and media frequently conflate Autopilot (highway driver assistance) with FSD (urban driving); these are different systems with different risk profiles	Common analytical error; Tesla has pushed back on conflation in NHTSA correspondence
Austin driverless launch	Tesla’s Austin supervised robotaxi launch (2025-2026) adds a new safety layer requirement: Texas AV framework requires manufacturer safety self-certification	Texas HB 1892 AV framework; Tesla self-certification model

The NHTSA data shows a pattern that is simultaneously reassuring and structurally distinct from Waymo’s: Tesla Autopilot-engaged driving shows approximately 10x lower crash rates than non-Autopilot driving in the same Tesla vehicles. But this is a supervised highway assist system compared to human driving in the same vehicle — a very different comparison from Waymo’s urban driverless performance versus all human drivers in urban conditions. Both data points show improvement over human baseline, but the risk profiles being measured are not equivalent.

The two OTA recalls are the most significant regulatory marks on Tesla’s record. Both were resolved without physical service visits (a major OTA advantage), but both also generated sustained negative media cycles and NHTSA scrutiny that created regulatory friction for Tesla’s broader driverless ambitions.

Section 3 — Safety Incidents: Head-to-Head Comparison

Category	Waymo	Tesla FSD	Notes
NHTSA major investigations	1 Special Order (2023, closed with no recall)	2-plus major investigations plus recalls (Autopilot 2021, FSD 2023)	Tesla has faced more significant NHTSA regulatory scrutiny
Published peer-reviewed safety studies	Yes — Nature Communications 2024 (Waymo-authored) showing 6.8x lower injury crash rate	No equivalent peer-reviewed study for driverless mode	Waymo has been substantially more proactive in publishing safety data
CA DMV disengagement reporting	Pre-driverless era reporting complete; Waymo phased out disengagement reports when fully driverless (no human driver = no disengagement to report)	Tesla never in CA AV pilot program; FSD supervised = not subject to CA AV disengagement reporting	Different regulatory regimes; direct comparison not straightforward
NHTSA recall history	No recalls	FSD beta recall 2023 (~362K vehicles); Autopilot recall 2022 (~300K-plus vehicles)	Tesla has more recall history; both resolved via OTA
At-fault driverless crashes	Rare in public record; most Waymo incidents involve being struck while stopped	Not applicable (FSD is supervised; human driver is legally responsible)	Structurally incomparable until Tesla achieves driverless commercial operation at scale
Human vs AV crash rate improvement	~6.8x lower injury crash rate than humans (driverless, Phoenix, Waymo-published study)	~10x lower Autopilot crash rate vs non-Autopilot in same vehicles (highway, NHTSA SGO data)	Both beat human baseline; methodologies, operating environments, and driving modes differ substantially

The comparison reveals a fundamental asymmetry: Waymo has more incident reports (because CA DMV requires reporting) but a better regulatory track record (no recalls, one closed investigation). Tesla has fewer public incident filings (because FSD is not subject to CA AV pilot reporting) but more significant NHTSA enforcement actions. This illustrates why raw incident count is a misleading safety metric — what matters is fault attribution, severity, and regulatory outcome.

Section 4 — Safety Data as Competitive Moat

Safety data is not simply a compliance cost for AV companies — it is a revenue-generating strategic asset. The company that can demonstrate the best safety record gets permits faster, enters more cities, earns lower insurance rates per mile, and builds public trust that accelerates consumer adoption. Safety performance directly determines the economic velocity of an AV business.

