Understanding Attitudes toward Autonomous Vehicle in Tort Liability: Randomized Experiments

This pre-registered study employs a between-subject legal vs. non-legal framing design combined with a common vignette-based judgment task.
[Legal Frame (N = 4,200)] Participants in the legal frame complete a 2×2 factorial information experiment followed by a liability judgment task. In the first stage, participants are randomly assigned to watch either (i) a video providing general information about civil litigation procedures or (ii) a video emphasizing the preventive function of tort law. Within each information condition, they are further divided into (a) a group that proceeds directly to the next stage after watching the video and (b) a group that answers brief questions about the preventive function of tort law before moving on. In the second stage, all participants act as mock judges evaluating traffic accident vignettes framed in legal terminology (e.g., comparative negligence, plaintiff, defendant, solatium).
[Non-Legal Frame (N = 2,000)] Participants in the non-legal frame receive no informational treatment — no video and no comprehension questions. They proceed directly to the vignette judgment task. All legal terminology is replaced with everyday language: references to comparative negligence are reframed as responsibility allocation, legal party designations are replaced with non-legal equivalents, and non-pecuniary damages are described using ordinary language. This condition serves as a baseline to identify whether the legal framing itself systematically affects fault attribution and damage award judgments.
[Common Vignette Task] In both frames, participants evaluate traffic accident cases presented as vignettes across 30 distinct scenarios that systematically vary whether the defendant operates a human-driven vehicle or an autonomous vehicle, whether the autonomous vehicle is equipped with state-of-the-art safety technology for accident prevention, whether its prediction-and-avoidance capabilities are superior to or equivalent to those of human drivers, and the plaintiff's baseline level of fault (equal to, lower than, or much lower than the defendant's fault). The 30 scenarios are categorized into five groups; for each participant, one scenario is randomly drawn from each group, and one additional scenario is randomly selected from the full set, so that each participant evaluates a total of six scenarios. For every scenario, participants allocate fault between the two parties, determine an amount for non-pecuniary damages, and report their confidence in each of these judgments.

We implement a 2×2 factorial design with four treatment arms in a legal framing condition (N = 4,200), followed by a common vignette-based judgment task. An additional non-legal framing condition (N = 2,000) was added post-registration (conducted March 10–31, 2026) as a between-subject comparison arm.

Information Content Manipulation (Factor 1)
Participants are randomly assigned to view one of two 30-40 second videos:

- Treatment A (Civil Litigation Information)
Video presents factual information about civil litigation procedures in South Korea, including case categorization by claim amount (small claims under 30 million won, single-judge cases from 30 million to 500 million won, and collegiate panel cases exceeding 500 million won), the three-tier appellate system (district court, high court or collegiate district court panel, and Supreme Court), and the progression of cases through these levels based on appeals and legal interpretation disputes.

- Treatment B (Preventive Function Information)
Video emphasizes tort law's preventive function in reducing accidents and promoting safety. Content explains that tort liability serves not only to compensate victims but also to incentivize accident prevention by establishing behavioral standards, imposing liability for violations that cause harm, and encouraging both individuals and firms to invest in safety measures including safer designs and quality control.

Engagement Manipulation (Factor 2)
Within each information content condition, participants are randomly assigned to:

- No Questions Condition
Participants watch the assigned video and proceed immediately to the vignette scenarios.

- Questions Condition
After watching the video, participants answer two multiple-choice questions testing support for the preventive function of tort law before proceeding to scenarios.

This creates four treatment arms with equal allocation: (1) Civil Litigation Information Only, (2) Civil Litigation Information + Questions, (3) Preventive Function Information Only, (4) Preventive Function Information + Questions.

Vignette-Based Judgment Task (Common to All Arms)

Following information treatment, all participants complete six traffic accident vignettes as mock judges. Each vignette describes a collision between a plaintiff and defendant vehicle (or pedestrian plaintiff), provides demographic information about parties, displays a 30-40 second animated video of the accident, and describes injury severity and economic damages.

