Rethinking What an AI Tutor Should Be.

‍

The default model of AI help had to go. Ask a question, get an answer. That model turns AI into a well-intentioned genie granting wishes instantly.

What students actually need is more like a personal trainer. Someone who spots, pushes, but doesn't lift the weight for them.

MENTOR was grounded in two educational philosophies.

First: Socratic questioning. Teaching by asking probing questions instead of giving lectures. A good human tutor often responds to a student's question with another question, guiding them to figure it out themselves. Frustrating sometimes. But it works. It stimulates deeper understanding.

Second: scaffolding theory from cognitive psychology. Providing support when a learner is struggling, then gradually removing that support as competence develops. Like scaffolds on a building under construction, taken away when the structure can stand on its own.

Combined: Socratic scaffolding. MENTOR never hands over the solution. It engages in dialogue. It asks "Why?" and "What if?" and "How might...?" It gives hints or context when truly needed, then steps back as confidence grows.

The goal: keep the student in that sweet spot of challenge. Not so hard that frustration takes over. Not so easy that passivity sets in.

There's another consideration: anthropomorphism. The tendency to treat machines like living beings. Studies show that if an AI tutor is too human-like or too friendly, students might develop parasocial relationships. Trusting it too much. Feeling emotional bonds. Engagement goes up, but so does dependency.

MENTOR was designed to avoid this. No cute avatar. No human name. The student should always remember: this is a tool, not a person.

Here's something interesting: using multiple specialized agents instead of one helps prevent anthropomorphism. One AI that chats might start feeling like a virtual buddy. But a team of AIs, each with a distinct role and voice, feels more like functional components of a system than a singular entity to befriend. Less Jarvis. More control room.

Three Groups. Same Challenge. Measured Against Real Metrics.

‍

A formal evaluation was organized to see if MENTOR delivered on its promises.

Participants were randomly assigned to one of three groups for a design challenge:

One group used MENTOR. The full multi-agent system.

A second group used a generic AI assistant. Direct answers. Help on demand. No Socratic features. No multi-agent architecture. This was the "active control" to see how the typical AI approach compares.

A third group had no AI assistance at all. Just the student, their brain, and the task.

Each participant had the same problem to solve. An open-ended design brief. Simple enough to be accessible, complex enough to require creativity and thinking. Fixed time. Every interaction collected.

What Was Measured and How:

Cognitive offloading was tracked through linguistic pattern detection. Phrases indicating dependency ("just give me the answer," "what should I do"), instances of copying AI suggestions verbatim, and the ratio of AI-generated content to student-originated thinking.

Deep engagement was quantified using natural language processing to assess reasoning complexity. Usage of causal language ("because," "therefore"), hypothetical scenario exploration ("what if," "alternatively"), and analytical terminology indicating strategic thinking rather than surface-level responses.

Design process quality was mapped using linkography, a method that represents the design process as a network of ideas. Each idea or design move is plotted as a node, and connections between ideas become edges. A richer, more interlinked map suggests thoughtful, iterative design. A sparse map with few connections suggests linear thinking or passive acceptance of given solutions. Link density was calculated for each session.

Human-AI relationship health was assessed through anthropomorphism detection. Language patterns indicating emotional attachment, human attribution ("the AI thinks," "it wants me to"), blind trust without questioning, and dependency signals were tracked. Subjective feedback was also gathered: Did the AI feel controlling or empowering? Did students maintain healthy skepticism?

All metrics were benchmarked against three baselines: the generic AI group (representing current standard practice), the no-AI control group (representing unassisted human capability), and predefined thresholds established from educational psychology research on healthy learning patterns.

What Happened in Practice:

With the generic AI group: A student asks "What should I include in a community center?" The AI responds with a comprehensive list: gymnasium, multipurpose rooms, kitchen, childcare center, computer lab. The student copies it. Done. Time elapsed: less than one minute.

With MENTOR: Same question. But instead of an answer, the Socratic Tutor Agent responds: "Who do you imagine using this space throughout a typical day?"

The student pauses. Thinks. "Well... elderly people in the morning?"

"Interesting. What might they need that's different from teenagers after school?"

Suddenly, the student isn't just listing rooms. They're thinking about human needs, daily rhythms, intergenerational dynamics. They're designing from understanding, not from a template.

The test was structured in three phases. An Ideation phase where the student sets up the groundwork of what was meant to be designed. A Visualization phase where the design starts getting conceptualized around the ideas discussed. A Materialization phase where those concepts get solidified into concrete form. Each serving its own purpose, the same way any creative work evolves from abstract conceptual ideation into specific detail.

At first, some students were taken aback. "This AI is answering my questions with more questions?"

But they didn't bail. They kept at it. Gradually the shift became visible in the session metrics: engagement with MENTOR's questions, bouncing ideas, refining.

Meanwhile, in the generic AI group, things played out generically. Those students would ask for an idea, the AI would give a decent suggestion, maybe a whole sample outline. Most ran with those AI-suggested ideas, only lightly tweaking them. Faster upfront. But they didn't dig as deep or explore alternatives.

The no-AI control group was mixed. Some struggled in silence. Others came up with great ideas on their own. Not having any guidance was challenging for many. A cognitive workout without a trainer. Some thrived. Some didn't. Most gave up.

‍