An international review identifies potential classroom benefits from generative AI, alongside unanswered questions about lasting understanding.
A review by the University of Sydney and the Barker Institute is drawing attention to an important distinction: helping a student complete an assignment is not necessarily the same as helping that student learn.
In its 3 September 2026 announcement, the University said the review covered 271 studies from more than 45 countries. Only four were Australian. The findings suggested benefits for engagement and task performance, but significant gaps remained in understanding effects on reasoning, self-regulation and longer-term development. University of Sydney announcement.
In the announcement, Professor Lina Markauskaite emphasised that outcomes depend on how students use the tools, the tasks they undertake and the support they receive. University of Sydney announcement.
What the evidence covers
According to the Barker Institute, researchers screened 975 publications before selecting the 271 included in the 52-page report, released in August 2026. The institute also made launch webinars available to help educators, policymakers and parents engage with the findings. Barker Institute report overview.
The report is a synthesis of existing research, not a single experiment involving 271 schools. That distinction matters when interpreting its scope: a large collection of studies can identify patterns and gaps without establishing one result that applies to every classroom.
When the assistance stops
An earlier experiment illustrates the distinction between performance and learning. In research published in PNAS in 2025, Hamsa Bastani and colleagues tested two GPT-4-based interfaces with nearly 1,000 mathematics students at one high school in Türkiye: a basic chat-style tool and a tutor with teacher-informed safeguards.
Both AI groups outperformed the control group during supported practice. Once assistance was removed, however, students assigned the basic tool scored about 17% below the control group on subsequent exams. The safeguarded-tutor group did not differ significantly from the control group on those exams. Bastani and colleagues’ research paper.
The experiment measured short-term outcomes in one school. It does not establish lasting harm from every AI system, but it shows why better assisted work should not automatically be equated with independent understanding. Study scope and limitations.
Learning is not the only consideration
UNESCO’s 2023 guidance on generative AI in education calls for protecting personal data, taking learners’ ages into account and preserving human agency. It also encourages institutions to assess the educational and ethical suitability of tools. UNESCO guidance.
Read alongside the classroom research, that guidance suggests two separate responsibilities: checking whether a tool supports learning and checking whether its use is appropriate for the students involved.
Three questions for parents and schools
One practical way to discuss this evidence is to ask:
- What should the student be able to explain or do independently?
- How will that understanding be checked after AI assistance ends?
- What age, privacy and supervision safeguards apply to the tool?
These are editorial discussion prompts, not a validated assessment checklist. They move the conversation beyond whether an assignment looks impressive, towards what the learner can demonstrate and the conditions in which the work was produced.
