OneTutor learning report
What students really use AI for when they learn.Measured, not assumed.
In 54% of their messages, students let the AI do the thinking. When do they think for themselves? It depends on the week, the subject and the question.
628,000 student messages from 32 higher education institutions
Why we publish this report and how we built it01 · When
When do students use AI?
From the first week of lectures on, and more and more as the exams approach. About half of a semester's messages fall in the five weeks around the exams. Exam weeks make up only a sixth of all weeks, but carry a third of all messages.
Christmas shows up in the data.
In winter, week 11 is the quietest week after the first one: for most students in the data it is the week of Christmas, followed by New Year in week 12. Week 9, the second week of December for most, is a small peak. One possible reason: midterm tests.
Select Winter 2025/26 to see it.
Exam weeks bring more practice, not more outsourcing.
In exam weeks, practice rises from 17% to 22% of messages, while the share where the AI does the thinking falls from 55% to 51%. 39% of all quiz feedback happens in exam weeks. Practice still bunches up around the exams, and we want to spread it across the semester.
Both semesters, all academic disciplines. Switch the chart to Who thinks to follow it week by week.
02 · Academic disciplines
Which academic disciplines use AI, and how differently?
Six academic disciplines send 85% of all messages: engineering, business, computer science, law, the natural sciences and maths. That mostly reflects which faculties first brought AI tutors into large courses, not where students need them most. We expect other disciplines to catch up as AI tutors reach their students too.
Every academic discipline learns differently. In maths, students do the most thinking themselves (28% of messages). In psychology, quiz practice plays the largest role (35%). In law, almost half of all messages ask the AI to explain something or look it up. Click a discipline to filter every chart on this page.
Comparing winter and summer? Read this first.
Which courses took part changed a lot between the semesters. Computer science fell from 21% to 8% of all messages, mostly because its largest winter courses had almost no messages in summer. Most other disciplines grew: law from 8% to 19%, engineering from 14% to 22%.
Differences in who took part, not trends.
This shift in courses explains most of the change between the semesters: practice falls from 23% to 15% of messages and students think rises from 16% to 22%. Read these differences as a change in who took part, not as a change in how students learn.
03 · Requests and thinking
What do students ask for, and who does the thinking?
Across both semesters and all academic disciplines, students mostly ask for information: explanations and facts make up 36% of messages, and working on their own answers adds up to 28%. In 54% of messages, the AI does the thinking.
How to read the four groups
The groups follow the ICAP framework (Chi and Wylie, 2014): the more actively students engage, from passive to active, constructive and interactive, the more they tend to learn. The example messages are made up.
Students think
ICAP: interactive or constructive
“I got x = 4. Is that right?”
Students bring their own attempt, reasoning or a precise question.
Practice
ICAP: active
“Quiz me on enzymes.”
Students recall and apply what they know, mostly through quiz questions.
AI thinks
ICAP: passive
“Explain the difference between mitosis and meiosis.”
Students receive an explanation, a fact, a solution or a text.
No learning request
Not a learning activity
“Thanks!” or “When is the exam?”
Greetings, organisational and technical questions.
How do learning request categories break down?
The flow of student message use cases into four primary cognitive categories • Winter 2025/26 and Summer 2026
Targeted questions are most common in maths, law and engineering.
16% of messages in maths, 13% in law and 12% in engineering are targeted questions, against 4% or less in psychology, education, the social sciences and the humanities. That does not prove an AI tutor helps most in these disciplines, or that their students use it best. But the gap is large enough to follow up in the next editions.
Quiz feedback follows how a subject is assessed.
Feedback on quiz answers makes up 29% of messages in psychology, but only 3% in law, 4% in maths and 7% in agriculture, food and life sciences. Instructors there set fewer quiz questions, because the skills they teach are hard to test with multiple choice or short answers.
What we found and what we take from it
Four findings. Two things we want to change.
Key results · both semesters, all academic disciplines
01 · When do students use AI?
From the first week on, rising towards the exams. About half of a semester's messages fall in the five weeks around them.
02 · Which academic disciplines use it?
Six academic disciplines send 85% of all messages, led by engineering, business, computer science and law. That mostly reflects which faculties started first.
03 · What do students ask for?
Mostly information: explanations and facts make up 36% of messages. Students working on their own answers add up to 28%.
03 · Who does the thinking?
The AI does, in 54% of messages. Students think for themselves in 20% and practise in 18%.
1
From passive intake to productive struggle.
Most of the time, students use AI to retrieve information and take it in passively. We take this as a reason to change OneTutor and how it is deployed, so that it promotes productive struggle and moves learning from passive to active, constructive and interactive.
2
From practice before exams to practice all semester.
Quiz questions already move students from passive to active, but mostly around the exams. We are looking for ways to keep practice going through the whole semester.
What comes next
A new edition every semester. And your critique.
We will publish this report every semester to follow how the use of AI in teaching and learning develops. We also want to measure better: quiz performance, the tutor's replies and, where possible, learning outcomes.
This edition is a starting point, not a verdict. Dissect our approach, question our readings and tell us what we missed. The goal is to find out together how and when AI improves learning and teaching.
