Introduction
As robotic systems increasingly integrate into domestic and healthcare environments, they encounter complex social scenarios where the optimal course of action is ethically ambiguous. A primary concern in human-robot interaction (HRI) is the use of "white lies" or robotic deception. This study, originally presented at TRAIT-HRI ’22, investigates the conditions under which deception might be considered acceptable by comparing the perspectives of the general public with those of ethics experts applying formal frameworks.
Methodology
The researchers employed a comparative survey design involving two participant groups:
- General Public ("Folk Morality"): 100 adults recruited via Amazon Mechanical Turk.
- Ethics Experts: 22 specialists tasked with applying five formal frameworks: Utilitarianism, Kantian Ethics, Ethics of Care, Virtue Ethics, and Social Justice Theory.
Participants evaluated robotic behavior across two distinct scenarios:
- High-Risk: A robot assisting an elderly individual with medication management (pill-sorting), where errors carry potentially fatal consequences.
- Low-Risk: A robot playing a game of Connect Four with a child.
The study further differentiated between two levels of user performance: Case 1 involved minor deficiencies (a single mistake), while Case 2 involved significant deficiencies (failing approximately 50% of the time). Statistical significance was determined using McNemar tests to compare the acceptability of deception between the high-risk and low-risk scenarios.
Results
Case 1: Minor Performance Deficiencies
In scenarios where the user made only minor mistakes, most ethical frameworks did not find a statistically significant difference in the acceptability of deception between the medical task and the game. The notable exception was the Ethics of Care framework, which suggested that deceiving a child in a game is more acceptable than deceiving an adult in a medical context, even when the risk is low.
Case 2: Significant Performance Deficiencies
When user performance dropped significantly, a strong convergence emerged across nearly all frameworks. In these instances, deception was viewed as significantly more acceptable in the low-risk game than in the high-risk medical task. The potential for "fatal consequences" in the pill-sorting task consistently outweighed the psychological benefits of encouragement through deception.
Social Justice Theory remained the outlier, showing no significant difference in either case (p = 0.25 and p = 0.125), suggesting this framework may be less sensitive to the specific risk-reward trade-offs inherent in these HRI scenarios.
Key Insights
- Risk Dominance: High-risk environments impose a strict requirement for honesty. The study indicates that as the severity of potential negative outcomes increases, the tolerance for "white lies" decreases, regardless of the user's emotional state.
- Performance Sensitivity: A robot's ethical decision-making must be dynamic. The findings suggest that ethical requirements for honesty intensify as a user's performance declines in critical tasks.
- Framework Convergence: The alignment of folk morality with most formal ethical frameworks in high-stakes, poor-performance scenarios provides a robust foundation for developing robotic "ethical architectures."
Conclusion and Future Work
The study highlights the necessity for robots to possess a context-aware ethical architecture. Such a system would allow a robot to transition between different ethical frameworks based on the task's risk level and the user's current state. Future research will focus on incorporating these findings into functional robotic models and exploring the influence of demographic factors on ethical preferences in HRI.
Summary updated