We systematically vary the following factors to construct a total of 30 distinct scenarios:
- Plaintiff type: human-driven vehicle, pedestrian
- Defendant vehicle type: human-driven vehicle, fully autonomous vehicle
- Plaintiff–defendant pairing:
.. human-driven vehicle vs. human-driven vehicle
.. pedestrian vs. human-driven vehicle
.. human-driven vehicle vs. autonomous vehicle
.. pedestrian vs. autonomous vehicle
- Potential accident preventability through advanced safety technology on the autonomous defendant vehicle (hereafter ‘preventability’ or ‘preventable’): whether the autonomous vehicle is equipped with state-of-the-art safety technology for nighttime accident prevention (e.g., a latest-generation infrared camera), yielding technology-equipped versus non-equipped autonomous vehicles
- Autonomous vehicle performance (prediction and avoidance capability) (hereafter ‘performance’): when the defendant is an autonomous vehicle, whether its accident prediction and avoidance capabilities are equivalent to, or superior to, those of human drivers
- Accident location: nighttime intersection settings
- Plaintiff’s baseline fault level:
.. Medium: for example, at night, at an intersection without traffic lights, the plaintiff attempts a right turn and collides with the defendant, who is making a left turn into the plaintiff’s lane.
.. Low: for example, at night, at an intersection without traffic lights, the plaintiff is driving straight through the intersection when the defendant attempts a left turn at the intersection and collides with the plaintiff.
.. Very low: for example, at night, at an intersection where traffic lights are installed only on one approach, the defendant enters the intersection by driving straight through a red light and collides with the plaintiff, who is turning right from the cross street that has no traffic light.

Each participant evaluates six scenarios selected through constrained randomization: one scenario with human-driven defendant, two scenarios with preventable autonomous defendant, two scenarios with non-preventable autonomous defendant, and one scenario randomly selected from the full set. This ensures within-subject variation in vehicle type while preventing participant fatigue.

For each scenario, participants make four judgments:

1. Fault allocation percentage: Using a slider from 0-100%, participants allocate responsibility between plaintiff and defendant, where higher percentages indicate greater defendant responsibility
2. Fault allocation confidence: 0-100% scale of certainty in the responsibility allocation judgment
3. Non-pecuniary damages amount: Monetary award for non-pecuniary damages (pain and suffering, emotional distress) in units of 10,000 Korean won, given fixed economic damages of 30 million won
4. Non-pecuniary damages confidence: 0-100% scale of certainty in the damage award

Pre-Treatment and Post-Treatment Measures

Before information treatment, participants complete pre-survey measures including eight questions assessing AI knowledge, AI attitudes, autonomous vehicle familiarity, and autonomous vehicle feature usage. After completing all vignette judgments, participants answer 27 post-treatment questions measuring litigation experience, trust in courts, traffic accident history, driving experience, openness to experience personality traits, risk attitudes, and fairness preferences.

Pilot Study

A pilot study with 300 participants (100 per information content condition) was conducted in November 2025. Pilot results informed refinement of vignette content, confirmation of randomization procedures, and validation of outcome measure distributions. Based on pilot analysis, we reduced the total number of unique scenarios from 45 to 30 by eliminating highway accident scenarios and modifying one high-plaintiff-fault scenario. The main study implements these refined materials with the full 2×2 factorial design including the engagement manipulation.

Non-Legal Framing Condition (Post-Registration Amendment)

A separate group of 2,000 participants completes the same six-vignette judgment task without any informational treatment (no video, no comprehension questions). All legal terminology is replaced with everyday language (e.g., comparative negligence → responsibility allocation, plaintiff/defendant → non-legal equivalents, solatium → ordinary damage language). This condition isolates whether the legal framing itself affects fault attribution and damage awards. The constrained randomization procedure and outcome measures are identical to the legal frame.

2025-12-01

2026-01-31

Defendant Responsibility Allocation

For each of the six vignette scenarios, we measure the percentage of responsibility allocated to the defendant using a continuous slider scale (0-100%), where 0% indicates the defendant bears no responsibility and 100% indicates the defendant bears full responsibility for the accident. This responsilibility ratio allocation determination directly measures whether participants assign different levels of responsibility to autonomous versus human-driven defendants.

