A tutor inside the course, lower grades
In autumn 2025 the University of Maryland did what few universities do before adopting a technology: it tested it with an experiment. Thirty instructors teaching different courses, from the sciences to the humanities, were split at random. In the courses of some of them, an AI study assistant based on GPT-4o appeared inside the university's learning platform, linked to the materials the instructor had chosen: slides, handouts, lecture transcripts. The others taught as usual. In all, 2,379 students took part.1
The results came out in October as a working paper from the Annenberg Institute at Brown University, and they point the opposite way to what many expected. Comparing sections of the same course, students with the tutor finished the semester with a final grade 4.12 points lower out of 100, or 0.37 standard deviations. Participation in the online activities set by the instructor, meaning quizzes, discussions and submissions, fell by 0.90 standard deviations. First-generation students, the first in their family to go to university, lost the most: across the whole sample the estimated effect on their grade was 5.1 points lower, against 2.2 for everyone else.
The researchers also read what students asked the tutor. Almost three requests in four were for information, explanations or ready-made solutions. Requests to practise or prepare for an exam were just over one in ten, and use clustered around tests.
What the study says, and what it does not
The authors themselves set out many limits.
It is a working paper, not yet through peer review at a journal. The fall in grades is clear only when comparing sections of the same course. Across the whole sample, where courses and grading differ, the effect on grades is not statistically significant, while the effect on participation stays large. And the grades are those given by instructors, not a common test.
One figure surprises the authors too. Only about 15% of the students who had the tutor opened it even once, for four sessions on average. An effect this large from use this low suggests that the tool was not the only thing at work. Everyone, including the control group, could already use ChatGPT and Gemini for free through the university. According to the authors, finding an AI assistant inside the course, introduced by the instructor, may have changed the mood of the class: more implicit permission to lean on AI, less time on course activities. In the surveys, students who used the tutor said they asked the instructor fewer questions and wrote to them less often.
Then there is the configuration. Instructors could choose between two modes: one that gives the answer directly and a "tutor" mode that walks the student towards the solution without handing it over. The first was the default, and only two instructors ended the semester with the second. None of the instructors who answered the survey had built it into assignments. The authors say so plainly: the results concern this assistant, set up this way, and say nothing about a tutor built to guide reasoning.
When the tutor asks questions
Other controlled experiments help explain why configuration matters so much.
In Türkiye a team led by Hamsa Bastani of the Wharton School gave GPT-4 to about a thousand secondary school students during maths practice, in two versions. The first resembled ChatGPT. The second, called GPT Tutor, had instructions written with teachers to give hints instead of solutions. During practice the first group did 48% better than students without AI, the second 127% better. Then the AI was taken away and a class test followed: those who had used the ChatGPT-like version did 17% worse than students who never had it, those who had used GPT Tutor about the same. The study appeared in PNAS in 2025 under a blunt title, which states that without guardrails generative AI can harm learning.2 It is the same gap between feeling and result seen among developers working with AI.
Practising with AI, then sitting the test without it
At Harvard, in autumn 2023, 194 students in the university's largest physics course were split at random into two groups: each had one lesson in class, taught with active learning, and one on their own with an AI tutor. The tutor had been written to follow the rules of good teaching: one step at a time, no free solutions, encouragement. With the tutor, median learning gains over the initial test were more than double, in a median time of 49 minutes against the lesson's 60.3
In March 2026, in the middle schools of Hamilton County, Tennessee, more than 6,000 pupils did a maths review on a platform with or without an AI tutor, assigned at random. The tutor was designed not to give answers. Those who had it did fewer exercises and worked more slowly, but after a mistake they got the next attempt right more often. A week later, among students who had to answer three questions in a row correctly before moving on, those with the tutor solved the question on the topic they had reviewed 40.2% of the time, against 37.0% for the others. That is three points, at the edge of statistical significance and again from a working paper, but in the right direction.4
In Nigeria, in Edo State, a six-week after-school programme in June and July 2024, with generative AI as a tutor and teachers alongside the students, produced a gain of about 0.3 standard deviations, which the World Bank compares to almost two years of conventional schooling.5
In January 2026 the OECD brought together the research published up to then in its Digital Education Outlook. Its conclusion: when the tool is designed or used without a teaching purpose, handing it the work improves performance but not learning, and the advantage seen while the AI is available disappears, sometimes reverses, in the exam without it.6
Four experiments, two ways to build a tutor
| Where | Who | The tutor | What happened |
|---|---|---|---|
| Maryland, university | 2,379 students | Answers questions on course materials | Lower grades in sections of the same course, less participation |
