AI Chatbot Boosts Financial Literacy in University Students, Pilot Study Finds

A one-week conversation with a carefully constrained artificial intelligence chatbot appears to have substantially improved university students’ financial literacy, according to a new pilot study published in Discover Education. Researchers at Bayburt University in Türkiye built a tutoring system on top of OpenAI’s GPT-4o model and asked 32 undergraduates to learn with it freely for seven days. When students were tested before and after the learning period, their overall financial literacy knowledge scores climbed from an average of 4.33 to 5.00 on a six-point scale, a change the authors describe as a large standardized effect, with a Cohen’s d of 1.09 and a 95 percent confidence interval running from 0.64 to 1.52. Every one of the six knowledge domains measured in the study showed a statistically significant gain after the researchers applied a Holm correction to guard against false positives across multiple comparisons.

The most eye-catching result concerned compound interest, one of the foundational concepts of personal finance and a topic that decades of research show people worldwide understand poorly. On an objective multiple-choice item requiring students to convert a monthly interest rate into its annual equivalent, the correct-response rate exploded from just 3.1 percent before the learning period to 81.2 percent afterward, a jump of more than 78 percentage points. Performance also improved on items covering simple interest, early savings growth, and budgeting, while an exchange-rate question remained flat at 65.6 percent. Because the same five objective items were administered twice, the authors caution that practice effects could account for part of the improvement, but the sheer scale of the compound-interest gain suggests that something more than mere familiarity with the test format was at work.

The chatbot’s design is arguably as interesting as its results. Rather than giving the model free rein, the researchers built it as a web application on the GPT-4o API and governed its behavior entirely through a structured system prompt, without retrieval-augmented generation or an external knowledge base. The system prompt encoded four pedagogical principles. First, scope restriction: the bot could discuss only financial literacy topics within the six measurement domains and was explicitly forbidden from giving personalized investment advice, real-time market data, or speculative forecasts. Second, pedagogical scaffolding: it was configured as a step-by-step learning assistant that answered every question through a fixed sequence of conceptual explanation, worked example, and reinforcement question. Third, safety redirects: questions about current tax law or specific investment products were referred to authoritative institutions such as Türkiye’s Capital Markets Board and Revenue Administration. Fourth, linguistic accessibility: all responses were delivered in plain Turkish with technical terms defined inline.

A documented example exchange shows the scaffold in action. When a student asked what compound interest means, the chatbot first explained the concept, noting that it means earning interest on both the principal and previously accumulated interest. It then offered a worked example: depositing 1,000 Turkish lira at 10 percent annual interest compounded monthly would yield roughly 1,105 lira after one year rather than 1,100. Finally, it posed a reinforcement question, asking the student to calculate what 500 lira would become after two years at the same rate. This explain-exemplify-elicit structure was written into the system prompt as the default response pattern, translating three decades of learning science into conversational practice.

That structure draws deliberately on three theoretical pillars. Constructivist learning theory, rooted in the work of Piaget and Vygotsky, holds that learners build knowledge actively rather than passively absorbing it, and the chatbot’s scaffolding was designed to operate within the learner’s zone of proximal development, the space between what a student can do alone and what they can achieve with guidance. The immediate feedback model, articulated by Hattie and Timperley, supplies the rationale for rapid, task-relevant corrective responses, which matter enormously in a domain like finance where a single misunderstood calculation can cascade. Self-determination theory, from Ryan and Deci, motivated the decision to let students direct their own questioning during the learning week, supporting the psychological needs for autonomy and perceived competence. The authors are careful to frame these theories as design rationales rather than tested causal mechanisms, since the study did not measure how students actually used the scaffolding.