Pain and Suffering Damages

For each vignette scenario, we measure the monetary award for non-pecuniary damages that participants determine the defendant must pay to the plaintiff, entered as a continuous amount in units of 10,000 Korean won (approximately $7.50 USD). Scenarios specify that economic damages (medical expenses, lost wages) total 30 million won, and participants determine appropriate additional compensation for pain, suffering, and emotional distress.

These outcomes directly test our core hypotheses about autonomous vehicle aversion in liability judgments. Responsibility allocation measures the fundamental determination in comparative negligence systems of how responsibility divides between parties. If participants systematically assign higher responsibility percentages to autonomous defendants than to human defendants in otherwise identical scenarios, this indicates bias that could create excessive deterrence for autonomous vehicle deployment.

Pain and suffering awards capture an independent dimension of liability that may reflect different psychological processes. Even if responsibility allocation appears neutral, differential damage awards could indicate that participants perceive harms differently when caused by autonomous vehicles, possibly due to violations of anthropomorphic expectations, reduced empathy for non-human defendants, or different intuitions about appropriate deterrence for algorithmic versus human errors.

The interaction between defendant vehicle type and information treatment tests our policy intervention hypothesis. If information emphasizing tort law's preventive function reduces the gap between autonomous and human defendant judgments, this suggests that framing liability as a safety incentive mechanism rather than purely compensatory affects how people evaluate autonomous vehicle responsibility. In addition, we examine interactions between the treatment indicators and key scenario characteristics. This allows us to analyze how bias against autonomous vehicles varies across combinations of treatment conditions and scenario features, such as potential accident preventability and autonomous vehicle performance.

Judgment Confidence Measures

For both responsibility allocation and non-pecuniary damages awards, we measure participants' reported confidence in their judgments on 0-100% continuous scales. These measures assess whether participants experience greater decision uncertainty for autonomous versus human defendant cases, which has implications for jury deliberation dynamics and settlement negotiation patterns even if average judgments appear similar.

Support for the Preventive Function of Tort Law

For participants in the Questions Condition, we measure performance on the two multiple-choice comprehension items testing support for the preventive function of tort law. This manipulation check verifies whether the engagement manipulation successfully increased processing depth and enables instrumental variables analysis treating support question as an endogenous mediator.

Baseline Autonomous Vehicle Attitudes

Pre-treatment measures include self-reported AI knowledge (5-point scale), AI interest level (5-point scale), autonomous vehicle feature usage experience (binary), and autonomous vehicle feature usage frequency (4-point scale). These measures enable analysis of whether pre-existing autonomous vehicle attitudes predict liability judgments and moderate treatment effects.

Scenario-Level Manipulation Checks

For selected scenarios, we measure participants' perceptions of accident preventability and vehicle performance characteristics to verify that manipulated features were perceived as intended. These measures test whether autonomous vehicle defendants described as having superior accident avoidance capabilities are indeed perceived as more technologically advanced than human drivers.

Confidence measures provide insight into decision processes beyond point estimates. Lower confidence for autonomous defendant cases would suggest that current liability frameworks provide insufficient guidance for evaluating algorithmic decisions, even if average fault allocations appear unbiased. This finding would indicate need for doctrinal development rather than simply correcting mean bias.

Attitudinal support questions enable analysis of treatment mechanisms. If the Questions Condition produces stronger effects than information-only exposure, this indicates that active processing is necessary for information to shift liability judgments. If attitudinal support questions mediate treatment effects, this suggests that inducing to think about the preventive role of tort law could enhance intervention effectiveness.

Baseline attitude measures test whether autonomous vehicle bias reflects domain-general technology attitudes or liability-specific concerns. If AI attitudes predict responsibility allocation but do not moderate treatment effects, this suggests that bias stems from stable technology preferences rather than misunderstanding of liability objectives. Conversely, if treatment effects are stronger for individuals with initially negative AI attitudes, this indicates that framing interventions can overcome prior dispositions.

