2026-06-18 👁️ — views

Physical AI Safety Record — Waymo 6.8x Safer Claim, Tesla FSD Incidents, NHTSA Investigations, and What the Data Proves

Waymo claims 6.8x fewer injury crashes across 30M-plus driverless miles. Tesla FSD disengagement rate falls annually. Both need more unsupervised miles.

Article 151 in the Physical AI Benchmark Series — Physical AI Safety Record: Waymo’s 6.8x Safer Claim, Tesla FSD Incident Data, NHTSA Investigations, and What the Numbers Actually Prove

Safety is the central commercial and regulatory case for autonomous vehicles. Waymo has published peer-reviewed studies claiming 6.8x fewer injury-causing crashes than human drivers. Tesla publishes quarterly safety reports on FSD miles per intervention. NHTSA has opened and closed multiple investigations into both Autopilot and FSD. Yet the public narrative around AV safety is frequently shaped by headlines that do not distinguish between at-fault and not-at-fault crashes, supervised and unsupervised driving, or statistical significance and sample size.

This article is Article 151 in the Physical AI Benchmark Series. It benchmarks the published safety data, the methodology behind each claim, what the numbers actually prove versus suggest, and the key regulatory events that have shaped the public safety narrative. All figures labeled “(est.)” are derived from public disclosures, industry research, analyst estimates, and reported data rather than independently verified primary data. This article does not constitute investment advice.

Section 1 — Waymo’s Published Safety Record

Reading Waymo’s Safety Numbers

The 6.8x figure is the most-cited data point in AV safety discourse, and it is important to understand both what it demonstrates and what it does not. The study was peer-reviewed and published in a Nature partner journal — this is a meaningful bar. The methodology adjusted for exposure by road type and time of day, which is more rigorous than a simple miles-driven comparison. The result is the strongest single piece of published evidence that a commercial AV operator has achieved meaningful safety advantages over aggregate human driving.

The limitation is structural, not methodological: Waymo operates in a self-selected operational design domain. The good-weather, well-mapped, urban San Francisco environment is inherently safer than the full range of human driving conditions (ice, fog, rural roads, impaired drivers, school zones). A more conservative reading of the 6.8x figure is that Waymo is significantly safer than the average human driver driving under a much wider range of conditions — which is meaningful, but not the same as being 6.8x safer than a human driving under identical conditions.

Section 2 — Tesla FSD Safety Data

Understanding the Supervised vs Driverless Distinction

Tesla’s quarterly safety reports are among the most detailed in the industry, and the declining disengagement rate is a genuine positive signal about supervised FSD maturity. However, the critical distinction for safety benchmarking is that supervised FSD data — where a licensed driver is present and legally required to monitor the system — is not a proxy for unsupervised driverless safety.

The reason is not merely technical but statistical: a supervised driver intervening before an accident prevents data on what the AV system would have done without intervention. The disengagement data tells us how often drivers felt the need to take over, which correlates with safety but does not equal it. Tesla’s Austin Robotaxi launch in 2026 represents the first real-world accumulation of unsupervised safety data — but weeks of data across tens of vehicles is too small a sample for meaningful comparison to Waymo’s 30M-plus miles.

Section 3 — NHTSA Regulatory Events Timeline

The NHTSA Investigation Lifecycle

NHTSA investigations are frequently reported as safety failures, but the investigation lifecycle tells a more nuanced story. An NHTSA investigation is opened based on a pattern of reported incidents — it does not imply that a defect has been found or that the system is unsafe. The majority of Tesla AV-related investigations have closed with OTA software updates rather than physical recalls, which is a structurally positive outcome: the safety issue was identified, addressed, and deployed to the entire fleet without owners needing to visit a dealer.

The 2022 Autopilot recall — 830,000 vehicles addressed via OTA — was the largest by vehicle count in automotive history. The fact that it required no physical action from owners and was completed in days rather than years represents a new model for automotive safety remediation. This OTA resolution model is both a Tesla competitive advantage and a template that regulators are still developing frameworks around.

The Cruise October 2023 incident deserves separate treatment because it is frequently cited alongside Waymo and Tesla incidents without noting that Cruise is a different company and a different system. The Cruise incident — in which a pedestrian who had already been struck by a human driver was subsequently dragged by a Cruise AV that failed to recognize the pedestrian under the vehicle — was a genuine software failure with serious consequences. Cruise suspended operations and the CA DMV revoked its permit. This incident had no equivalent at Waymo or Tesla, but it materially damaged public and regulatory confidence in the AV sector broadly.

Section 4 — How to Read AV Safety Statistics: Methodology Matters

The Statistical Honesty Required for Fair AV Safety Assessment

The most important methodological point in AV safety benchmarking is the difference between absence of evidence and evidence of absence. Waymo’s zero at-fault fatalities in 30M-plus commercial driverless miles is a meaningful positive result, but it is not statistical proof that Waymo is dramatically safer than human drivers on fatalities. At the US human baseline of approximately 1 fatality per 100M miles, Waymo’s 30M-mile fleet would statistically expect 0.3 fatalities — and zero observed is consistent with that expectation. The result would be more statistically meaningful at 1B miles, and decisive at 10B miles.

This is not a criticism of Waymo’s safety record — zero at-fault fatalities in commercial driverless operations is an extraordinary achievement. The point is epistemic: extraordinary claims (6.8x safer, zero fatalities) require the statistical sample to support the confidence level of the claim. The AV industry is still in the data-accumulation phase, and honest safety benchmarking requires acknowledging this limitation alongside celebrating the genuine progress.

Section 5 — Safety Benchmark Scorecard

Overall Verdict

Waymo has the most rigorous published AV safety record in the world: zero at-fault fatalities in 30M-plus driverless commercial miles, a peer-reviewed 6.8x safety advantage study (with methodology caveats), and 4-plus years of incident disclosure to CA DMV. Tesla FSD has a strong supervised safety record with declining disengagement rates, but its unsupervised commercial safety record is 2026-early-stage with insufficient data for comparison.

The honest conclusion: Waymo has demonstrated meaningful safety in its operational design domain; Tesla has demonstrated improving supervised safety; both need more unsupervised driverless miles to make statistically robust claims about human-level safety parity. The next 24 months — as Tesla’s Austin Robotaxi operation accumulates commercial driverless miles — will be the most important period yet for closing the comparison gap.

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

Sources

• Waymo safety study — peer-reviewed, Nature partner journal ↗
• Tesla quarterly vehicle safety report — Tesla ↗
• NHTSA AV incident database — NHTSA ↗
• California DMV AV incident reports — CA DMV ↗
• Cruise pedestrian dragging incident — NHTSA investigation ↗

Tags