Ethical Asymmetry in Human-Robot Interaction - An Empirical Test of Sparrow's Hypothesis

The ethics of human-robot interaction (HRI) have been discussed extensively based on three traditional frameworks: deontology, consequentialism, and virtue ethics. We conducted a mixed within/between experiment to investigate Sparrow’s proposed ethical asymmetry hypothesis in human treatment of robots. The moral permissibility of action (MPA) was manipulated as a subject grouping variable, and virtue type (prudence, justice, courage, and temperance) was controlled as a within-subjects factor. We tested moral stimuli using an online questionnaire with Perceived Moral Permissibility of Action (PMPA) and Perceived Virtue Scores (PVS) as response measures. The PVS measure was based on an adaptation of the established Questionnaire on Cardinal Virtues (QCV), while the PMPA was based on Malle et al. [39] work. We found that the MPA significantly influenced the PMPA and perceived virtue scores. The best-fitting model to describe the relationship between PMPA and PVS was cubic, which is symmetrical in nature. Our study did not confirm Sparrow’s asymmetry hypothesis. The adaptation of the QCV is expected to have utility for future studies, pending additional psychometric property assessments.

💡 Research Summary

This paper presents the first empirical test of the “ethical asymmetry” hypothesis originally proposed by Sparrow (2022), which claims that people condemn negative actions toward robots (e.g., kicking, destroying) far more strongly than they praise positive actions of comparable moral weight (e.g., petting, helping). The authors adopt a virtue‑ethics framework, focusing on the four cardinal virtues—prudence, justice, courage, and temperance—and develop two novel measurement instruments tailored for human‑robot interaction (HRI).

First, they adapt the Questionnaire on Cardinal Virtues (QCV) to create a Perceived Virtue Scores (PVS) scale. Each virtue is measured with six Likert‑type items (1 = strongly disagree, 10 = strongly agree), yielding highly reliable subscales (Cronbach’s α = 0.94–0.97; McDonald’s ω ≈ same). Second, they modify Malle et al.’s moral permissibility instrument into a three‑item Perceived Moral Permissibility of Action (PMPA) scale, also on a 10‑point Likert scale.

The experimental design is a mixed within‑/between‑subjects study. The between‑subjects factor is Moral Permissibility of Action (MPA), operationalized as ten discrete levels (scores 1–10). Participants are randomly assigned to one of these ten MPA conditions. Within‑subjects, each participant evaluates four vignette sets, each designed to target one of the cardinal virtues. In total, 40 textual vignettes (19 adapted from prior human‑human corpora, 21 newly authored) describe robot‑related scenarios ranging from highly permissible (e.g., assisting a robot) to highly impermissible (e.g., sabotaging a robot). After reading each vignette, participants first rate the relevant virtue using the PVS items, then rate the same vignette on the PMP A scale.

A total of 146 adult native‑English speakers from the United States completed the online questionnaire (average age ≈ 39 years; gender balanced). Power analysis (G*Power) indicated a minimum sample of 124; the final sample exceeds this threshold, providing adequate power (0.80) to detect small effect sizes (ΔR² ≈ 0.07).

Statistical analysis proceeds in three stages. (1) ANOVA confirms that the manipulated MPA level significantly influences both PMP A scores (F > 12, p < .001) and PVS scores (same significance), indicating that participants perceive higher moral permissibility as associated with higher virtue attribution. (2) Reliability analyses show excellent internal consistency for all four virtue subscales. (3) To examine the shape of the relationship between PMP A (independent) and PVS (dependent), the authors fit linear, quadratic, and cubic regression models. Model comparison using adjusted R² and Akaike Information Criterion (AIC) identifies the cubic (third‑order) model as best‑fitting (adjusted R² ≈ 0.68, lowest AIC). Crucially, the cubic curve is symmetric around its inflection point, meaning that increases in perceived moral permissibility produce comparable increases in virtue scores for both positive and negative actions. This symmetry directly contradicts Sparrow’s asymmetry prediction, which would manifest as a steeper slope for negative actions (condemnation) than for positive actions (praise).

The authors therefore conclude that, at least within the constraints of vignette‑based, self‑report methodology, there is no empirical support for ethical asymmetry in HRI. Instead, participants appear to apply a balanced virtue‑based evaluation to both harmful and benevolent robot‑related behaviors.

Beyond hypothesis testing, the study contributes methodological tools to the HRI community. The adapted QCV‑based PVS scale offers a concise, high‑reliability instrument for assessing perceived virtues of others (in this case, human actors interacting with robots). Its brevity (24 items total) makes it feasible for inclusion in larger experimental batteries. The PMP A adaptation likewise provides a scalable measure of perceived moral permissibility that can be applied across diverse HRI contexts.

The paper acknowledges several limitations. The sample is limited to U.S. English speakers, restricting cross‑cultural generalizability. Vignette scenarios, while carefully calibrated, cannot fully capture the embodied, affective intensity of real‑world robot interactions; thus, the observed symmetry might differ in live settings where physical harm or affection elicits stronger emotional responses. Self‑report measures are susceptible to social desirability bias, and the binary‑to‑Likert conversion for moral permissibility may alter the underlying construct. Moreover, the operationalization of “equal moral weight” between positive and negative actions remains somewhat subjective; future work could develop objective metrics (e.g., quantified damage, resource cost) to standardize this dimension.

Future research directions proposed include (a) cross‑cultural replications to test whether negativity bias or cultural norms modulate asymmetry; (b) live‑robot experiments that record physiological indices (skin conductance, facial EMG, EEG) alongside self‑reports; (c) refinement of the “moral weight” parameter to enable more precise matching of positive and negative scenarios; and (d) longitudinal studies tracking how virtue attributions evolve with repeated robot exposure.

In sum, this study provides a rigorous, mixed‑design empirical assessment of Sparrow’s ethical asymmetry hypothesis, finds no supporting evidence, and delivers validated measurement tools that can advance systematic inquiry into moral cognition within human‑robot interaction.