This pre-registered online survey experiment is conducted with 4,200 adults in South Korea aged 20 to 60, recruited from a professional online panel. Sample sizes are allocated across age groups to reflect the underlying population structure. Each participant is individually and randomly assigned to one of four treatment arms in a 2×2 factorial design that combines information content (general information about civil litigation vs. information emphasizing the preventive function of tort law) and the presence or absence of support questions about the preventive role of tort law. Participants first answer pre-treatment questions on AI and autonomous vehicles, then receive their randomly assigned information treatment, and subsequently read standardized explanations of key legal concepts such as fully autonomous vehicles, responsibility allocation, and non-pecuniary damages. They are then presented with six traffic accident vignette scenarios, randomly drawn from a set of 30 scenarios that systematically vary the defendant vehicle type, the presence of advanced safety technology for accident prevention (a latest-generation infrared camera), the accident-prediction and avoidance performance of the autonomous vehicle, and the plaintiff’s baseline fault level (medium/low/very low). For each scenario, participants determine the comparative responsibility allocation and the pain and suffering damages amount and report their confidence in these judgments. Finally, they complete a post-treatment questionnaire measuring demographic characteristics, litigation and driving experience, and psychological traits.

Post-registration amendment (March 2026): A non-legal framing condition was added as a between-subject arm with 2,000 new participants (conducted March 10–31, 2026). In this condition:

All legal terminology is removed (e.g., "comparative negligence" → "responsibility allocation"; "plaintiff/defendant" → everyday equivalents)
No informational video or text about civil litigation procedures is shown
Participants evaluate the same six vignette scenarios as the legal framing conditions
The outcome measures (fault allocation percentages and solatium amounts) remain identical

This addition enables a between-subject comparison of legal vs. non-legal framing to test whether the legal context itself moderates fault attribution toward autonomous vehicle defendants.

We conduct a pre-registered online survey experiment with a total of 6,200 participants total (4,200 in legal framing conditions + 2,000 in non-legal framing condition). The sample consists of adults aged 20 to 60, and is allocated evenly across the four treatment arms while taking into account the demographic composition of the target population. Participants are assigned with equal probability to one of the four treatment groups using a computer-based randomization algorithm embedded in the survey platform.

Sampling and Recruitment

Participants are recruited through the online panel of a professional survey firm and allocated to meet target quotas for each demographic stratum. Eligibility is restricted to adult men and women in their 20s to 60s who are capable of driving.

Survey Flow

The survey proceeds in six phases:

1. Consent and Screening (2 minutes): Participants provide informed consent following institutional review board protocols, confirm age and employment eligibility, and report their primary device for survey completion (mobile phone, tablet, or computer).

2. Pre-Treatment Measures (3 minutes): Participants complete eight questions measuring AI knowledge, AI interest, AI usage experience, AI performance evaluations, autonomous vehicle knowledge, autonomous vehicle interest, autonomous vehicle feature usage experience, and autonomous vehicle feature usage frequency.

3. Information Treatment (1-2 minutes): Participants are randomly assigned to one of four conditions in the 2×2 factorial design. They view their assigned 30-40 second video. In the Questions Condition, they answer two multiple-choice attitudinal questions before proceeding.

4. Concept Explanation (2 minutes): All participants receive standardized definitions of three legal concepts: fully autonomous vehicles (vehicles that recognize all road conditions and drive independently without human intervention), comparative negligence (the legal doctrine allowing courts to adjust damage awards based on plaintiff fault by considering both parties' intent, negligence degree, illegality, and causal contribution to harm occurrence and expansion), and pain and suffering damages (compensation for non-pecuniary damages including emotional distress beyond economic losses such as medical expenses).

5. Vignette Scenarios (15-18 minutes): Participants evaluate six scenarios presented sequentially. Each scenario displays party information (plaintiff and defendant demographics, occupations, vehicle types), an animated video depicting the accident, accident circumstances description, and injury details. After reviewing each scenario, participants make four judgments: responsibility ratio allocation, responsibility allocation confidence, non-pecuniary damages amount, and non-pecuniary damages confidence. Participants cannot return to previous scenarios after submitting judgments.

