Answers are grounded in this paper and cite their sources.
Public crash reports can identify injury patterns within reported automated driving system (ADS) incidents, but cannot establish exposure-normalized safety. We analyzed the National Highway Traffic Safety Administration ADS file released on 15 July 2026, covering reports received from 16 June 2025 to 15 June 2026. After version selection, incident deduplication, engagement restriction and exclusion of unknown injury outcomes, 1,148 incidents remained. Any alleged injury occurred in 116 (10.1%); 39 (3.4%) involved hospitalization or fatality. Injury was alleged in 13 of 28 incidents involving vulnerable road users (46.4%; 95% confidence interval 29.5–64.2%) and 95 of 807 involving light motor vehicles (11.8%; 9.7–14.2%). The adjusted odds ratio was 6.48 (2.96–14.18). Repeated nested cross-validation yielded an area under the receiver operating characteristic curve (AUROC) of 0.605 and an area under the precision recall curve (AUPRC; average precision) of 0.161, compared with a no-skill AUPRC of 0.101. Class weighting did not materially improve discrimination. These incident-level associations support further investigation of vulnerable road user scenarios, but do not identify who was injured, establish causation, quantify crash incidence or demonstrate deployment utility.
No takes yet. Share an insight, caveat, or question.
Viktor et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: