Would you trust an autonomous car to handle an emergency without touching the brakes? For many passengers, the instinct to intervene may be difficult to resist. Yet new research suggests that taking control at the wrong moment can interfere with an automated vehicle's response and potentially increase danger.

A study published in Multimodal Transportation in September 2026 examined how passenger behaviour affects the safety of connected and autonomous vehicles. Researchers developed a framework that evaluates not only how vehicles perform but also how people respond to their decisions.

Testing Autonomous Cars In Emergencies

The researchers used the CARLA driving simulator to test a fleet of three connected and autonomous vehicles across six safety-critical scenarios. These included emergency braking, conflicts involving pedestrians and complicated merging situations.

Thirty participants took part in human-in-the-loop experiments, experiencing cautious, normal and aggressive automated driving modes.

The team assessed three aspects of performance: safety, smoothness and autonomy. They also recorded when participants intervened, examining whether those decisions helped or disrupted the automated driving system.

Why Cautious Driving Scored Higher

Driving style made a substantial difference to the results.

The cautious configuration achieved an overall intelligence score of 90.7, approximately 27.6% higher than the lowest score of 71.1 recorded for aggressive driving.

However, performance also depended on the vehicle's position within the fleet and the algorithm controlling it.

Following vehicles achieved higher autonomy scores because passengers intervened less frequently. Meanwhile, the lead vehicles demonstrated smoother movement thanks to their more sophisticated control strategy.

These differences suggest that evaluating automated driving requires more than measuring collision avoidance or vehicle speed. Passenger reactions and the quality of the journey also matter.

When Human Intervention Backfires

Across 27 valid experiments, researchers recorded 115 passenger interventions. Collisions occurred in 33.3% of those experiments.

Surprisingly, many interventions were unnecessary or counterproductive. Participants sometimes braked too early or applied insufficient braking force, disrupting the automated system's intended response.

This highlights a difficult aspect of shared control. A passenger may recognise a developing hazard but misunderstand how the vehicle plans to respond.

Intervening prematurely can therefore prevent an automated system from completing a manoeuvre that might otherwise have been effective.

Understanding Passenger Trust

Traditional assessments of autonomous vehicles largely concentrate on technical performance. However, even a technically capable system may struggle to gain acceptance if passengers find its behaviour uncomfortable or unpredictable.

The researchers argue that passenger perceptions and intervention behaviour should form part of vehicle safety assessments.

Their proposed framework combines these human factors with measurable driving performance, offering developers a more comprehensive way to evaluate autonomous systems.

What This Means For Robotaxis

The findings could inform how manufacturers design automated driving modes and communicate vehicle decisions to passengers.

However, the experiments involved only 30 participants in a simulated environment. The results cannot establish how frequently unnecessary interventions occur in everyday traffic or whether the same patterns would emerge during real emergencies.

The research nevertheless identifies an important challenge for autonomous mobility: a vehicle must perform reliably while helping passengers understand its behaviour.

Future safety assessments may therefore need to consider not only whether an automated car can handle a dangerous situation, but also whether the people inside are prepared to let it do so.