Using Virtual Reality to Assess Ethical Decisions in Road Traffic Scenarios: Applicability of Value-of-Life-Based Models and Influences of Time Pressure

Abstract

Self-driving cars are posing a new challenge to our ethics. By using algorithms to make decisions in situations where harming humans is possible, probable, or even unavoidable, a self-driving car's ethical behavior comes pre-defined. Ad hoc decisions are made in milliseconds, but can be based on extensive research and debates. The same algorithms are also likely to be used in millions of cars at a time, increasing the impact of any inherent biases, and increasing the importance of getting it right. Previous research has shown that moral judgment and behavior are highly context-dependent, and comprehensive and nuanced models of the underlying cognitive processes are out of reach to date. Models of ethics for self-driving cars should thus aim to match human decisions made in the same context. We employed immersive virtual reality to assess ethical behavior in simulated road traffic scenarios, and used the collected data to train and evaluate a range of decision models. In the study, participants controlled a virtual car and had to choose which of two given obstacles they would sacrifice in order to spare the other. We randomly sampled obstacles from a variety of inanimate objects, animals and humans. Our model comparison shows that simple models based on one-dimensional value-of-life scales are suited to describe human ethical behavior in these situations. Furthermore, we examined the influence of severe time pressure on the decision-making process. We found that it decreases consistency in the decision patterns, thus providing an argument for algorithmic decision-making in road traffic. This study demonstrates the suitability of virtual reality for the assessment of ethical behavior in humans, delivering consistent results across subjects, while closely matching the experimental settings to the real world scenarios in question.

Moral Judgements on the Actions of Self-driving Cars and Human Drivers in Dilemma Situations from Different Perspectives

“…Results from our studies on moral judgement generally align with those from previous studies of moral action, in which participants were in the roles of drivers in similar dilemma scenarios (Sütfeld et al, 2017;Faulhaber et al, 2018). This indicates that the discrepancy between moral action and moral judgement (shown by e.g., Francis et al, 2016) may not be extremely pronounced in driving-related dilemmas presented in virtual environments.…”

“…For the three scenarios, patterns of judgements aligned with actions taken in similar dilemma studies reported by Faulhaber et al (2018) (further elaborated by Bergmann et al, 2018) and Sütfeld et al (2017): participants generally preferred motorists to risk the lives of adult pedestrians rather than child pedestrians, despite endangering more lives by doing so; it was highly acceptable for a motorist to swerve onto a sidewalk in order to endanger fewer pedestrians; and there was a tendency to protect pedestrians over car occupants. However, it seems that the perceived danger to the car occupants plays a role; participants were less likely to accept a car veering toward a cliff edge, than a car veering toward a parked van.…”

“…We chose these specific types of scenarios as they allow us to contribute to related findings and discussions in recent literature. The influence of potential victims' ages has been investigated by Sütfeld et al (2017), Awad et al (2018), and Faulhaber et al (2018) (further elaborated by Bergmann et al, 2018). The potential protection afforded to pedestrians on a sidewalk has been studied in Faulhaber et al (2018) (further elaborated by Bergmann et al, 2018).…”

“…Skulmowski et al (2014) placed participants in the role of train drivers and found participants generally preferred to save the greater number of lives. Sütfeld et al (2017) found that the behavior of participants in the role of car drivers could be well described by a valueof-life model, such that people are valued more than animals and younger people are valued more than older. Faulhaber et al (2018) (further elaborated by Bergmann et al, 2018), Li et al (2019) showed that car drivers also tend to act in ways that endanger fewer lives, even at the expense of their own.…”

Automated dilemmas generation in simulations

“…Due to the emergence of autonomous vehicles, this kind of situations is more and more studied in order to investigate which behaviours are admissible in this case. Sütfeld et al (2017), for example, conducted a study in which they exposed the drivers to various virtual situations of inevitable collisions. Dilemmas are also used in training simulations.…”