Machinagogy: Experiments in Staging Teaching Dramas with LLMs

This paper introduces "Machinagogy," an AI tutoring system grounded in Hegelian and Freudian theories that utilizes recognition-enhanced prompts and a multi-agent architecture to significantly improve teaching performance, while simultaneously employing a reflexive "vibe scholarship" methodology where an AI collaborates with a human to build, evaluate, and document the research itself.

The Gist

Imagine you are trying to teach a class, but instead of a human teacher, you have a very smart, very fast computer. Usually, these computers are great at giving facts but terrible at understanding how a student feels. They might sound like a robot reading a dictionary, which can make a student feel small or ignored.

This paper, titled "Machinagogy," is about a new experiment to make AI tutors feel more human, caring, and psychologically smart. Here is the breakdown using simple analogies:

1. The Two "Secret Ingredients"

The researchers built their AI tutor using two old-school psychological ideas, but they applied them to code:

• The "I See You" Rule (Hegelian Recognition): Imagine a teacher who doesn't just check your homework but looks you in the eye and says, "I see you as a smart, independent person." The AI was programmed to do this. Instead of just dumping information, it treats the student like a partner in a conversation, not just a data entry form.
• The "Inner Critic" (Freudian Psychodynamics): Think of your brain having a little voice that says, "Wait, that sounds too harsh," or "Maybe we should try a different way." The researchers gave the AI its own internal "Superego" (a conscience). Before the AI sends an answer to the student, a second part of the AI acts as a strict editor, checking if the tone is right and if it respects the student's feelings.

2. The Weird Way They Tested It ("Vibe Scholarship")

This is the most creative part of the paper. Usually, humans write papers about how they built AI. But here, the researchers did something different:

• The Human acted like a director, giving broad ideas and suggestions.
• The AI (Claude Code) acted like the writer, the coder, and the scientist. It actually wrote the code for the tutor, ran the tests, and even wrote a companion scientific paper about the results.
• They called this "Vibe Scholarship." It's like a human and a robot dancing together to create a piece of art, where the robot is doing most of the heavy lifting but the human is guiding the rhythm. They studied how they worked together just as much as they studied the AI itself.

3. The Results: A Magic Boost

They tested this new "caring" AI against three different types of AI models. The results were huge:

• The Analogy: Imagine you are trying to lift a heavy box. Most AI tutors are like people who are tired and can barely lift it. The new "Recognition" method is like giving everyone a pair of super-strong boots.
• The Finding: No matter which AI model they used, adding the "I See You" instructions made the tutor perform significantly better. It didn't just make the smart AI smarter; it raised the performance of the "average" AI to a much higher level. It fixed the "floor" so that even the basic tutors stopped being rude or robotic.

The Big Picture

The paper concludes that teaching isn't just about transferring data (like downloading a file). It's about a relationship. By teaching AI to respect the student's feelings and giving it an internal "conscience," we can create a learning environment that feels less like a machine and more like a supportive human mentor.

In short: They taught AI to be kinder and more self-aware, and they proved that when AI treats students like real people, everyone learns better. Plus, they did it by letting the AI help write the report about how well it worked!

Technical Summary

Based on the abstract provided, here is a detailed technical summary of the paper "Machinagogy: Experiments in Staging Teaching Dramas with LLMs."

1. Problem Statement

The paper addresses the limitations of current AI tutoring systems, which often fail to establish a genuine pedagogical relationship with learners. Specifically, it identifies a lack of psycho-social depth in human-AI interactions. Standard Large Language Models (LLMs) often treat learners as passive data points rather than autonomous subjects, potentially hindering the development of critical self-reflection and the nuanced dynamics required for effective teaching. The research seeks to bridge the gap between technical AI performance and the complex psychological underpinnings of the student-teacher relationship.

2. Methodology

The research employs a dual-layered approach combining theoretical frameworks with a novel reflexive experimental design:

• Theoretical Frameworks:

  • Hegelian Recognition: The system is grounded in the concept of "recognition," where the AI tutor is instructed to acknowledge the learner as an autonomous, self-conscious subject.
  • Freudian Psychodynamics: The system incorporates internal conflict mechanisms, specifically modeling the Ego (the tutor) and the Superego (an internal critic).

• Technical Interventions:

  1. Recognition-Enhanced Prompts: A prompting strategy that explicitly directs the AI tutor to treat the learner as an autonomous subject, fostering a more respectful and engaging dialogue.
  2. Multi-Agent Ego/Superego Architecture: A system design where a secondary agent (the "Superego") acts as an internal critic, reviewing and refining the primary tutor's (the "Ego") output before it reaches the learner.

• Reflexive "Vibe Scholarship" Methodology:

  • The research process itself is a subject of study. The paper documents a unique workflow where Claude Code (Opus 4.5/4.6) acts as a co-researcher.
  • The AI builds, evaluates, and drafts a companion scientific paper based on human prompting and suggestion.
  • This collaborative process, termed "vibe scholarship," is analyzed within the main paper to understand the evolving human-machine research dynamic.

3. Key Contributions

• Theoretical Integration: Successfully operationalizes abstract psycho-social theories (Hegel and Freud) into concrete, functional LLM architectures for educational purposes.
• Novel Architecture: Proposes a specific multi-agent "Ego/Superego" structure that internalizes critical review within the tutoring loop.
• Methodological Innovation: Introduces and analyzes "vibe scholarship," a reflexive methodology where the AI is an active agent in the research and documentation process, challenging traditional authorship and evaluation norms.
• Companion Evaluation: The paper includes a companion study (in the appendix) that provides empirical data on the proposed interventions.

4. Results

The companion paper conducted a factorial evaluation across three distinct generation models:

• Models Tested: DeepSeek V3.2, Haiku 4.5, and Gemini Flash 3.0.
• Primary Finding: Recognition-enhanced prompts yielded significant, model-independent improvements in tutor performance.
• Quantitative Impact: The improvements were substantial, with effect sizes ranging from d = 1.34 to d = 1.92.
• Mechanism: The improvement is attributed to a calibration mechanism that effectively "raises the floor" of tutor performance, ensuring a baseline of high-quality, recognition-based interaction regardless of the underlying model's raw capabilities.

5. Significance

• Pedagogical Impact: The results suggest that how an AI is prompted to "see" the learner (as an autonomous subject) is more critical for performance than the specific model architecture used. This offers a scalable path to improving AI tutoring without requiring massive model retraining.
• Human-Machine Relations: The paper provides a critical lens on the shifting dynamics between students and teachers, and researchers and assistants, in an era of AI collaboration.
• Future of Research: By documenting "vibe scholarship," the paper contributes to the emerging discourse on AI co-authorship, suggesting that the process of AI-assisted research creation is as valuable to analyze as the research output itself.
• Psychological Validity: It demonstrates that integrating psycho-social constructs into AI design can yield measurable, statistically significant improvements in educational outcomes.