When Students Trust the Machine: Gemini Boosts Programming Grades in Nigerian Classrooms

Generative artificial intelligence has swept through classrooms faster than almost any educational technology before it, but a nagging question has trailed the hype: do students actually trust these systems enough to learn from them? A new quasi-experimental study from researchers at the University of South Africa, Hong Kong Metropolitan University and the University of Nigeria offers one of the most detailed answers yet. Christian Basil Omeh and colleagues found that undergraduates who used Google Gemini inside a problem-based programming course scored dramatically higher on achievement tests and reported far greater trust in AI than peers who tackled identical problems without machine assistance. The effect was large, statistically robust, and, strikingly, identical for men and women.

The study, published in Discover Education, focused on sixty-two second-year computer science education students drawn from two government-owned universities in a developing African context. Programming is notoriously difficult to teach; introductory courses are frequently marked by frustration, disengagement and dropout, because learners must simultaneously absorb an unfamiliar non-natural language, its syntax and semantics, and the abstract art of decomposing problems into algorithms. In many Nigerian institutions, instruction still relies on traditional lecturing, an approach the authors argue has done little to improve the learning experience. The team instead paired problem-based learning, a constructivist method in which students collaboratively analyse authentic problems before writing code, with structured access to Gemini.

The design was deliberately rigorous for a classroom setting. Because intact classes could not be randomly shuffled without disrupting academic schedules, the researchers used a non-equivalent control group pre-test–post-test design. Twenty-nine students received Gemini-enhanced problem-based learning while thirty-three followed the same problem-based sequence without AI support. Crucially, the control condition was not passive lecturing: both groups covered the same curriculum, completed equivalent tasks, received equal instructional time and were assessed with identical rubrics. The only meaningful difference was that the experimental group could consult Gemini as a scaffold, while the control group relied on course materials, peer collaboration, manual debugging and instructor feedback.

Over eight weeks, the intervention followed a carefully scripted progression. Week one established baselines using a thirty-one-item Multi-Dimensional Measure of Trust and a thirty-five-item computer programming achievement test. Subsequent weeks moved from basic syntax through algorithm design and pseudocode, data structures, modular functions and a regression computation, culminating in a collaborative capstone project. In the experimental group, Gemini was introduced only after students had attempted to interpret each problem and draft an initial solution, a rule designed to prevent the chatbot from replacing independent reasoning. Students then used the tool to explain unfamiliar syntax, critique algorithms, diagnose errors, generate test cases and review function structures, recording whether each suggestion was accepted, modified or rejected. Gemini never completed graded work or assigned marks.

The quantitative results were striking. On the pre-test, the two groups were statistically indistinguishable, with means of roughly twenty-three and twenty-four points. By the post-test, the Gemini-enhanced group averaged 73.90 against 57.37 for the control group, a difference that produced an effect size of nearly 0.79 in the analysis of variance. Because variance differed between groups, the team supplemented conventional analysis of covariance with HC3 robust estimation and bias-corrected bootstrapping. The adjusted mean difference held at 16.42 points with a 95 percent confidence interval of 14.16 to 18.67, confirming that the achievement gap was not an artefact of unequal spread or influential outliers.

Trust told a parallel story. Measured across four dimensions drawn from McKnight’s framework, performance trust, moral trust, ethical trust and transparency, trust scores climbed from around 1.6 on a five-point scale at baseline to above 4.2 in the Gemini group, while the control group rose only to the mid-2s. The analysis of covariance confirmed that instructional condition, not gender, drove the difference: gender itself and the condition-by-gender interaction were both statistically nonsignificant. That finding matters in a context where women remain under-represented in computing and where scepticism about technology adoption is often framed through a gendered lens. Here, female students in the AI-supported group actually posted the largest trust gains, rising 2.68 points on average compared with 1.41 for their counterparts without AI.

To understand why the intervention worked, the researchers interviewed eight experimental-group students, four men and four women, using inductive thematic coding. Participants described Gemini as valuable across the full arc of programming work: clarifying problems, developing algorithms, debugging code and enabling self-directed learning. One interviewee linked the approach to improved problem-solving, algorithm skills and structured program design; another said it motivated engagement in coding classes and described natural-language interaction as increasing their enthusiasm for coding. Others associated trust with transparent prompts, responsible data handling and clearly defined lines of responsibility between what the AI generates and what is expected of the student, suggesting that trust coexisted with human oversight rather than blind acceptance.

The authors ground these findings in an unusual theoretical pairing, combining Bourdieu’s concept of habitus, the idea that dispositions are formed through social and educational experience, with McKnight’s technology-specific trust framework. Their interpretation is that trust is not merely a pre-existing personality trait but something that can be cultivated through structured pedagogical encounters with AI. When instructors explained how Gemini worked, how prompts shaped outputs and how data were protected, students appeared to calibrate their reliance on the tool rather than either rejecting it or deferring to it uncritically. The nonsignificant gender effects further suggest that socially formed dispositions did not uniformly determine outcomes in this sample.

The researchers are candid about the limits of their evidence. Intact classes, non-random assignment, different institutions and facilitators, a small qualitative sample of volunteers, and self-reported trust all constrain causal claims, and the study does not establish that problem-based learning is superior to lecturing or that Gemini alone produced the gains. All four quoted interviewees were favourable, and one closed-ended ethical concern went unexplored. Still, for educators and policymakers weighing whether generative AI belongs in formal curricula, the study offers a compelling template: embed the tool inside active, problem-centred pedagogy, require students to reason first, verify AI output against course materials, and make transparency and accountability explicit. Under those conditions, the authors conclude, carefully structured AI support may meaningfully complement programming education, and trust in the machine may be something teachers can actually teach.

Subject of Research: The effect of AI-integrated problem-based learning on programming achievement and student trust in artificial intelligence in higher education

Article Title: Exploring trust in artificial intelligence in a problem-based computer programming course

Article References: Omeh, C. B., Ezeamuzie, N. O., & Olelewe, C. J. (2026). Exploring trust in artificial intelligence in a problem-based computer programming course. Discover Education, 5(1), Article 1172. https://doi.org/10.1007/s44217-026-02200-7

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02200-7

Keywords: generative AI, Google Gemini, trust in AI, problem-based learning, programming education, computer science education, quasi-experimental study, Nigeria, higher education, gender differences, educational technology, achievement