Designing an AI-guided learning platform that scaffolds thinking instead of shortcutting it.
An evaluation-driven redesign of conversational AI for higher education — grounded in Bloom's Taxonomy and Microsoft's Human–AI Interaction Guidelines.
Generative AI is no longer an emerging trend — it's embedded in students' daily academic workflow. But the tools they use were never designed to teach.
of higher-education students have already used ChatGPT for coursework.
students surveyed across 109 countries and territories.
cite ChatGPT as their most-used AI tool — followed by Gemini (50%).
The pedagogical gap
Students receive polished responses that bypass the productive struggle on which understanding and retention depend. We see four converging problems.
Quick, complete answers displace the foundational thinking that assignments are designed to develop.
Emerging AI tutor tools — Study Mode, Guided Learning, Khanmigo — need systematic evaluation. Best practice has not been established.
Good practice should be theory-driven, grounded in established educational frameworks rather than ad-hoc UX heuristics.
Learning styles differ; a one-size-fits-all chat response ignores half a century of research on differentiation.
Research question
How can the interface of a general-purpose LLM be redesigned to support genuine learning across Bloom's Taxonomy — instead of accelerating answer retrieval?
5 testers · 4 tools · 6 cognitive levels · 2 subject domains · 1 rubric. Every interaction logged end-to-end.
12 prompts — 6 Bloom levels × 2 domains — written to a fixed role-verb-context structure.
Gemini, Claude, Microsoft Copilot, Sovi.AI — the most-used and emerging LLM tutors.
Each tester completed full interactions on every tool. Every turn logged for traceable analysis.
15 criteria across 5 dimensions, scored on a 4-point scale — Insufficient → Exemplary.
Affinity mapping of qualitative findings → 4 insight themes → 5 design interventions.
Scaffolding · Not Shortcutting · For College Students
Six cognitive levels, from basic recall up to original creation. As the demand climbs, tasks require thinking a fluent answer can't fake — exactly where LLM tutors tend to fail.
account_treeEach task followed a fixed structure — Bloom level × verb × role × subject keyword → final prompt.
We designed one task at each level, across two subject domains, then mapped exactly where LLM-as-tutor capability collapses — and targeted interventions at those collapses.
Every design decision in this project is traceable to one of two well-established bodies of work — Microsoft's Human–AI Interaction Guidelines and the practice of Responsible AI. Neither is decorative; both are load-bearing.
Existing LLM benchmarks measure accuracy. We needed an instrument that measures pedagogy, trust, interaction quality, integrity, and accessibility — 15 criteria across 5 dimensions, each scored on a 4-point scale: 1 Insufficient · 2 Developing · 3 Proficient · 4 Exemplary.
The strongest degradation appears in pedagogical support and interaction quality — not perfectly linear at every step, but the downward trend is consistent across all four tools.
Accuracy scores cluster more tightly. Variation is widest in Pedagogical Quality and Interaction — where Sovi.AI's gap from the others is most visible.
Gemini leads overall; Copilot scores highest on Usability; Claude is strongest at Evaluate-level reasoning. The matrix shows trade-offs, not a dominant system.
Each cell averages 5 testers across both subject domains, on a 4-point scale. Warmer = stronger.
We affinity-mapped every observation into four themes — each mapped to Microsoft's Human–AI Interaction Guidelines, giving every intervention an evidence-based provenance.
Tools treat every learner as the same hypothetical user. None asked about prior knowledge, preferred pace, or coaching tone before responding.
Tools frequently asked 3–4 clarifying questions in a single turn, overwhelming users mid-task and causing thread abandonment.
Most tools default to prose paragraphs. Learners who think in tables, diagrams, or dialogue cannot select these formats explicitly.
Hallucination risk is not surfaced, and finished responses make it trivial to bypass learning entirely. No friction protects integrity.
Every feature is anchored to a design scope and a Microsoft HAI Guideline. We deliberately resisted novelty for novelty's sake — each intervention closes one of the four mapped gaps, demonstrated inside the interface students already use.
Topic level (Beginning / Intermediate / Advanced), coaching tone, and response language captured up front — calibrating every subsequent response to the learner.
Complex prompts trigger a stepped clarification flow — one question at a time, sequenced by context, never stacked. The thread stays calm and finishable.
An inline, dismissable format picker — text paragraph, comparison table, or diagram view. The learner directs the modality instead of accepting the default prose.
A confidence rating (High / Medium / Low) and a visible source count sit beside every response — with one-tap source inspection to verify before trusting.
Final-practice prompts gate the answer: "Give me a hint" comes before "Reveal explanation." Friction is the feature — it protects the learning loop.
This project reframes LLM tutoring as a UX problem, not only a model problem. The interface determines when to pause, what to surface, what to withhold, and how to invite reflection. Those are designable.
Our 15-criterion rubric is tool-agnostic and openly documented — any LLM tutor can be re-scored as the landscape evolves.
The matrix locates failure at upper Bloom levels and in interaction quality — not factual accuracy — redirecting future investment.
Each provably grounded in both Bloom's Taxonomy and Microsoft HAI Guidelines.
