The research behind Inkling

We start with a simple question: what helps students learn?

Here is the research shaping how we think about great tutoring - and how we can make more of it possible.

01 — The foundation

One-to-one tutoring works

In 1984, Benjamin Bloom found that students receiving one-to-one tutoring performed two standard deviations better than students taught conventionally. An average tutored student outperformed around 98% of the classroom group.

Bloom called this the Two Sigma Problem. Since then we've knew how powerful individual tutoring can be, but not how to make it available to everyone.

One classroom of 100 students

98 students the average tutored student outperformed

The 2 they didn’t

The papers

Benjamin Bloom, 1984

The 2 Sigma Problem

Bloom’s seminal paper established the potential of individual tutoring—and the challenge of reproducing it at scale.

Benjamin Bloom, 1984

The 2 Sigma Problem

Bloom’s seminal paper established the potential of individual tutoring—and the challenge of reproducing it at scale.

Nickow, Oreopoulos and Quan, 2024

The Promise of Tutoring for PreK–12 Learning

A systematic review and meta-analysis of experimental tutoring programmes.

Nickow, Oreopoulos and Quan, 2024

The Promise of Tutoring for PreK–12 Learning

A systematic review and meta-analysis of experimental tutoring programmes.

02 — The opportunity

AI can now deliver tutoring at scale

AI can hold open-ended conversations, respond to individual students and adapt explanations in real time. Early controlled studies suggest that carefully designed AI tutors can already produce meaningful learning gains.

But an intelligent model is not automatically an intelligent tutor. Other research has found that unrestricted AI can help students complete work without helping them perform independently afterwards - and that thoughtful pedagogical guardrails are needed to reduce this effect.

In a 2025 Harvard trial, students using an AI tutor learned more in less time than students receiving an in-person active-learning lesson covering the same material.

Kestin et al., 2025

AI Tutoring Outperforms In-Class Active Learning

A randomised controlled trial of a deliberately structured AI tutor in an undergraduate physics course.

Kestin et al., 2025

AI Tutoring Outperforms In-Class Active Learning

A randomised controlled trial of a deliberately structured AI tutor in an undergraduate physics course.

Bastani et al., 2025

Generative AI Without Guardrails Can Harm Learning

A large field experiment showing why AI must be designed to support independent learning, not simply task completion.

Bastani et al., 2025

Generative AI Without Guardrails Can Harm Learning

A large field experiment showing why AI must be designed to support independent learning, not simply task completion.

03 — The craft

Teaching takes more than intelligence

A powerful language model can explain almost anything. That does not make it a great tutor.

Inkling is designed to reproduce the behaviours of top human tutors: leading the learning, uncovering how each student thinks and adapting continuously.

Scroll across

I

They lead
the learning

Students should not need to know the right question to ask. Great tutors use Socratic questions and scaffolding to move the lesson forward without taking the thinking away.

Kurt VanLehn, 2011

The Relative Effectiveness of Tutoring Systems

A major review finding that systems offering guidance throughout problem-solving achieved substantially stronger results than answer-based systems.

II

They uncover how
each student thinks

A final answer shows whether a student was right. Conversation reveals how they were thinking: surfacing misconceptions, uncertainty and connections that would otherwise remain hidden.

Chi, de Leeuw, Chiu and LaVancher, 1994

Eliciting Self-Explanations Improves Understanding

A seminal study showing how prompting students to explain their reasoning can deepen understanding.

III

They adapt
continuously

Every response changes what a great tutor does next: the question they ask, the explanation they choose, the challenge they set and the pace at which they move.

Black and Wiliam, 1998

Assessment and Classroom Learning

A foundational review of using evidence of student understanding to continually adapt teaching.

04 — The human side

Learning is social

Learning is shaped not only by the content of instruction, but by the social context in which it happens.

01

Social presence matters

02

Learning depends on a wider support system

03

Human relationships are essential

Learning depends on a wider support system

Even excellent tutoring is only one part of a student’s education.

Learning outcomes improve when students, parents and educators are all pulling in the same direction, and all share information, expectations and next steps.

Inkling works with all of these parties to build these compounding relationships.

Smith et al., 2020

The Effects of Family–School Partnership Interventions

A meta-analysis of 77 interventions examining how collaboration between families and schools affects student outcomes.

Human relationships are essential

In person teaching will always remain key.

An important purpose of education is where students learn with and from other people. Cooperative learning, peer learning, and learning about how to exist as part of a society are factors that technology should create more room for, not try to replace.

AI can provide more individual explanation, practice and immediate feedback, freeing teachers to spend more time noticing, challenging, encouraging and connecting.

Roorda et al., 2011

Teacher–Student Relationships, Engagement and Achievement

A meta-analysis of 99 studies examining how the quality of teacher–student relationships relates to engagement and achievement.

05 — The standard

Led by evidence.
Measured by outcomes.

An AI tutor should not be judged by how intelligent its conversation sounds, or how long students spend using it. It should be judged by what they can do afterwards.

AI tutoring is still a rapidly emerging field. We believe progress should be driven by evidence: combining learning science, expert evaluation, real-world data and measured student outcomes to understand what works.

Our own evidence

In a controlled study with 119 GCSE students, students using Inkling improved on the topics they were tutored in.

+26.6%

on tutored topics

+9.3%

overall

As the field develops, we will continue combining learning science, emerging AI research, expert review and real student outcomes. The evidence will decide what we build next.