Automated driving promises safer roads, yet advanced technology cannot eliminate every dangerous situation.

A 2026 study published in Accident Analysis & Prevention reveals that vehicle safety depends not only on automation capabilities but also on when drivers intervene and the level of risk they face when taking control.

Testing Different Automation Levels

Researchers led by Detong Qin at Tsinghua University investigated how drivers interact with automated vehicles during safety-critical situations. Using high-fidelity driving simulators, they examined 2,582 experimental trials involving different levels of automation and potentially dangerous traffic scenarios.

The experiments covered SAE Levels 2, 3 and 4, representing progressively greater automated-driving capabilities. The researchers compared three aspects of safety: human driving performance, the automated system's independent ability to avoid hazards and the outcomes produced when humans interacted with automation.

This approach allowed them to examine whether increasingly sophisticated technology genuinely improves safety when driver behaviour is taken into account.

How Automation Changes Safety

Collision rates differed substantially across the experimental conditions. Conventional human driving produced a collision rate of 13.73%, compared with 25.19% at Level 2, 19.94% at Level 3 and 15.37% at Level 4.

Although the human-driven reference recorded the lowest collision rate, safety performance improved progressively across the three automated-driving levels examined.

Higher automation also changed driver behaviour. Participants intervened less frequently as the systems became more capable, and their interventions generally occurred later.

However, the researchers discovered an important pattern: among situations involving driver intervention, collision risk remained relatively consistent at approximately 26%, regardless of the automation level.

Why Intervention Timing Matters

The findings suggest that the circumstances in which drivers regain control may be more important than the sophistication of the automated system alone.

Researchers examined the risk present when intervention began, the driver's response delay and the remaining opportunity to prevent a collision.

Their modelling indicated that collision risk was primarily associated with the risk state at the beginning of intervention and the time drivers took to respond.

They also identified similar patterns in intervention effectiveness across automation levels. Although more advanced systems changed how frequently and how late drivers intervened, they did not eliminate the underlying risks associated with taking control during an emergency.

An intervention that begins after a hazardous situation has developed may leave insufficient time for effective corrective action, even when the driver responds appropriately.

Designing Safer Automated Vehicles

The research highlights why evaluating automated vehicles exclusively through their technical capabilities provides an incomplete picture of safety.

Developers must also consider how drivers perceive hazards, recognise the need for intervention and respond when automation reaches its operational limits.

Better-designed warning systems and transitions of control could potentially help drivers intervene while enough time remains to avoid a collision.

The findings also support evaluating automated-driving systems according to the safety achieved during actual human–automation interaction, rather than relying solely on how effectively the technology performs independently.

Important Research Limitations

The experiments used driving simulators rather than everyday road conditions. Consequently, the observed collision percentages should not be interpreted as real-world crash probabilities.

The approximately 26% collision risk was specific to experimental situations involving driver intervention, not every journey made in an automated vehicle.

Nevertheless, the study identifies an important consideration for future vehicle development: increasing automation capabilities does not automatically resolve the challenges of human intervention. Understanding when drivers take control, and whether sufficient time remains to respond, could help engineers develop safer systems that account for both human behaviour and technological performance.