The measurement itself was unusually detailed. The primary instrument comprised 32 factual financial statements, eleven of them deliberately worded as misconceptions, which students rated on a six-point agreement scale before being keyed against correct answers. Items were organized into six subscales: Basic Economics and Finance, Individual Banking, Retirement and Insurance, Financial Statements, Investment, and Tax and Legislation. Five objective multiple-choice questions on exchange rates and financial mathematics served as a complementary performance check. The researchers chose paired-samples t-tests for subscales whose difference scores passed Shapiro-Wilk normality tests and Wilcoxon signed-rank tests for those that did not, reporting bootstrap confidence intervals for the rank-based effects. The largest subscale gains appeared in Financial Statements, with a mean increase of 1.08 points and an effect size of 0.95, and Retirement and Insurance, which rose 0.83 points with an effect of 0.91, while Individual Banking showed the smallest change at 0.34 points, plausibly because students started from a high baseline of familiarity with internet banking and credit cards.

Not every data point moved in the expected direction, and one reversal is particularly instructive. Agreement with the false statement that inflation in Türkiye is below 10 percent actually rose during the study, producing the largest single-item change in the entire instrument once the item was reverse-coded, a decline of 1.50 points. Thirteen students answered this item less accurately at post-test, sixteen were unchanged, and only three improved. The authors offer a telling explanation: the chatbot was deliberately configured without access to real-time market data and instructed to redirect questions about current economic conditions to authoritative sources, so it could explain what inflation is but could not state what the current rate actually is. The decline on this item is therefore consistent with the system’s scoping rather than contradicting it, though the researchers acknowledge that students may also have generalized from worked examples that used round illustrative rates, or that the item itself may simply be unstable.

The study’s honesty about its own limits is notable. With no control or comparison group, the observed changes cannot be causally attributed to the chatbot; testing effects, maturation, concurrent coursework, novelty, and regression to the mean all remain plausible contributors. Interaction logs were never systematically recorded, so session duration, question counts, and topic coverage cannot be reconstructed. Several subscales showed weak or even negative pre-test internal consistency, with Cronbach’s alpha values as low as minus 0.22 for the Investment subscale, meaning dimension-specific findings must be treated as exploratory. The agreement-scale format also conflates knowledge with confidence, and the small volunteer sample of 32 students from a single university constrains generalizability. The authors explicitly position the work as a design and feasibility record rather than evidence of effectiveness, and they lay out a roadmap for follow-up research: randomized controlled trials, verified knowledge sources such as retrieval-augmented generation for regulation-sensitive content, systematic privacy-aware logging, validated instruments with established factor structures, delayed post-tests to separate practice effects from retained learning, and outcome measures that capture actual financial behavior rather than agreement-rated knowledge alone.

Even with those caveats, the findings land at a moment when both educators and technologists are searching for scalable answers to a persistent problem. Financial literacy is strongly linked to retirement planning, wealth accumulation, and resistance to fraud, yet surveys consistently show that even basic understanding of compound interest, inflation, and risk diversification remains shockingly low worldwide. University students face a particularly acute transition, juggling credit cards, student loans, and housing costs as they approach financial independence, while traditional instruction cannot offer continuous, individualized explanation on demand. A scaffolded chatbot can, at low marginal cost. The Bayburt team’s work suggests that the recipe matters as much as the technology: constrain the scope, forbid personalized advice, route regulation-sensitive questions to authoritative institutions, and build the pedagogy directly into the prompt. Meta-analyses of educational chatbots report modest pooled effects, with a 2026 synthesis of 35 experimental studies finding a moderate effect of ChatGPT on learning outcomes, so the large uncontrolled changes observed here demand cautious interpretation. But as a blueprint for how large language models might be tamed into patient, safe, and pedagogically disciplined financial tutors, this pilot offers one of the clearest demonstrations yet, and a compelling reason for controlled trials to follow.

Subject of Research: A pilot pre-post study of a scaffolded GPT-4o AI chatbot for financial literacy education among university students

Article Title: A pilot pre–post study of a scaffolded AI chatbot for financial literacy education among university students

Article References: Ayberkin, D., & Yirik, E. (2026). A pilot pre–post study of a scaffolded AI chatbot for financial literacy education among university students. Discover Education, 5(1), Article 1068. https://doi.org/10.1007/s44217-026-02226-x

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02226-x

Keywords: financial literacy, AI chatbot, GPT-4o, large language models, university students, educational technology, scaffolding, pre-post study, compound interest, Türkiye, higher education, pilot study