Debates about the ethics of autonomous vehicles often begin with dramatic crash dilemmas, but the moral decisions that shape everyday road behaviour are usually far less cinematic. A driver may bend a rule to help someone, obey the law but still cause a collision, or act with good intentions and produce a bad outcome. A newly published pair of experiments suggests that people weigh these ingredients differently, and that the way a traffic situation is presented can substantially change the moral judgment that follows.
The research, published on 28 September 2026 in Frontiers in Psychology, examined moral judgment through the Agent, Deed and Consequence framework. The model separates three questions that are often blended together in ordinary moral reasoning: what kind of person is acting, whether the action follows an accepted rule, and whether the outcome is good or bad.
Across two studies, Dario Cecchini, Michael Pflanzer, Sam Cacace and Veljko Dubljević tested these dimensions in realistic traffic scenarios. The first study involved professional ethicists and compared written vignettes with desktop audio-visual versions of the same situations. The second placed university students in an immersive 360-degree visual environment and exposed them to multiple variations of everyday traffic situations.
From philosophical theory to ordinary traffic
Study 1 began with invitations sent to nearly 4,000 professional ethicists. A total of 796 people entered the study, with 386 receiving audio-visual scenarios and 410 receiving text. After exclusions for incomplete responses, low response variability and automated bot detection, the final analytical sample contained 457 participants. Of these, 226 were in the audio-visual condition and 231 in the textual condition.
Participants were randomly assigned to combinations that varied the moral character of the driver, whether the driver followed or violated a traffic rule, and whether the situation ended with a positive or negative consequence. The scenarios deliberately focused on lower-stakes road situations rather than life-or-death trolley problems. Examples included a mother taking a child to school and a person transporting a cat to a veterinary hospital.
The researchers also measured participants’ preferences for virtue ethics, deontology and consequentialism. This allowed them to test whether people who consciously favoured different moral theories would systematically focus on character, rules or outcomes when judging the scenarios.
The driver’s character produced the largest difference
Across the professional-ethicist sample, the portrayal of the agent had a particularly large association with moral acceptability. Positively characterised drivers received an average rating of 5.49, compared with 3.10 for negatively characterised drivers. The 2.39-point difference was statistically significant and corresponded to a large effect size of Cohen’s d = 1.21.
Consequences mattered as well. Scenarios with positive consequences averaged 4.82 compared with 3.77 for negative consequences, a 1.05-point difference with an effect size of d = 0.53. When all presentation formats were combined, however, the overall difference between rule-following and rule-violating deeds was only 0.14 points and was not statistically significant.
That overall result concealed one of the study’s most interesting patterns. Presentation format substantially changed how participants reacted to the moral components. In text, the positive-versus-negative agent difference was 1.06 points. In audio-visual form, it widened to 3.72 points. The direction of the deed effect also changed: textual scenarios produced a 1.37-point difference in the opposite direction, while audio-visual scenarios produced a 1.66-point advantage for rule-following behaviour.
Consequences showed the reverse pattern. In written scenarios, positive outcomes were rated 1.97 points more favourably than negative outcomes. In the audio-visual condition, the difference fell to only 0.13 points and was not statistically significant. The same moral ingredients were therefore not weighted identically when participants read about a situation and when they watched it unfold.
A second experiment tested the pattern in immersive scenes
The second study shifted both the population and the presentation. Researchers recruited 206 undergraduate and graduate students, retaining 146 after applying the study’s completion and response-quality criteria. Participants viewed traffic scenarios in a gallery equipped with eight high-definition projectors that created a 360-degree audio-visual environment.
Instead of seeing only one version of a scenario, each participant assessed 15 vignettes drawn from five traffic situations. These included the two scenarios from Study 1 as well as situations involving a cow crossing the road while a man drove his pregnant wife to hospital, a driver yielding to an ambulance on a narrow mountain road, and a fire marshal managing a riot.
In this immersive experiment, all three components were associated with moral judgment. Positively portrayed drivers received social acceptability ratings averaging 6.58 versus 3.75 for negatively portrayed drivers, a difference of 2.83 points. For participants’ personal moral judgments, the gap was even larger at 3.27 points. Both comparisons had effect sizes of d = 1.16.
Rule compliance mattered, but less strongly. Rule-following deeds were rated 0.90 points higher for social acceptability and 0.88 points higher for personal acceptability, with effect sizes around d = 0.31 to 0.32. Positive consequences produced differences of 0.83 points for social judgment and 0.73 points for personal judgment, with smaller effect sizes of d = 0.29 and d = 0.26 respectively.
The researchers’ sensitivity analysis indicated that the 146-person within-participant design could detect effects of approximately d = 0.23 with 80% statistical power. This matters because the deed and consequence effects were considerably smaller than the effect associated with the driver’s portrayed character.
Why presentation format matters for autonomous-vehicle ethics
The findings complicate any attempt to derive a single moral rule for automated driving from questionnaire responses. A written description leaves room for readers to imagine motives, context and possible justifications. Video constrains some of that interpretation by showing a particular sequence of events, while immersive presentation adds further perceptual detail.
This is especially important for autonomous-vehicle research because ethical preferences are often measured through hypothetical vignettes. If the same underlying scenario produces different moral weighting depending on whether it is described in words or presented visually, designers and regulators should be cautious about treating one survey format as a transparent measure of public moral expectations.
The experiments also found limited evidence that people’s declared preference for virtue ethics, deontology or consequentialism neatly determined how they judged the corresponding character, rule and outcome dimensions. The authors interpret the broader pattern as evidence that participants can converge on several core dimensions of traffic morality even when their explicitly endorsed ethical frameworks differ.
The limits are important
The study measured third-person moral judgments, not actual driving decisions. Participants evaluated what they saw or read rather than controlling a vehicle under pressure. The authors argue that third-person judgments remain relevant because autonomous vehicles will ultimately be assessed by passengers, other road users, regulators and the wider public, but they are not equivalent to behaviour behind the wheel.
The agent manipulation was also deliberately strong. Drivers were portrayed with clearly positive or negative character cues, which may have magnified the apparent importance of character relative to rule compliance and consequences. Study 1 additionally lost a substantial share of its initial respondents through quality exclusions, and Study 2 used students from one university rather than a representative population.
The scenarios were intentionally low-stakes. That improves relevance to routine road use, where autonomous systems will encounter thousands of ordinary interactions for every extreme emergency, but it leaves open whether the same hierarchy of moral considerations would survive in a genuinely life-threatening situation.
The practical lesson is therefore not that autonomous vehicles should simply be programmed to imitate these judgments. Instead, the research shows how sensitive moral evaluation can be to context and presentation. Building ethically acceptable automated systems may require more than asking people which outcome they prefer. It may also require understanding how observers interpret the character of an action, its compliance with rules and its consequences when those elements are experienced in realistic settings.
Source Information
Study Title: Goodness behind a wheel: moral judgment in realistic traffic simulations featuring Agents, Deeds, and Consequences
Authors: Dario Cecchini, Michael Pflanzer, Sam Cacace and Veljko Dubljević
Journal: Frontiers in Psychology
Year: 2026
DOI: 10.3389/fpsyg.2026.1885843