Moat Dimension	Waymo	Tesla	Strategic Implication
Data volume	10M-plus driverless commercial miles (est.); slower accumulation rate	5-6B supervised miles (est.); much higher volume but supervised mode = different risk profile	Tesla has raw volume; Waymo has driverless-mode quality; both matter for different things
Regulatory currency	Each new city permit depends on demonstrated safety; Waymo’s track record is its primary permit application asset	Texas driverless permit uses self-certification; CA driverless would require more (Tesla not yet applying)	Waymo’s published safety data is its fastest path to new city permits
Insurance pricing advantage	Demonstrated driverless safety record → lower insurance cost per mile → better unit economics	No driverless safety record yet; Austin robotaxi fleet insurance cost is an open question	Waymo’s safety record directly improves its cost structure per revenue mile
Liability shield	Published safety data creates a legal defensibility argument: “our system is demonstrably safer than human drivers”	No equivalent published data for driverless mode; FSD liability rests with human driver	Waymo’s safety publications function as pre-litigation defense in any future mass-tort scenario
Public trust	Waymo incident data largely transparent via CA DMV; builds long-term public trust through consistent transparency	Tesla NHTSA investigations and recalls generated negative media cycles even when OTA-resolved	Transparency in safety data → long-term public trust → faster consumer adoption
Sensor redundancy argument	Lidar + camera + radar = three independent sensing modalities; failure of one is caught by others	Camera-only = single sensing modality; proponents argue AI model redundancy replaces sensor redundancy	Unresolved fundamental debate; lidar proponents cite safety; Tesla cites human drivers using vision only

The insurance moat is underappreciated. Traditional specialty insurance for AV fleets (Lloyd’s syndicates, etc.) prices a poorly understood risk category conservatively. A company with a 10M-mile, peer-reviewed safety record can negotiate lower premiums per mile than a competitor without one — and at fleet scale of 50,000 or 100,000 vehicles, that per-mile cost difference compounds into a structural economic advantage.

The liability shield dimension compounds over time. As AV litigation becomes inevitable at commercial scale, the company with the most robust published safety record — especially peer-reviewed publications that independent courts can evaluate — has a qualitatively different legal exposure than a competitor relying on internal data. Waymo’s Nature Communications publication is not just a PR move; it is establishing an evidentiary baseline.

Section 5 — Safety Benchmark Scorecard

Dimension	Waymo	Tesla	Edge
Published driverless safety data	Yes — peer-reviewed, covers 10M-plus driverless miles	No equivalent for driverless (FSD is supervised)	Waymo
NHTSA investigation severity	1 Special Order, closed with no recall	Multiple investigations plus 2 major OTA recalls	Waymo
Incident transparency (CA DMV)	Full reporting (pre-driverless era) plus annual safety reports	Not in CA AV pilot = no CA reporting obligation	Waymo (voluntary transparency strategy)
Crash rate vs human baseline	~6.8x lower injury crash rate (driverless, Phoenix, Waymo-published)	~10x lower Autopilot crash rate (highway only, NHTSA data, supervised)	Both beat human baseline; Waymo’s is driverless-mode
Regulatory permit track record	Full driverless commercial permits in AZ, CA, TX	Self-certified in TX; no CA driverless permit yet	Waymo
Safety as a permit moat	High — published safety data, permit track record, established regulatory relationships	Growing — Austin driverless launch will begin building the driverless safety record	Waymo holds decisive edge today; Tesla is closing as Austin data accumulates
OTA recall capability	No recalls issued	2 major OTA recalls resolved without physical service visits	Tesla demonstrates OTA fix velocity; Waymo has not needed it
Sensor stack for safety	Lidar + camera + radar (three-modality redundancy)	Camera-only (FSD v12-plus; no lidar)	Waymo on redundancy; ongoing debate on necessity vs AI sufficiency

The scorecard shows a consistent pattern: Waymo has a decisive safety data advantage in every dimension that matters for regulatory permitting today. Tesla’s advantage is structural for the future — the Austin driverless launch is the beginning of a driverless safety record that will accumulate over time, and Tesla’s OTA capability means that when regulators identify issues, Tesla can respond within days rather than months.

The key strategic question is whether Tesla can accumulate enough driverless safety data in Texas (a self-certification state) to support driverless permit applications in California and other states with more rigorous regulatory frameworks before Waymo’s geographic lead becomes insurmountable. If Tesla’s Austin operation runs cleanly for 12-18 months and accumulates meaningful driverless miles, it will be able to publish safety data that begins to close the gap. If the Austin operation encounters significant incidents in its early months, it will extend Waymo’s regulatory lead substantially.

Safety data is not a static moat — it is an accumulating one. The company that generates more driverless commercial miles with lower at-fault incident rates will continuously extend its regulatory advantage. Today, that company is Waymo. Tomorrow depends on Austin.

Note: All figures labeled “(est.)” are derived from public market information, company disclosures, analyst estimates, and industry reports as of mid-2026. This article does not constitute investment advice.