A review of 36 studies involving more than 7,000 university students found that generative AI generally improved learning, but the strongest results depended on how the technology was built into teaching rather than which platform or AI model students used.
A university student can ask an AI chatbot to write an answer, explain an unfamiliar idea, challenge an argument or help a group compare several possible solutions.
All four activities use broadly the same technology.
They do not necessarily involve the same amount of learning.
That distinction is becoming more important as universities move beyond the first question raised by generative AI, whether students should be using it at all, and begin confronting a more difficult one: what should students actually be doing with it?
New research suggests the answer may matter considerably.
Researchers combined results from 36 studies involving 7,229 higher-education students and found that generative AI had an overall positive effect on learning. Yet when they examined why some educational uses performed better than others, the technology itself was not the strongest differentiator.
The teaching method was.
AI produced its strongest gains in collaborative learning
The researchers analysed 132 separate effect sizes and compared several aspects of AI-supported learning, including the subject being taught, type of task, length of the intervention, AI platform, model version and teaching method.
Overall, generative AI produced a moderate positive effect on learning.
The strongest gains appeared in outcomes related to understanding, cognition and creativity, followed by higher-order learning. Improvements in conventional attainment measures, such as academic performance, were positive but smaller.
The more unusual result emerged when researchers compared teaching approaches.
Collaborative learning produced the largest measured benefit, followed by blended learning, where technology was deliberately combined with face-to-face teaching.
Inquiry-based, personalised and traditional teaching approaches also produced positive results, but the gains were smaller.
Teaching method was the only factor examined where the differences between groups reached statistical significance.
The same chatbot can play very different roles
The result makes more sense when the classroom activity is considered rather than the software.
In the studies classified as collaborative learning, students had to work with other students through activities such as group projects, collaborative writing or shared problem-solving.
AI could therefore become something to discuss.
A chatbot might generate an explanation, but students still have to decide whether it is convincing. It may suggest an argument, but the group can challenge it, improve it or compare it with another response.
That creates a different learning experience from asking the same system to complete an individual task from beginning to end.
The researchers argue that these interactive environments create more opportunities for students to explain, critique and revise ideas instead of simply accepting an answer.
A better AI model did not automatically mean better learning
Perhaps the most interesting result for universities investing in AI is what did not appear to make a significant difference.
The researchers found no statistically significant moderating effect for the AI platform students used or the model version.
They also found no significant differences according to discipline, measurement method, intervention length or whether students were mainly acquiring knowledge or solving problems.
That does not mean all AI systems are equally capable.
It means that, within the studies included in this review, differences in the technology did not explain differences in learning outcomes as clearly as the way educators organised the learning around it.
For universities, this shifts attention away from a purely technological question.
Buying access to a more advanced AI system is not the same as designing a better learning experience.
Better work is not always better learning
There is still an important risk.
AI can make it easier for students to produce stronger work without necessarily developing the abilities that the work is supposed to demonstrate.
A well-written essay, correct calculation or polished presentation can show what a student submitted. It does not always reveal how much thinking happened before submission.
The researchers warn that excessive reliance on AI may allow students to offload cognitively demanding work to the technology. AI systems can also produce inaccurate or misleading information with considerable confidence.
Assessment therefore becomes part of the same problem.
If an assignment can largely be completed by an AI system, universities may need to reconsider whether that assignment still measures the learning it was originally designed to test.
South African universities face the same design question
For South African higher education, the finding offers a useful way of framing the AI debate.
The choice does not have to sit only between allowing AI and prohibiting it.
Universities can also ask where AI belongs inside the learning process.
A lecturer may use it to create discussion, compare explanations or help students interrogate an argument. A student may use it to obtain feedback before revising their own work. Groups may use generated responses as material to analyse rather than answers to submit.
In each case, the educational value comes partly from what happens around the technology.
That matters in an environment where access to the newest AI model can change quickly, while the quality of teaching remains a much more fundamental educational question.
The evidence is encouraging, but still young
The review provides stronger evidence than any single classroom experiment because it combines findings across dozens of studies.
It does not settle the issue.
The researchers only included peer-reviewed studies published in English or Chinese, and some learning outcomes were represented by relatively few studies. They also call for more long-term research, particularly into areas such as critical thinking and motivation.
Generative AI itself is changing faster than universities can study its effects.
But the evidence is beginning to move the education debate in a useful direction.
The important question may no longer be whether a student has access to the most powerful AI.
It may be what the student is being asked to do once the chatbot is open.
Source Information
Study Title: Exploring the effect of GenAI on learning outcomes in higher education: a three-level meta-analysis
Authors: Changxin Fan, Lele Ke, Zexiong Chen and Pin Lv
Journal: Frontiers in Psychology
Published: 15 May 2026
DOI: 10.3389/fpsyg.2026.1758670
