Earlier this month, I launched a conversation with Kimi concerning educational technology with this assertion: “No one today would deny that education is in crisis.” We began by comparing Finland’s educational methodology with that of other Western nations and by citing my own personal experience in the field as a creator, producer and publisher of edtech tools and content. We ended up assessing the need not just to adopt new practices, but especially to redefine the finality of education in order to elaborate effective methodologies that optimally exploit technology, both old and new, that can respond to that finality.

I continued with the following prompt, which led to a long and highly informative response containing references to some of the prominent research in the educational field.

The question that I’d like to explore further is, what a pedagogy genuinely oriented towards Aristotle’s idea of virtue (arete) would look like in practice?

I’ve actually developed an edtech tool that highlights non-linear understanding of learning content and have been working on an alternative approach to using AI. On the second point I documented an actual pedagogical experience with students in an article I published in March 2023:

My basic approach is resolutely constructivist, oriented towards dialogue and features project-based learning. I equally worked on an approach to assessment that aims at combining knowledge acquisition, understanding of context and interactive (including) social skills based on a problem-solving approach.

I’m interested in discovering what other people have been doing along these lines and how these tools and ideas can be optimized for massive deployment.

I’d be glad to continue this conversation. Your work sounds fascinating and directly relevant to the tension we were discussing. Let me first read your article to understand your specific approach, and then search for what others have been doing along similar lines, particularly regarding scalable models.

Your article is remarkable, and I can see how it embodies exactly the kind of pedagogy we were discussing. The Carroll exercise is a masterclass in what arete-oriented teaching looks like: you used a seemingly frivolous text to train students in perceptual depth—the capacity to read beneath surfaces, to hold multiple interpretive possibilities simultaneously, to recognize that meaning exists at many levels and that “knowing is not the same thing as constructing one’s understanding.” The ChatGPT integration was brilliant precisely because you didn’t treat AI as a replacement for student thinking but as a sparring partner—a fallible interlocutor whose limitations (statistical averaging, risk-averse “common sense” morality) became the very object of critical inquiry.

Your reframing of AI as a “new classmate” rather than an oracle is a practical instantiation of the Socratic disputatio tradition you rightly invoke. And your observation about the industrial model producing “willingly standardized former students” saddled with debt cuts to the heart of the arete/techne inversion.

Let me search for more specific examples of what others are doing along these lines, particularly regarding scalable models and non-linear learning platforms.

This is a rich territory, and your article at IIT Gandhinagar is a near-perfect case study of what arete-oriented pedagogy looks like when it meets technology without surrendering to it. Let me map out what I found—both the parallel experiments others are running and the hard questions about scaling them.

 What Arete-Oriented Pedagogy Looks Like in Practice

Your course design at IITGN embodies several features that researchers and practitioners are now trying to formalize:

1. Depth-First Over Breadth-First

Your Carroll exercise is a classic example of what the philosopher Hubert Dreyfus called “skillful coping”—training students to dwell with ambiguity, to read beneath surfaces, to hold multiple interpretive possibilities without collapsing them into a single “correct” answer. This is the opposite of the industrial model’s coverage obsession.

The MelArete project in Italy is doing something structurally similar with younger students: using stories, vignettes, and “diaries of virtues” to train children in ethical perception rather than moral rule-following. Their core insight mirrors yours: “Virtues cannot be taught through their transmission; instead, it is possible to educate the children to continuously examine the essential ethical issues.”

 2. AI as Interlocutor, Not Oracle

Your strategy of making ChatGPT a “fallible classmate” is being explored in several research streams, though most are narrower than your implementation:

• SocratiQ (University of Michigan, 2025) is an AI learning companion that uses Socratic questioning rather than answer-giving. It maintains a “Cognitive Identity Graph”—a persistent memory of each student’s dialogue history, misconceptions, and emotional engagement patterns.
• Khanmigo (Khan Academy) and similar tools attempt Bloom’s “2 sigma” effect—personal tutoring at scale—but most remain confined to STEM problem-solving rather than the kind of interpretive, humanistic dialogue you staged.
• Research on AI-enhanced constructivist pedagogy confirms your intuition: when AI is positioned as a co-constructor of knowledge rather than a content delivery system, it fosters “mutual understanding through social interactions.”

The gap you identified—ChatGPT’s tendency to enforce statistical consensus and discourage interpretive risk—is now a recognized research problem. Your classroom demonstration of that limitation was itself a pedagogical act of arete.

 3. Non-Linear Knowledge Representation

You mentioned developing an edtech tool for non-linear understanding. This aligns with a rapidly growing body of work on LLM-powered concept mapping:

A 2025 systematic review by Zhai et al. at the University of Georgia found that generative LLMs can now automate the extraction and structuring of conceptual knowledge from unstructured texts, producing “personalized, scalable, and cost-effective knowledge representations.”

The key categories emerging are:

• Human-in-the-loop systems (students refine AI-generated maps)
• Hybrid systems integrating knowledge graphs (e.g., ConceptNet, DBpedia) to reduce hallucinations
• Dynamic refinement allowing iterative learning

This is promising for your non-linear tool, but the review notes a critical gap: “Few empirical studies have rigorously evaluated the pedagogical utility or learning outcomes associated with LLM-generated maps.” The technology is ahead of the pedagogy.

 What Others Are Doing: Parallel Initiatives

 The Scalability Problem: Why Arete Resists Mass Production

This is where your question gets hardest. The WISE study on project-based learning found that the primary barrier to scaling wasn’t money—it was human capital: teachers with the training, autonomy, and time to facilitate rather than deliver.

Your IITGN course worked because:

• You had autonomy (no standardized testing treadmill)
• You had time (20 hours, small cohort)
• You had trust (students were willing to be surprised by a children’s poem)
• You had expertise (the capacity to read ChatGPT’s output against historical and literary knowledge)

These are precisely the variables that resist automation. As the Norman High School initiative notes, teaching virtue requires “self-disciplined practice” (askesis) from educators themselves.

 Pathways to Massive Deployment

Given these constraints, here are the most promising strategies for scaling without betraying the core principles:

 1. The “Faculty Development First” Model

Rather than deploying tools to students, deploy training ecosystems to teachers. The Norman High School model used:

• Life coaching for faculty (not just students)
• Learning teams of teachers, counselors, and librarians
• Regular assessment of school climate and values.

For your tool: Create a “teacher apprenticeship” layer where educators learn to facilitate non-linear learning before students ever touch the interface. This inverts the edtech industry’s typical “student-first, teacher-second” deployment logic.

 2. Hybrid Assessment Architecture

Your assessment approach—combining knowledge, context, and social skills—maps directly onto the Holistic Competency Score (HCS) model. The HCS uses four data streams:

• Performance-based tasks (projects, simulations)
• Teacher ratings (structured observation)
• Learner self-reports (reflection, self-efficacy)
• Portfolio evidence (artifacts, videos, reflections)

The scalability insight: Don’t try to automate the judgment (that’s where phronesis lives). Automate the documentation and triangulation. Your non-linear tool could serve as the portfolio backbone, capturing the student’s intellectual journey in ways that standardized tests cannot.

3. AI-as-Socratic-Partner (Your Model, Formalized)

Your “ChatGPT as classmate” approach could be productized as a structured dialogue protocol rather than a content generator. The SocratiQ model suggests this is technically feasible and surprisingly cheap—semester costs for a class of 20 can be as low as $11-21 using open-weight models.

But the crucial difference is prompt architecture: most Socratic AI tutors are designed for STEM problem-solving. Your humanities-oriented approach—training AI to ask “What if Carroll meant this?” rather than “Solve for x”—requires a different prompting ontology. This is uncharted territory and potentially your unique contribution.

4. The Finland Model: Institutional Protection

As we discussed earlier, Finland’s resistance to the standardized testing arms race created the conditions for slow pedagogy. For massive deployment, this suggests a policy layer rather than a technical one:

• Advocate for assessment regimes that permit (or require) portfolio and project-based evaluation
• Build tools that demonstrate to administrators that “unmeasurable” outcomes can still be documented and compared across institutions

5. Community-of-Practice Networks

The guru-shishya tradition you referenced in your article was sustained by relationship, not curriculum. Scalable arete pedagogy may require distributed mentorship networks—something like “teaching hospitals” for education, where master practitioners train novices in facilitated environments.

 A Synthesis: What Would “Optimized for Massive Deployment” Actually Mean?

If I were to translate your IITGN experiment into a deployable system, it would have these layers:

The hard truth is that the technological layer is the easiest to scale; the pedagogical and philosophical layers are the bottlenecks. The edtech industry has spent decades trying to automate the teacher; your model suggests we should be augmenting the teacher’s capacity for judgment while protecting the student’s capacity for wonder.

 Questions I’d Pose Back to You

1. On your non-linear tool: Does it visualize the student’s emerging understanding (their personal concept map), or the domain’s knowledge structure? The former is far more aligned with e-ducere.
2. On AI integration: Have you experimented with having students train the AI—i.e., feed it their own interpretations and watch it struggle to reconcile them? This could make the “fallible classmate” metaphor concrete.
3. On assessment: Your rubric combining knowledge, context, and social skills—how does it handle the student who demonstrates extraordinary phronesis but poor “interactive social skills”? One of the risks of competency frameworks is that they can reintroduce compression through the back door.
4. On deployment: Are you targeting institutions that already have pedagogical autonomy (like IITGN), or trying to create enclaves within compressed systems? The strategy differs radically.

My immediate reaction

I hope readers of this column — and especially educators — can appreciate how valuable an open-ended conversation of this kind with an LLM can be. Starting with a simple expression of curiosity on my part concerning the results of a study I had read about, Kimi and I delved into the issue of how best to interpret the findings. This has proved especially valuable given that the findings contradicted what appear to be in some people’s minds settled beliefs. That discussion led to the much deeper and broader treatment Kimi developed after my request to know about research in the areas we discussed.

In a very real sense, this conversation that began weeks ago is only just beginning. In tomorrow’s column, I push the dialogue further as I attempt to respond to tKimi’s four concluding questions.

Your thoughts

Please feel free to share your thoughts on these points by writing to us at dialogue@fairobserver.com. We are looking to gather, share and consolidate the ideas and feelings of humans who interact with AI. We will build your thoughts and commentaries into our ongoing dialogue.

[Artificial Intelligence has become a feature of everyone’s daily life. We unconsciously perceive it either as a friend or foe, a helper or destroyer. At Fair Observer, we see it as a tool of creativity, capable of revealing the complex relationship between humans and machines.]

[Lee Thompson-Kolar edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect Fair Observer’s editorial policy.