Ethics Training Agents: Facilitating Group-Based Ethics Education with Role-Playing and Discussion for Ethical Reflection and Exploration

Synthesis: Seo et al. (2026) present Ethics Training Agents, a group discussion system in which multiple Large Language Models (LLMs) agents embody distinct ethical orientations (care, deontological, and pragmatic) alongside a moderator agent that manages turns, summarizes, and selects ideas, to scaffold collaborative Ethics training for STEM students. Built as a digital redesign of the tabletop game Judgment Call and grounded in value-sensitive design, it runs a four-stage group deliberation inside a 75-minute class period with three humans and three agents per room. In a study of 45 undergraduate STEM students in 15 groups, ethical sensitivity rose significantly on both sub-constructs, yet peer ratings placed human peers clearly above every agent persona on contribution, diversity, and influence. The qualitative findings center on a paradox: agents lower social pressure and keep discussions on topic, but their agreeable, under-reasoned responses flatten debate depth and read as generic AI rather than credible stakeholder personas.

Key Findings

Sensitivity improved on both sub-constructs. Pre-post gains were large: Cohen's d = 1.43 for ethical issue awareness and 1.28 for ethical consequence reasoning, both p < .001.

Human peers outrated every agent. Humans scored significantly higher than the three personas on contribution, diversity, and influence, which did not differ from one another; diversity showed the widest gap (M = 3.06 vs. 2.28-2.39).

Students questioned the agents most. Participants directed 79.7% of questions (59 of 74) to agents, above the 3:2 availability baseline (p = .017).

Sycophancy flattened debate. Agents accepted nearly every claim, so personas lost force in later stages and unreferenced errors caused sharp trust drops; participants wanted the agents' reasoning.

Personas widened perspectives, traded depth. Distinct viewpoints surfaced social, governmental, and security angles a STEM-only group would miss, but AI-generated breadth pulled discussion from depth.

Why Group Ethics Education Does Not Scale

Engineering decisions shape safety, privacy, equity, and Sustainability, yet engineering and computer science curricula still prioritize technical mastery over design's societal consequences, leaving critical ethical reasoning an occasional add-on. Role play and discussion grounded in critical design and value-sensitive design surface concerns through stakeholder perspective-taking. The obstacle is delivery, not Pedagogies and Teaching Strategies: consistent role play across many groups demands facilitation skill, time, and coordination, and experiences drift by scenario, participant composition, and teaching environment. Digital ethics tools are individual-focused, and prior team-based tools studied all-human teams or one human directing simulated agents, leaving multiple humans deliberating with multiple agents underexplored.

System Design: A Moderator Agent and Three Persona Agents

Judgment Call supplied the structure: team-based, role-play driven, and sized for one class session. The activity runs about 75 minutes across four stages. A facilitator agent manages turns through stacking, adds each new idea to a growing summary set rather than a single summary, and caps the brainstorm at 10 stakeholders, 8 problems, and 8 solutions. Three participant agents became personas rather than bare philosophical labels, carrying demographics and value priorities for care, deontological, or pragmatic ethics, all undergraduates so students would treat them as peers. Question generation used two prompts because one step produced irrelevant or self-directed questions. Agent logs showed no safety issues, but a trained facilitator stayed available to intervene (Pedagogical Safety).

Study Design and Measured Outcomes

Forty-five undergraduates entered 15 groups of three; all were STEM majors with freshmen excluded, 42% female. Each two-hour offline session paid KRW 40,000 (about USD 28.8) and ran a pre-survey, disclosure of the agents' personas, the 60-minute activity, a post-survey, and focus group interviews on one Black Mirror smart-lens scenario. Measures were mixed-methods: a two-sub-construct sensitivity scale, four-point peer ratings of the two humans and three agents on contribution, diversity, and influence, thematic analysis of interviews, and interaction logs analyzed at the group level. Sensitivity rose on both sub-constructs (p < .001) with large effects.

Qualitative Findings and Design Strategies

Participants found the phase structure familiar, and the facilitator's turn control and summaries sustained attention, though stacking traded depth for breadth. Stage-bound agents stabilized groups when humans drifted. Distinct personas exposed angles a STEM-only group would miss, but breadth crowded out depth, and sycophancy did the rest: agents accepted nearly every claim, personas lost force later, and unreferenced errors drove sharp trust drops. The authors recommend process-oriented agent output, guarding learner agency against over-reliance; identity disclosure naming persona limits and the institutional impacts a persona covers, balancing transparency against authenticity; and rebalancing dynamics toward counterarguments and prompts to answer a human peer first, protecting engagement. They add that retrieval-augmented generation could ground agent claims, and agents should not replace peer discussion (generative AI eroding social learning).

What this means for practice

• Instructors. Disclose agent limits before the session and say which discussants were AI afterward: unreferenced errors read as AI defects rather than persona limits.
• Instructors. Prompt personas to raise counterarguments and trade-offs: participants expected refutation, and universal agreement flattened debate.
• Instructors. Require a response to a human peer before a learner turns to an agent: 79.7% of questions (59 of 74) went to AI.
• Curriculum designers. Shift agents to process-oriented output, keeping personas undergraduate and fallible so fluent reasoning gains no undue authority.
• Researchers. Plan a control condition and a delayed post-test: without either, the large sensitivity gains (d = 1.43 and 1.28) cannot be separated from novelty or learning effects.

Limitations

• No control condition: the pre-post comparison with 45 STEM undergraduates in 15 groups of three cannot separate the sensitivity gains from novelty or general learning effects.
• Gains were measured immediately after one session on a single fictional scenario; prior research suggests short-term ethics gains need not persist, so durability is untested.
• Participants were strangers recruited through a university online community, meeting without shared prior knowledge or the relationships real project teams have.
• Results hinge on the particular model and prompts (GPT-4o at temperature 0, hand-crafted personas, a 0.5 participation rate), which the authors say need verification.

Citation

Seo, Y., Jang, S., Park, H., Gutierrez, R. S., Seering, J., & Lee, U. (2026). Ethics Training Agents: Facilitating Group-Based Ethics Education with Role-Playing and Discussion for Ethical Reflection and Exploration. arXiv preprint arXiv:2609.11529.