Rooted in research, built with intention
Every design decision in Aristotle traces back to peer-reviewed learning science. Here's the research behind how we teach.
The Foundation
One-on-one tutoring is the most effective form of education ever measured
In 1984, educational psychologist Benjamin Bloom described what he called the “2 Sigma Problem.” Students who received one-on-one tutoring learned far more than those in typical classroom instruction, and no group teaching method came close.
Decades of research since have pointed to the same conclusion: personalized tutoring is one of the most effective forms of instruction ever studied. The challenge has always been making it available to every student.
The Evidence
AI tutoring works
Intelligent tutoring systems have been studied for decades. The consistent finding: they approach the effectiveness of human tutors. Recent trials with modern AI show the gap closing further.
AI tutoring outperforms active learning (Harvard, 2025)
A randomized trial of 194 physics students at Harvard found that students using an AI tutor learned more, and in less time, than students in active learning classrooms with peer instruction and real-time feedback.
Read the paper →
AI tutoring matches human tutors (Google DeepMind, 2025)
Across five UK secondary schools, students tutored by an AI system performed as well as those tutored by humans, and were more likely to solve novel problems on their own afterward.
Read the paper →Why It Matters
Most AI gets teaching fundamentally wrong
General-purpose AI is optimized to answer questions. A good tutor is optimized to build understanding. These are fundamentally different objectives.
Our Approach
The science behind every design decision
Each feature in Aristotle traces back to a specific finding in learning science.
Socratic method, not answer delivery
We guide through questions, never give the answer. This outperforms direct instruction by 12%+.
Standard LLMs have a strong bias toward revealing solutions directly. Research in pedagogical reinforcement learning shows that overcoming this bias requires explicit training against answer-giving. SocraticLM, presented at NeurIPS 2024, demonstrated that an LLM fine-tuned for Socratic dialogue outperformed GPT-4 by over 12% on teaching performance metrics. Aristotle is built on this principle: every response asks a guiding question rather than providing the answer.
Expert-level error remediation
When a student errs, we follow the same 3-step process expert tutors use, not the novice shortcut.
Research on the Bridge framework shows that expert tutors follow a structured process when students err: identify the specific error, choose a remediation strategy, and form a pedagogical intention before responding. Novice tutors skip these steps, jumping straight to correction. When LLMs are guided by this expert decision-making structure, tutoring quality improves dramatically. Aristotle’s error handling mirrors this expert process.
Misconception diagnosis
We model the reasoning pattern behind the mistake and address the root cause, not just the error.
The MISTAKE framework uses cycle consistency to model the relationship between incorrect answers and their underlying misconceptions. This allows the system to anticipate why a student is confused — not just that they are — and to address the specific conceptual gap. This is fundamentally different from simply marking an answer wrong and re-explaining the procedure.
Voice-first = deeper learning
The ICAP framework proves it: interactive engagement produces deeper learning than passive consumption.
The ICAP framework, one of the most cited works in learning science, establishes a clear hierarchy of cognitive engagement: Interactive > Constructive > Active > Passive. Typing a question into ChatGPT and reading the answer is passive. Speaking through a problem with a tutor who responds to your reasoning is interactive — the highest level. Aristotle’s voice-first design is a direct application of this principle: students explain their thinking out loud, producing the constructive and interactive engagement that decades of research shows produces the deepest learning.
Turn-by-turn verification
Every tutor response is checked against pedagogical criteria before it reaches your child.
The TRAVER framework combines knowledge tracing with turn-by-turn verification to ensure each tutor response proactively guides students toward understanding, not just toward the answer. Research shows this achieves significantly higher task completion rates than unverified tutoring approaches. Additionally, work on conversational uptake demonstrates that when tutors build on student contributions — acknowledging, reformulating, and extending what the student said — student achievement improves. Aristotle verifies every response for pedagogical quality before delivery.
Evaluation
How the field measures AI tutoring quality
The research community has developed rigorous frameworks for evaluating whether AI tutors actually teach. We build against these standards.
Pedagogical evaluation taxonomy
The field has converged on 8 pedagogical dimensions for evaluating AI tutors, with MRBench providing the gold-standard benchmark of 192 conversations and 1,596 expert-annotated responses.
Read the paperScale AI’s TutorBench
A benchmark of 1,490 expert-crafted prompts evaluating adaptive explanations, actionable feedback, and hint generation. No frontier model scores above 56%, showing that raw language ability alone doesn’t make a good tutor.
Read the paperGoogle’s LearnLM framework
Google’s evaluation-driven approach to educational AI identified key pedagogical principles and showed that educators and learners consistently preferred pedagogically fine-tuned models over standard LLMs.
Read the paperThe AI Teacher Test
Research testing GPT-3 and Blender as teachers found both were quantifiably worse than human teachers, especially on helpfulness. Raw language ability does not equal teaching ability — pedagogical fine-tuning is essential.
Read the paperGive your child the advantage they deserve
Set up your family's Aristotle account in under two minutes. Tailored, research-backed tutoring, ready when they are.
Get Started