| Türkiye, secondary | about 1,000 students | ChatGPT-like version | Better practice, 17% worse on the test without AI |
| Türkiye, secondary | same experiment | Gives hints, not solutions | Better practice, test on a par with students without AI |
| Harvard, physics | 194 students | Guides one step at a time | Median gains more than double those of the class lesson |
| Tennessee, middle school | over 6,000 pupils | Gives no answers and slows students down | Three points higher a week later, at the edge of significance |
In Italy: phones out, AI in
In Italy phones in class have been banned in primary and lower secondary schools since a ministry note of 11 July 2024, and in upper secondary schools since a circular of 16 June 2025, which rules them out during school hours "even for teaching purposes".7 On 31 March 2026 the ministry said that over 90% of schools had already written the ban into their rules, and that almost 97% had adopted or were adopting the guidelines on artificial intelligence.8 The minister signed the guidelines on 9 August 2025.9
Over the same period AI came into the classroom by another door. In September 2024 the education minister, Giuseppe Valditara, announced a trial in 15 classes across Calabria, Lazio, Tuscany and Lombardy.10 Students worked with Gemini, NotebookLM and Google Workspace, teachers were trained by Google staff, and INVALSI, the national school assessment institute, compared the pilot classes with "mirror" classes that did not have those tools. As the newspaper Avvenire reported at the end of May, in most classes the real work began only in April 2025, when teacher training ended. And Paolo Branchini, the ministry adviser who designed the trial, admitted the biggest problem: AI had reached the comparison classes too, whose students talked to chatbots almost every day.11 It is the same difficulty as in the Maryland study, where the control group also had ChatGPT and Gemini. A class without AI to compare against is now hard to find.
On 23 September the president of INVALSI, Roberto Ricci, told the Culture Committee of the Chamber of Deputies that the first results "seem encouraging", especially in helping students past some difficulties, but that "we do not see direct effects on learning in some contexts". Further analysis will be needed.12 On 29 September an SWG survey of 796 students in 13 schools in the trial found differences of 5-10 points between pilot and control classes in questionnaire answers: for example, 38% of students in pilot classes said their relationship with teachers had improved, against 30%. Outside the pilot classes, meanwhile, AI is already there: in an Indire survey of 5,342 lower and upper secondary students, 94% say they use it at least occasionally, 62% every week or every day, 84% also for studying.13
For teachers, the AI decree approved by the Council of Ministers on 4 August, in force from 23 October, sets aside 100 million euros for training.14 At university level each institution makes its own choice. CRUI, the conference of Italian rectors, has signed a three-year contract with OpenAI, starting on 14 October 2025, that universities can join to get ChatGPT Edu licences.15 On 25 September the Politecnico di Torino announced Gemini for Education free of charge for its whole community, students included, starting in October, while ChatGPT licences remain reserved for permanent staff.16 These are decisions about which tool to buy. How to set it up for studying gets far less attention.
The tutor that thinks for you and the one that makes you think
The Maryland study should not be read as a verdict against AI in studying. It shows that putting a chatbot next to a course's materials, without changing how the course is taught, can be enough to make students study less. The tutor was built to answer, and students used it to get answers. Where the tool was built to ask for an attempt, give one hint at a time and start again from the mistake, the harm disappeared and in some cases learning improved.
A good tutor, human or artificial, hands the question back to the student instead of closing it with a ready answer. In October 2025 Valditara said: "Woe betide us if we were to treat AI as a prosthesis that risks atrophying the limb, that is, human intelligence."12 I agree. But whether it becomes a prosthesis or a training ground is not decided by the technology itself. It is decided by the settings and by how the teacher uses it, and in Maryland almost every instructor kept the default.
What worries me most is the result for first-generation students. They are the ones with fewer people to ask for help at home, which is exactly who a tutor available at any hour should help most. In Maryland they lost more than twice as much as the others. If the Italian trial is extended, I would like to see the results broken down by students' starting level, not only the average.
What to ask the school
To a parent or a teacher, when a school proposes an AI tool for studying, I suggest four questions.
- SettingsDoes it answer or ask?A tool that gives the solution straight away should be set up to ask for an attempt and guide with hints. In Maryland almost no instructor changed the default.
- ActivitiesDoes the teacher use it in assignments?A chatbot left at the side of the course ends up replacing its activities. It works better when the teacher decides which exercise to use it for, and why.
- TestingHow is it measured?Work done with AI says little. What counts is a class test without the tool, as in the experiment in Türkiye.
- FairnessWho falls behind?Ask for results broken down by groups of students: in Maryland first-generation students lost more than twice as much.
In May Avvenire reported that in the 2026/27 school year the trial would be repeated with more classes and with teacher training.11 Before extending it further, the ministry and INVALSI should be asked two precise things: which mode Gemini and the other tools will be set to in the new classes, and which tests without AI will be used to measure what students have learnt.