Summary
Mollick and Mollick propose seven ways to assign AI in educational settings, each grounded in learning science. An AI-tutor provides personalized direct instruction (Bloom’s two-sigma problem). An AI-coach prompts metacognition — reflection, self-assessment, pre-mortems — without handing out answers. An AI-mentor gives balanced, ongoing feedback on student work. An AI-simulator creates realistic practice scenarios. An AI-teammate provides alternate viewpoints. An AI-student lets learners teach (a powerful learning technique). An AI-tool helps accomplish tasks faster. Each role carries specific risks: confabulation is highest for Tutor (factual instruction), personality mismatch for Coach (tone may not suit the student), outsourced thinking for Tool.
The paper provides complete prompts for each role, annotated with design rationale — where the prompt establishes role, goal, step-by-step instructions, personalization, and constraints. The authors emphasize that students must remain “human in the loop,” critically assessing AI output rather than passively accepting it.
What it means for our work
This paper is the pedagogical foundation for Persona. Today, each langlearn-tts prompt manually combines three Mollick roles — Tutor, Coach, and Simulator — into a named character (Profesor Garcia teaches Spanish through communicative exercises, gives pronunciation feedback, and creates immersive dialogue scenarios). The MCP server acts as the Tool layer, generating audio the tutor weaves into lessons.
Persona extracts the character definition into a standalone building block so any domain tool can adopt it. Z Spec gets a formal methods mentor who teaches the Z notation through guided specification exercises. PR/FAQ gets a product strategist who coaches the Working Backwards process. Use Cases gets a requirements analyst who simulates stakeholder interviews. The tool doesn’t just do the work — it teaches the underlying discipline, adapting to the learner’s frontier (modeled after ALEKS and Knowledge Space Theory).
The seven-role taxonomy also shapes how we think about the /tool:tutor command that every grounding tool implements. A tutor session isn’t a single role — it’s a composition: direct instruction (Tutor) when introducing a concept, reflection prompts (Coach) when reviewing work, simulated scenarios (Simulator) when practicing application.