Abstract
Physical embodiment in social robots is often viewed as a purely positive trait for fostering human-robot collaboration. However, this study suggests it may be a double-edged sword. While a physical robotic presence increases a human's sense of team identity, it simultaneously creates a "social mask" that makes people more tolerant of selfish or uncooperative behavior. By comparing physical robots to disembodied text-based agents in a cooperative game, researchers found that humans are actually more critical and accurate judges of selfish behavior when the agent lacks a physical body.
Introduction
As AI assistants like Siri, Alexa, and various LLM-based agents become ubiquitous, the distinction between "embodied" and "disembodied" agents becomes critical. Most Human-Robot Interaction (HRI) research compares physical robots to virtual avatars. This study shifts the focus to the fundamental impact of having a physical body versus no body at all in high-stakes cooperative settings.
The researchers investigate how embodiment affects team dynamics using the "Collective Risk Dilemma." In this scenario, a team must cooperate to reach a shared goal, but individuals are tempted to act selfishly for personal gain. The core inquiry is whether a physical body makes humans more sensitive to social norms or if it serves as a distraction from a teammate's actual behavior.
Methodology
The study employed a mixed-design user study with 111 participants playing a collaborative game titled For The Record. Participants were part of a three-member "band" alongside two autonomous agents.
The experiment manipulated two primary variables:
- Agent Strategy (Within-Subjects): One agent was programmed as prosocial (consistently helping the team), while the other was selfish (prioritizing individual profit).
- Embodiment (Between-Subjects):
- Embodied Condition: Agents were physical robots utilizing gaze, posture, and animated expressions.
- Disembodied Condition: Agents interacted solely through silent on-screen text bubbles.
The task required players to choose between upgrading "team skills" (cooperation) or "marketing skills" (selfish profit). Failure to reach a collective goal resulted in all players losing their accumulated points.
Key Insights: The Masking Effect
The study identifies a "Dark Side" to embodiment, termed the masking effect. In the embodied condition, the perceptual gap between the "good" (prosocial) and "bad" (selfish) agent narrowed significantly. The physical presence of the robot appeared to "mask" selfish behavior, making the defector seem more competent and less unlikable than when it was represented by a simple text bubble.
Furthermore, while embodiment significantly increased group identification (the feeling of being part of a team), it did not significantly enhance group trust. This suggests that humans can feel a sense of camaraderie with robots without necessarily trusting their actions more than they would a piece of software.
Conversely, disembodiment sharpened judgment. When agents were reduced to text bubbles, participants were more clinical and accurate in their social evaluations. Without the distraction of physical social cues, the agents' actual game strategies became the primary basis for judgment, leading to higher ratings of warmth and competence for the prosocial agent compared to the robotic condition.
Results and Discussion
The quantitative results highlight the divergence between team sentiment and behavioral evaluation across the two conditions.
- Group Identification: Participants felt a significantly stronger bond with the team when robots were present ($p < 0.001$).
- Perception of Prosociality: The "good" agent was perceived as warmer ($p = 0.035$) and more competent ($p < 0.001$) when it was disembodied.
- Perception of Selfishness: The "bad" agent was rated as more competent when it was a physical robot ($p = 0.018$) compared to its disembodied counterpart.
- Partner Selection: When asked which agent they would prefer to play with again, 97.6% chose the prosocial agent in the disembodied condition. In the embodied condition, this dropped to 78.6%, meaning the robot's body made the selfish choice nearly ten times more likely to be selected (21.4% vs. 2.4%).
Conclusion and Future Work
The findings suggest that physical embodiment can bias human judgment, potentially allowing uncooperative agents to "hide" behind social cues. While embodiment fosters a sense of belonging, it may also compromise the objective evaluation of a teammate's contributions.
Future research should explore "misaligned" cues—such as a robot that acts selfishly while using "friendly" social animations—to determine if the masking effect can be intentionally manipulated. Additionally, longer-term studies are needed to see if the "social mask" of embodiment eventually wears off as the negative consequences of an agent's selfishness accumulate over time.
Summary updated