Automated vehicles are designed to reduce human error, yet their safety in dangerous situations depends on more than software performance alone.

A 2026 study published in Accident Analysis & Prevention found that outcomes during safety-critical events are shaped by both the capability of the automated system and the way drivers decide when to intervene.

Testing Several Automation Levels

Researchers from Tsinghua University and collaborating institutions created a unified driver-in-the-loop framework to examine human–automation interaction across SAE Levels 2, 3 and 4. Using high-fidelity driving simulators, they exposed drivers to representative safety-critical scenarios and recorded intervention behaviour, timing and collision outcomes.

The study analysed 2,582 experimental trials. To separate the influence of automation from human behaviour, the researchers compared three elements: a human-driven reference, the hazard-mitigation capability of the automated system operating on its own, and the safety actually achieved when drivers and automation interacted.

This distinction matters because a technically capable automated system does not necessarily produce the same result once a human begins supervising, hesitating or taking control.

Higher Automation Changed Driver Behaviour

Overall collision rates differed across levels. The human-driven Level 0 reference had the lowest rate at 13.73%. Under human–automation interaction, the collision rate was 25.19% at Level 2, declined to 19.94% at Level 3 and fell further to 15.37% at Level 4.

As automation capability increased, drivers intervened less often and generally did so later. Overall safety performance also improved at the higher automation levels.

That does not mean intervention itself became safer, however. Among cases in which drivers did intervene, the probability of collision remained at a relatively stable non-zero level of about 26% across the automation levels studied.

Timing May Matter More

The researchers found that two factors were especially important: the driver's risk state when intervention began and the delay before responding. Model-based validation showed that collision risk was primarily associated with those conditions rather than simply with the nominal automation level.

The intervention-onset risk state and the so-called controllability boundary also showed broadly similar patterns across Levels 2–4. In practical terms, this suggests that increasing automation mainly changed whether drivers stepped in and how late they did so, while the basic relationship between intervention timing and the remaining opportunity to avoid a collision stayed relatively consistent.

The researchers' analysis indicates that takeover performance should therefore not be judged only by whether a driver resumed control. Intervention can occur after a hazardous situation has already progressed to a point where the remaining opportunity for successful mitigation is limited.

Why This Matters For Design

These findings could influence how automated-driving systems are evaluated. Safety testing often focuses heavily on what automation can achieve independently, but realised performance can also depend on supervision, delayed reactions and the transition of control between machine and driver.

A system that recognises when the remaining window for effective human intervention is narrowing could potentially support earlier warnings or better-designed transitions of control. The authors say their findings provide a driver-centred basis for understanding takeover limits and improving human-centred safety evaluation.

Because the evidence comes from controlled high-fidelity simulator experiments rather than everyday road use, the approximately 26% collision risk following intervention should not be interpreted as a universal probability for automated vehicles.

Instead, the research highlights a broader principle: safer automation is not simply a matter of building a more capable driving system. Designers also need to understand the moment when human judgement re-enters the loop — and whether enough time remains for that intervention to make a meaningful difference.