6. Post-Treatment Measures (8-10 minutes): Participants answer 27 questions measuring demographic characteristics, litigation experience, judicial trust, traffic accident involvement, driving history, openness to experience, risk preferences, and fairness attitudes.

Vignette Assignment Mechanism

Each participant's six scenarios are selected through constrained randomization that ensures: (1) exactly one scenario features a human-driven defendant, (2) exactly two scenarios feature accident-preventable autonomous defendants, (3) exactly two scenarios feature accident-non-preventable autonomous defendants, (4) exactly one scenario is randomly selected from the full set of 30 scenarios, (5) no two scenarios are identical, and (6) both car-to-car and car-to-pedestrian scenarios appear when possible. Within these constraints, scenario order is randomized across participants. This design provides within-subject variation in defendant vehicle type while maintaining statistical power through repeated measures.

Attention and Data Quality Measures

The survey includes two attention check items embedded within the post-treatment questionnaire asking participants to select a specific response option. Participants who fail both attention checks are flagged for sensitivity analysis. Survey completion time is recorded to identify potential non-attentive responding (e.g., completion in under 50% of median time). The survey platform prevents multiple submissions from the same IP address and requires completion on devices with screen resolution sufficient to display vignette videos clearly.

Pilot Study and Design Refinements

A pilot study with 300 participants was conducted in November 2025 using three information conditions (no preventive function emphasis) without the engagement manipulation. Pilot results indicated: (1) median survey completion time of 23 minutes, (2) substantial variance in fault allocation and solatium judgments providing sufficient power for main effects, (3) some evidence of autonomous vehicle bias in fault attribution but not damage awards, (4) no evidence of information treatment effects, suggesting need for the engagement manipulation, (5) scenario comprehension challenges for highway scenarios leading to their exclusion, and (6) ceiling effects in one high-plaintiff-fault scenario leading to its revision.

Based on these findings, we revised the main study to include the engagement manipulation testing whether active processing through comprehension questions is necessary for information to affect judgments. We reduced scenarios from 45 to 30 by eliminating highway cases and modifying one pedestrian scenario. We shortened individual scenario descriptions to reduce participant fatigue. These refinements ensure the main study tests the engagement hypothesis while addressing practical implementation challenges identified in piloting.

Random assignment in the experiment is conducted by a computer program. At the point when participants complete the pre-survey and proceed to the information treatment stage, each participant is independently assigned to one of the four treatment conditions. The platform’s randomization algorithm is designed to maintain approximate balance across the four treatment groups, taking into account key demographic characteristics of the Korean population (such as age, gender, and region). Participants in the non-legal framing condition were recruited as a separate sample and were not randomized into the 2×2 information treatment design. Within the non-legal condition, each participant was assigned six vignette scenarios through the same constrained randomization procedure used in the legal framing conditions, ensuring comparable exposure to different vehicle types.

Individual participant. Each respondent is independently assigned to treatment with no clustering.

No

Not applicable (individual-level randomization without clustering)

6,200 participants (Legal 4,200 + Non-legal 2,000)

* Legal frame (4,200 participants)
Approximately 1,050 participants per treatment arm:
- Arm 1: Civil Litigation Information Only (1,050 participants)
- Arm 2: Civil Litigation Information + Comprehension Questions (1,050 participants)
- Arm 3: Preventive Function Information Only (1,050 participants)
- Arm 4: Preventive Function Information + Comprehension Questions (1,050 participants)

Since each participant evaluates six vignette scenarios, this yields a total of approximately 25,200 scenario-level observations (about 6,300 per treatment arm).

* Non-Legal Frame (2,000 participants)
All participants are assigned to a single condition:

Arm 5: No legal information, no legal terminology — all legal terms replaced with everyday responsibility language (2,000 participants)

Since each participant evaluates six vignette scenarios, this yields a total of approximately 12,000 scenario-level observations.