What Crash Reports Actually Say: A Text Analysis

Every crash reported to NHTSA under the Standing General Order includes a narrative — a free-text description of what happened, written by the manufacturer or the vehicle operator. These narratives are a goldmine of qualitative data, but they're rarely analyzed systematically. We applied natural language processing to 4,218 crash narratives to identify the most common terms, failure categories, and patterns.

The Words That Define AV Failures

The most frequent term across all crash narratives is "lane change", appearing in 21.1% of reports. This encompasses both the automated system making an unsafe lane change and the system failing to respond when another vehicle changes lanes. It's followed by "rear-end" at 17.4% — the classic crash type where one vehicle strikes the back of another — and "braking" at 16.3%, covering both insufficient braking and the phantom braking phenomenon.

"Pedestrian" appears in 9.8% of narratives, a sobering frequency given that pedestrian crashes carry some of the highest injury and fatality rates. "Stationary object" at 9.2% directly relates to the emergency vehicle problem and general failure to detect stopped vehicles in the travel lane.

Six Categories of Failure

We classified narratives into six primary failure categories based on the described scenario:

Lane Keeping Failure (27.4%) is the dominant category. The vehicle drifts out of its lane, swerves unexpectedly, or fails to track lane markings through curves, construction zones, or faded paint. This is the most fundamental function of any ADAS system — staying in the lane — and it fails more than a quarter of the time.

Object Detection Failure (23.4%) means the system failed to see something it should have seen: a stopped car, a pedestrian, a cyclist, a piece of debris. This is the category that includes emergency vehicle crashes, construction zone failures, and pedestrian near-misses.

Inappropriate Braking (19.5%)covers both extremes: braking too hard (phantom braking) and not braking enough (failing to slow for traffic or obstacles). Phantom braking is particularly insidious because it trains drivers to distrust the system's brake warnings — and then one day the braking is real and necessary, but the driver overrides it.

Right-of-Way Violations (12.7%) include running red lights, failing to yield at intersections, and cutting off other vehicles. These are particularly concerning because they represent the system making active decisions that violate traffic law.

Vulnerable Road User incidents (9.8%) involve pedestrians, cyclists, and motorcyclists — the road users most at risk of serious injury. Every one of these represents a potential fatality that the automated system should have prevented.

Environmental Confusion (7.3%) captures cases where weather, lighting, road conditions, or unusual infrastructure confused the system: sun glare, rain, snow, construction, temporary lanes, and unlit roads.

Who Gets the Blame?

In 67.3% of narratives, the automated system is clearly described as the proximate cause of the crash. The vehicle did something wrong — veered out of lane, failed to brake, struck an object it should have detected. In 18.2% of cases, another driver is blamed — they rear-ended the AV, cut it off, or ran a red light. The remaining 14.5%are ambiguous — the narrative doesn't clearly assign responsibility.

This blame distribution is important context for the industry claim that "most AV crashes are caused by other drivers." That may be true for some ADS systems in urban environments, but the aggregate data across all systems shows the opposite: in two-thirds of reported crashes, the automated system bears responsibility.

Reading Between the Lines

The narratives also reveal patterns that don't show up in structured data. Terms like "unresponsive driver" (2.7%) tell a story about driver monitoring failures — situations where the human backup was asleep, distracted, or had checked out. "Construction zone" (5.5%) highlights a known weak spot that requires dynamic understanding of temporary road configurations.

Text analysis can't replace engineering investigation. But when applied to thousands of reports, it reveals the forest that individual crash reports — the trees — can obscure. The message from 4,218 narratives is consistent: lane keeping, object detection, and braking decisions remain unsolved problems for autonomous and driver-assist systems.