AI tutors: help, risks and independent learning

In the studies below, structured AI systems supplied on-demand explanations, hints or feedback on a learner's attempt. They could help a learner locate a mistaken step, work at an individual pace, or ask a question at the moment it arose. These are findings about particular systems, subjects and comparisons; they do not establish that any AI chat will teach effectively.

This page keeps three claims separate. External evidence reports what the cited studies measured. AXIOVUM design guidance explains how that evidence informs a proposed learning interaction. AXIOVUM learner outcomes would require a separate evaluation of this platform. AXIOVUM currently has no learner-outcome result to report, so this page gives no AXIOVUM effect estimate or success rate.

An answer that looks fluent can still be wrong, mistimed or educationally unhelpful. Kestin and colleagues reported that a prompt alone did not reliably scaffold multipart problems in sequence, so their platform combined expert-prepared materials with a fixed progression through the activity. A learner should therefore check definitions, assumptions and algebra against the course material rather than treating fluency as verification.

Assisted success can conceal dependence. In Bastani and colleagues' study, unrestricted GPT access improved practice performance but students performed worse once the tool was removed. Chen and colleagues' interview participants also raised concerns about over-reliance. Copying a solution, recognising an explanation and constructing a new argument are different achievements.

Platform progress is not automatically durable learning. Oreopoulos and colleagues found that a three-correct-in-a-row rule increased attainment of that platform threshold, but mastery progression alone did not produce a clear gain on the one-week delayed assessment. Completion counts, answer streaks and time on the platform should therefore remain process measures, not substitutes for an independent task.

Kestin et al., Scientific Reports (peer-reviewed, 2025): 194 eligible students in an introductory university physics course took part in a randomised crossover comparison across two lessons. The custom tutor and an in-class active-learning lesson used identical content; each lesson had a pre-test and an immediate post-test. Post-test performance was higher with the structured tutor, and median tutor time was 49 minutes versus an assumed 60 minutes of in-class learning time. The study did not test delayed retention, and the authors did not claim that the result must extend to work requiring complex synthesis or higher-order critical thinking.

Bastani et al., PNAS (peer-reviewed, 2025): nearly 1,000 secondary-school mathematics students were assigned by class to no AI, a GPT-4 chat interface, or a GPT-4 tutor with teacher-designed hints and safeguards. Within each session, assisted practice was followed by an unassisted exam. Both GPT conditions raised practice grades; after AI was removed, the unrestricted-chat group scored 17% lower than the control group. The safeguarded tutor largely removed that harm, but its unassisted exam score was not significantly higher than control. This was a short-term study of mathematics review in one school context, not a test of long-term transfer across subjects.

Chen et al., arXiv:2509.16778 (preprint, 2025): 155 volunteers from an introductory university discrete-mathematics course were randomised to earlier or later access to a proof-review tutor plus chatbot. After the first withdrawal checkpoint 148 remained, and 145 remained by the final examination, so the analysed sample varies by outcome. The earlier-access group scored higher on the two homeworks before the second midterm, while no statistically significant condition difference appeared on the midterm or final examinations. The measurements spanned course homework and examinations rather than a single immediate test. As a preprint from one course with voluntary participation, it needs replication and peer review. A non-significant exam difference does not prove that the true effect is zero.

Oreopoulos et al., EdWorkingPaper 26-1552 (working paper, 2026): more than 6,000 students in grades 6–8 across 20 schools were individually randomised across AI versus otherwise comparable computer-assisted learning, mastery versus non-mastery progression, and one of two mathematics topic bundles. A one-week delayed test covered practised and unpractised material. In the mastery workflow's Exercise 1, AI improved next-attempt correctness after a mistake and reduced the attempts needed to recover, while slowing progress and reducing the number of questions completed. The clearest delayed result was about a three-percentage-point advantage for AI within the mastery workflow on practised Exercise 1 material; it was only marginally statistically significant, with essentially no difference on the paired unpractised item. That is limited, context-specific evidence, not a stable or general learning-effect claim.

Use a five-step hand-back routine. First, if you have an attempt, show it; if you have not started, name the exact idea or line you cannot enter and ask for an explanation or example. Second, ask for the smallest useful help or for the condition that permits the next step; a complete explanation is available when needed, but it should not end the work. Third, explain the reason back in your own words and check it against the course source. Fourth, close the explanation and produce the next step yourself. Fifth, try a changed condition or a new example without AI. If you cannot reproduce or adapt the reasoning, treat the topic as still in progress.

AXIOVUM design illustration—not an observed learner interaction or evidence of effectiveness: while minimising ‖Ax − b‖², a learner asks why the normal equation follows. The tutor asks which subspace contains the residual and what must be true at the closest point. The learner states that b − Ax must be orthogonal to every column of A, derives Aᵀ(b − Ax) = 0, and then considers what changes if A's columns are linearly dependent. The important event is the learner completing and testing the reasoning after the prompt.

The proposed AXIOVUM approach is to explain a difficult step, then return the next inference to the learner and later check an unaided variation. This is a design choice informed by the external evidence above. It remains a hypothesis until AXIOVUM measures immediate independent work, delayed transfer and possible adverse effects in an ethically reviewed study.