Beyond Prompts and Answers: What Really Turns ChatGPT Use Into Academic Growth

Access to generative AI may be transforming university life, but a major international study suggests that academic gains depend less on the technology itself than on how students feel, think and learn while using it.

Titled "Positive Affect and Academic Skill Development Through ChatGPT in Higher Education," the research was published in the European Journal of Investigation in Health, Psychology and Education. It was authored by Bonginkosi A. Thango, Lerato Matshaka, Alaa M. S. Azazz and Ibrahim A. Elshaer, researchers affiliated with the University of Johannesburg in South Africa and King Faisal University in Saudi Arabia.

Analysing survey responses from 12,035 active ChatGPT users across 135 countries, the study finds that students who feel curious, enthusiastic and engaged while using the tool tend to report stronger academic skill development, but emotions are only part of the story. The strongest predictor is whether students regard ChatGPT as useful and academically valuable, while the most self-directed learners appear less dependent on either enjoyment or a favourable opinion of the technology.

Overall, the study asserts that AI does not produce learning simply by being available. Its educational value is shaped by motivation, judgement, emotional engagement and the level of intellectual challenge surrounding its use.

The Missing Link Between AI Access and Academic Ability

Most higher education debates about ChatGPT have focused on adoption, cheating, plagiarism, reliability and assessment integrity. These concerns remain important, but they leave a deeper issue unresolved. Even when students use AI legitimately, what determines whether the interaction strengthens their writing, research, analysis and critical thinking?

The study approaches this question by examining three connected factors: positive affect, attitude toward ChatGPT and learning orientation. Positive affect refers to emotions such as curiosity, excitement, confidence and engagement during AI use. Attitude captures whether students see ChatGPT as useful, effective and beneficial for academic purposes. Learning orientation reflects a student's motivation to master difficult material, deepen understanding and treat academic work as an opportunity to build competence.

Students do not approach ChatGPT as neutral information processors. One learner may use it to interrogate competing explanations, identify weaknesses in an argument and revise several drafts. Another may use the same system to produce a quick answer with minimal reflection. Both have access to the technology, but only one may be converting that access into learning.

The researchers used data from the Global ChatGPT Student Survey, coordinated by CovidSocLab at the University of Ljubljana and administered between October 2023 and February 2024. The final sample included only students who actively used ChatGPT and had complete responses for the variables analysed.

Respondents came from 135 countries, and 66% were drawn from countries classified in the paper as lower-to-middle-income. Undergraduates made up 81.8% of the sample, while the gender split was close to even. The international coverage gives the findings relevance beyond wealthy university systems.

For institutions across the Global South, where generative AI may help address shortages of tutors, feedback and specialised learning support, the promise is considerable. However, the research also warns against assuming that widespread access will automatically close educational gaps. Without effective learning design, AI could simply create a faster route to task completion rather than a stronger route to knowledge.

Attitude Opens the Door, but Motivation Changes the Payoff

The study found that positive emotional engagement was significantly associated with students' perceived academic skill development. Students who experienced greater curiosity, enthusiasm and engagement during ChatGPT use were more likely to report improvements in research, writing, analysis, critical thinking and field-specific competence.

The statistical relationship was positive but not overwhelming. Positive affect had a standardised coefficient of 0.199 in predicting perceived skill development. It also predicted more favourable attitudes toward ChatGPT, with a coefficient of 0.267. The strongest direct relationship, however, came from attitude. Students who considered ChatGPT useful, effective and beneficial for academic work were more likely to report skill development, with a coefficient of 0.354.

This does not mean that simply liking ChatGPT makes students more capable. A more plausible interpretation is that favourable attitudes encourage sustained and purposeful use. Students who believe the tool has academic value may be more willing to question its answers, refine prompts, compare outputs and use feedback iteratively.

The researchers also found that attitude partially mediated the link between positive emotions and skill development. In other words, a stimulating or confidence-building interaction may contribute to learning partly because it strengthens the student's belief that ChatGPT is worth using as an academic resource.

However, the most revealing finding concerns learning orientation. Among students with a strong desire to master difficult material, the influence of positive emotions and favourable attitudes became weaker. These students appeared less dependent on enjoying the technology or regarding it especially positively. Their underlying commitment to learning may have been enough to sustain demanding, reflective engagement.

