An Experimental Ethics Approach to Robot Ethics Education

Thomas Emrys Williams, Qin Zhu, Daniel H. GrollmanAAAI 2020 · DOI

Introduction

Computer Science education frequently prioritizes technical proficiency over ethical considerations. When ethics is addressed, it is often through passive lectures or standalone philosophy courses that feel disconnected from engineering practice. This paper explores a "learning by doing" approach to robot ethics, where undergraduate students act as researchers in a real-world Human-Robot Interaction (HRI) experiment. By integrating rigorous social science methods into a technical curriculum, the authors aim to help students internalize the ethical implications of the systems they build.

Methodology

The study employed a "nested" experimental design within a Robot Ethics course at the Colorado School of Mines. The curriculum followed a four-phase process: formal ethics certification (CITI training), classroom lectures on experimental moral psychology, hands-on lab training, and active research participation.

The HRI Experiment

Students conducted a study investigating whether the choice of "wakewords" (e.g., "Hey Pepper" vs. "Excuse me, Pepper") influenced how politely human participants treated both the robot and a human bystander. The experiment utilized a SoftBank Pepper robot as the interaction platform.

The SoftBank Pepper robot used as the primary interaction platform in the student-led HRI experiment.

Educational Assessment

The instructors studied the students themselves, who were randomly assigned to one of three roles:

  • Experimenters: Managed participants and handled debriefing.
  • Confederates: Acted within the experimental scenarios.
  • Control Group: Attended lectures only.

The cohort consisted of 32 undergraduate students. Evaluation was conducted via an ungraded "pop quiz" measuring recall of key concepts, including research hypotheses, metrics, experimental design, and consent procedures.

Results and Key Insights

The most significant finding was that the specific role a student played significantly impacted their learning outcomes. Students in the "Experimenter" role, who were required to repeatedly explain the study's goals and ethics to participants, demonstrated a measurably better grasp of the research hypotheses.

Comparison of research hypotheses recalled across the three experimental roles, showing higher recall for the Experimenter group.

Bayesian Analysis of Variance provided moderate evidence (Bayes Factor = 6.606) that Experimenters recalled more key hypotheses (mean = 2.56) than Confederates (mean = 1.7). However, for basic facts like general metrics or consent procedures, there was weak evidence against the idea that active participation outperformed traditional lectures. This suggests that while lectures are sufficient for factual knowledge, active roles are superior for deep conceptual understanding.

Additionally, graduate students serving as supervisors benefited from an "apprenticeship" model, gaining experience in data analysis and scientific co-authorship.

Discussion and Limitations

While the "learning by doing" model shows promise, it faces significant logistical hurdles. Coordinating dozens of students to run multiple research sessions created a substantial administrative burden. Furthermore, the small sample size (n=32) limited the statistical power to find quantifiable benefits across all learning objectives.

Future iterations might focus on "pilot-oriented" research modules. By having students design and pilot experiments without full-scale data collection, educators can provide the benefits of experimental design and ethics training without the logistical complexity of managing hundreds of human participants.

Summary updated

BibTeX
@inproceedings{williams2020experimental,
  author = {Thomas Emrys Williams and Qin Zhu and Daniel H. Grollman},
  title = {{An Experimental Ethics Approach to Robot Ethics Education}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i09.7067},
}
Figure