A Note on Tesla's Revised Safety Report Crash Rates

Goodall, Noah · 2023 · arXiv

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the lack of transparent, standardized data regarding the crash rates of Level 2 automated driving systems (ADS), specifically focusing on Tesla’s Autopilot. Because police records and government databases often fail to capture whether ADS was active during a crash, researchers rely heavily on manufacturer-reported data. The study investigates significant discrepancies in Tesla’s quarterly safety reports, which were revised in January 2023. Prior to March 2021, Tesla categorized crashes into three groups: with Autopilot engaged, without Autopilot but with active safety features, and without either. In 2023, Tesla consolidated these into two categories: “using Autopilot technology” and “not using Autopilot technology,” while also adjusting for double-counting and excluding crashes that did not involve airbag or restraint activation. The author analyzes pre-revision and post-revision crash rates from Q3 2018 to Q4 2023 using data archived by the Internet Archive. The analysis reveals that while Autopilot crash rates changed modestly (an average 3.6% reduction), the non-Autopilot category showed counterintuitive shifts. The revised “not using Autopilot” crash rates align closely with the historical “without active safety features” rates, despite evidence that drivers rarely disable active safety systems like automatic emergency braking, which are enabled by default. Statistical analysis of relative weights suggests that for the revised rates to hold, vehicles would need to have traveled approximately 6.3 times further without active safety features than with them, a scenario deemed highly unlikely given driver behavior patterns. Furthermore, the variance in crash rates for the “with active safety features” group was lower than for the “without” group, contradicting the expectation that the larger sample size (presumably the “with” group) should exhibit less variability. The findings indicate that Tesla’s revised reporting methodology is heavily biased toward crashes occurring without active safety features, contrary to operational realities. The author evaluates potential explanations, such as universal quantification of features or differing classification thresholds, but finds them insufficient due to the incompatible operational speed ranges of various safety features. The most plausible explanation is that Tesla excluded or undercounted a portion of crashes involving active safety features during the revision. The paper concludes that these discrepancies raise serious concerns about data transparency and sample size integrity. It emphasizes the need for clearer reporting methodologies from manufacturers to ensure that researchers and regulators can accurately assess the safety performance of advanced driver assistance systems.

Key finding

Tesla's January 2023 retroactive revision left Autopilot crash rates roughly unchanged (mean -3.6%) but shifted the non-Autopilot category to closely match the prior 'no active safety features' subgroup, despite HLDI evidence that ~93% of drivers keep front crash prevention enabled and that Tesla's automatic emergency braking is on by default each ignition; the revised non-AP benchmark therefore appears biased upward, distorting Autopilot-vs-non-AP safety comparisons.

Methodology

field_study

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via discover_arxiv on 2026-05-04 (3 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• incidence prevalence
• adas effectiveness
• naturalistic crash near crash

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes, observational prevalence
• Methodological Resource: dataset resource