The study interprets this as a "desirable difficulty" effect. Highly motivated learners may continue examining flawed answers, testing alternatives and revising their work even when the process is frustrating. For them, difficulty is not necessarily a barrier; it can become part of the learning process.

Students with weaker learning orientation may need more emotional and evaluative scaffolding. A supportive introduction, confidence-building prompts and clear demonstrations of usefulness may help them remain engaged long enough to benefit. This challenges the idea that every AI-assisted learning experience should be made as effortless and enjoyable as possible.

The Danger of Designing Education Around Frictionless AI

Technology companies often compete by reducing friction. Faster answers, smoother interfaces and more effortless assistance are treated as evidence of progress. In education, however, the removal of difficulty can become a problem. Learning often requires struggle: evaluating evidence, recognising errors, revising assumptions and working through uncertainty. When AI completes too much of that process, students may feel productive without necessarily becoming more capable.

The study warns universities pursuing AI strategies centred on convenience. A system designed only to make students feel positive may help those who need early encouragement, but it may also weaken the intellectual challenge required for deeper learning.

For highly learning-oriented students, the research suggests that institutions should protect productive difficulty. ChatGPT could be paired with tasks requiring students to verify claims, identify bias, compare AI-generated arguments or explain why an answer is wrong. Students might first use AI support and then complete a similar task independently. Such designs would treat ChatGPT as a tool for practice and reflection rather than as a substitute for reasoning.

The findings also support differentiated AI pedagogy. Students with lower confidence or weaker mastery motivation may benefit from structured guidance and carefully designed introductory exercises. More advanced learners may need less motivational support and more demanding evaluation tasks. A uniform policy is unlikely to serve both groups equally well.

The study's model explained 54.2% of the variation in perceived academic skill development, suggesting that emotion, attitude and learning orientation together provide a powerful account of why students report different outcomes from ChatGPT use.

Yet "perceived" remains the crucial word. The research did not measure whether students objectively became better writers, researchers or critical thinkers. It measured whether they believed that they had improved.

The authors acknowledge that self-reported gains may not correspond closely with performance assessed through grades, standardised tasks or instructor evaluations. Because all variables were collected through the same survey at the same point in time, generally positive response patterns may also have strengthened the observed relationships.

The cross-sectional design further prevents causal conclusions. Students may feel positive because they are developing skills, rather than developing skills because they feel positive. Similarly, those who already perform well may be more likely to regard ChatGPT favourably.

The sample also came from convenience recruitment and purposive filtering. It provides extensive international coverage but should not be treated as statistically representative of all university students.

The paper's abstract and structural-results section refer to bootstrapping with 200 resamples, while the analytical-methods section states that 5,000 resamples were used. This discrepancy does not automatically invalidate the findings, but it should be corrected because the number of resamples affects the precision of statistical estimates.

From AI Rollout to Learning Design

The research has an important implication for governments and universities: the next phase of educational AI policy must move beyond access. Institutions have often treated technology deployment as an infrastructure question: buying licences, issuing guidelines and training staff to operate new systems. The study suggests that successful adoption also requires attention to learning psychology.

Universities need AI literacy programmes that teach students how to verify information, test reasoning, recognise uncertainty and retain responsibility for academic work. Faculty members need support in designing tasks that reward interpretation and judgement rather than the reproduction of AI-generated content.

Governments and development agencies should also resist measuring success only by the number of institutions or students using generative AI. More meaningful indicators would examine whether AI-supported learning improves independent performance, reduces educational inequality and strengthens durable skills.

This is particularly relevant to progress toward SDG 4 on inclusive and equitable quality education. Generative AI may expand access to explanation and feedback in resource-constrained settings, but access without pedagogy could produce shallow gains. It could also widen disparities between students who know how to question AI and those who accept its outputs uncritically.

Future research should thus combine perception surveys with objective assessments of writing, analysis and problem-solving. Longitudinal studies could test whether improvements persist after students stop using AI, while experiments could compare different types of scaffolding and intellectual challenge. The authors similarly call for objective performance measures and research capable of addressing reverse causation.