AI predicts car crashes: Johns Hopkins engineers say system spots collisions before they happen

Show summary Hide summary

• How the system learns to predict crashes
• What the model outputs and why it matters
  • Types of answers users receive
• How this differs from traditional machine learning
• Key risk drivers uncovered by the AI
• Relevance to Maryland roads and rising fatalities
• Adapting the AI to different regions and cultures
  • How customization works in practice
• Practical benefits for policymakers and engineers
• Ongoing research and publication

A team at Johns Hopkins has built an AI that can estimate the change in crash risk when planners tweak a roadway. Change a signal from 20 to 30 seconds, and the system will predict whether crashes rise or fall. The goal is to give engineers and city officials a clearer, data-driven view of traffic safety.

How the system learns to predict crashes

The tool, called SafeTraffic Copilot, uses large language models to read and reason over vast accident records. Researchers fed it detailed descriptions of more than 66,000 collisions. Inputs included scene photos, satellite imagery and numerical measures such as blood alcohol readings.

• Text descriptions from crash reports
• On-site photography and satellite views
• Numeric indicators like speed and intoxication levels
• Contextual details such as weather and road design

By combining these inputs, the AI forms a richer picture than typical black-box algorithms. It interprets interactions among factors, not just single variables.

What the model outputs and why it matters

SafeTraffic Copilot produces scenario-based forecasts. Planners can ask “what if” questions about signals, lanes, or signage. The AI then estimates how many more or fewer crashes would likely occur.

The model also returns confidence scores. These numbers tell decision-makers how much trust to place in each prediction. That transparency addresses a major barrier to using AI in public-safety decisions.

Types of answers users receive

• Estimated change in crash counts for a selected time window
• Risk breakdown by cause, such as alcohol or aggressive driving
• Confidence level that quantifies prediction reliability
• Suggestions for data to collect to raise confidence

How this differs from traditional machine learning

Conventional machine-learning models rely on patterns in past examples. If a new scenario looks too different, those systems often fail to provide a prediction. Generative AI can imagine plausible outcomes for novel combinations of conditions.

That capability lets SafeTraffic Copilot answer customized “what-if” queries. For example, it can simulate the effects of changing traffic-signal timing or redesigning an intersection. This is not just classification. It is causal-style reasoning based on learned context.

Key risk drivers uncovered by the AI

The model’s analyses highlight which factors most strongly increase crash likelihood. In the Johns Hopkins study, two causes stood out.

• Alcohol-impaired driving contributed far more to recorded crashes than many other causes.
• Aggressive driving behaviors were also a major multiplier of crash risk.

Researchers report these causes contributed roughly three times the crash burden compared with other common factors in their dataset.

Relevance to Maryland roads and rising fatalities

The team developed the system partly to help local communities. Maryland has seen a troubling rise in highway deaths over the past decade. Local traffic-safety data show a steady increase in fatalities, with recent yearly totals climbing into the hundreds.

Public officials could use model outputs to prioritize interventions where they matter most. Safety investments could then target the highest-return changes, such as enforcement efforts or signal retiming backed by predictive evidence.

Adapting the AI to different regions and cultures

Because the system understands descriptive text, it can be tuned for other countries and driving cultures. A short paragraph explaining local vehicle mixes or rider habits helps the model adapt.

For example, in many South and Southeast Asian settings, motorcycles are a large share of vehicles. Their crash patterns differ from car-dominated regions. SafeTraffic Copilot can incorporate such differences when provided with localized context.

How customization works in practice

• Provide a written description of local driving behaviors.
• Supply a modest set of region-specific crash reports or photos.
• The model reweights factors and refines confidence scores.
• Outputs then reflect local vehicle mixes and cultural patterns.

Practical benefits for policymakers and engineers

Planners can use the tool to compare options before spending on infrastructure. A single simulation can show trade-offs among safety, traffic flow and cost. That supports faster, evidence-driven choices.

• Evaluate signal timing changes without physical trials
• Choose interventions that cut the most crashes per dollar
• Identify locations where further data collection will reduce uncertainty

Ongoing research and publication

The team published their methods and results in a peer-reviewed journal. Their work explores both the technical performance of generative approaches and their practical utility for road-safety programs.

Researchers emphasize that predictions improve with richer local data. They also stress the need for careful validation when applying the model to high-stakes decisions.