July 20, 2026

AI adaptive learning doubled programming gains for disadvantaged students, Hungarian study finds

A 13-week study at four Hungarian universities found AI adaptive learning doubled programming gains for disadvantaged students (effect size 1.40).

AI adaptive learning doubled programming gains for disadvantaged students, Hungarian study finds

A controlled study at four Hungarian universities found that AI adaptive learning more than doubled the programming gains of socially disadvantaged students compared with conventional teaching. The research, published in the journal “TechTrends” by József Katona and Klára Ida Katonáné Gyönyörű of the University of Dunaújváros, followed 122 students over 13 weeks and reported a large effect on test scores and engagement through AI adaptive learning.

What the study tested

The team identified 122 socially disadvantaged students, aged 18 to 21, using a 15-item “Social Condition Index,” then randomly assigned them to an experimental group of 61 and a control group of 61.

The experimental group learned programming through an AI adaptive learning system built with ChatGPT that provided real-time feedback, used Bayesian Knowledge Tracing to model each learner’s knowledge, and adjusted task difficulty and content to the individual.

The control group covered the same curriculum – control structures, functions, loops, and algorithms – through traditional, non-adaptive teaching, with the same instructor, laboratories, hardware, and internet access.

AI adaptive learning results

On the post-test, the experimental group’s mean score rose by 2.0 points (from 5.1 to 7.1), while the control group rose by 1.0 point (from 5.0 to 6.0). Within-group effect sizes reached Cohen’s d of 1.40 for the experimental group and 0.74 for the control group, both at p < 0.001.

GroupPre-test meanPost-test meanGainLearning effect size (Cohen’s d)
AI adaptive — experimental (n = 61)5.17.1+2.01.40 (large)
Traditional — control (n = 61)5.06.0+1.00.74 (moderate)

Post-test results. Source: Katona & Katonáné Gyönyörű (2025), “TechTrends”, DOI 10.1007/s11528-025-01088-8.

The authors wrote that “The students in the EG learned much better (p < 0.001) with a large effect size of 1.40, while the CG learned moderately better.” The experimental group also reported higher engagement across behavioural (d = 1.10), emotional (d = 1.00), and cognitive (d = 1.40) measures, each at p < 0.001. An analysis of covariance found the AI advantage held independent of students’ socio-economic background, and the engagement questionnaire returned a Cronbach’s alpha of 0.89.

Effect sizes (Cohen’s d) for AI adaptive learning versus traditional teaching Effect sizes (Cohen’s d): AI adaptive vs traditional teaching 0 0.5 1.0 1.5 Learning — AI adaptive (EG) 1.40 Learning — traditional (CG) 0.74 Engagement — behavioural 1.10 Engagement — emotional 1.00 Engagement — cognitive 1.40
Cohen’s d by measure. Green and grey: within-group learning gains; teal: between-group engagement (experimental vs control). Source: Katona & Katonáné Gyönyörű (2025), “TechTrends”.

Stated limitations

The authors note the 13-week duration limits any conclusion about long-term retention and call for multi-semester studies. They also flag potential selection bias, because participants were recruited through survey responses rather than full random sampling of the student population.

Background on AI Adaptive learning systems

Adaptive learning systems use data on each learner’s responses to adjust pace, difficulty, and content in real time, an approach often paired with techniques such as Bayesian Knowledge Tracing. Interest has grown as schools and universities test whether such tools can narrow gaps for students with fewer resources rather than widen them.

Winss Solutions has in the past tracked this question through coverage of AI + Learning Differences, a blueprint for inclusive learning without boundaries and of policies for the digital transformation of school education set out by the OECD, which examines the digital divide in schools.

Commercial tools have moved in the same direction, including Google’s “Learn Your Way,” which generates AI-augmented textbooks for personalized learning. The Hungarian experiment, published in the journal “TechTrends”, adds a controlled measurement to that record by testing the approach specifically with socially disadvantaged students and reporting the size of the effect.


Become a Sponsor

Our website is the heart of the mission of WINSS – it’s where we share updates, publish research, highlight community impact, and connect with supporters around the world. To keep this essential platform running, updated, and accessible, we rely on the generosity of you, who believe in our work.

We offer the option to sponsor monthly, or just once choosing the amount of your choice. If you run a company, please contact us via info@winssolutions.org.

Select a Donation Option (USD)

Enter Donation Amount (USD)

What do you feel about this?

AI adaptive learning doubled programming gains for disadvantaged students, Hungarian study finds

A 13-week study at four Hungarian universities found AI adaptive learning doubled programming gains for disadvantaged students (effect size 1.40).

A controlled study at four Hungarian universities found that AI adaptive learning more than doubled the programming gains of socially disadvantaged students compared with conventional teaching. The research, published in the journal “TechTrends” by József Katona and Klára Ida Katonáné Gyönyörű of the University of Dunaújváros, followed 122 students over 13 weeks and reported a large effect on test scores and engagement.

What the study tested

The team identified 122 socially disadvantaged students, aged 18 to 21, using a 15-item “Social Condition Index,” then randomly assigned them to an experimental group of 61 and a control group of 61. The experimental group learned programming through an AI-based adaptive system built with ChatGPT that provided real-time feedback, used Bayesian Knowledge Tracing to model each learner’s knowledge, and adjusted task difficulty and content to the individual. The control group covered the same curriculum — control structures, functions, loops, and algorithms — through traditional, non-adaptive teaching, with the same instructor, laboratories, hardware, and internet access.

AI adaptive learning results

On the post-test, the experimental group’s mean score rose by 2.0 points (from 5.1 to 7.1), while the control group rose by 1.0 point (from 5.0 to 6.0). Within-group effect sizes reached Cohen’s d of 1.40 for the experimental group and 0.74 for the control group, both at p < 0.001.

Group Pre-test mean Post-test mean Gain Learning effect size (Cohen’s d)
AI adaptive — experimental (n = 61) 5.1 7.1 +2.0 1.40 (large)
Traditional — control (n = 61) 5.0 6.0 +1.0 0.74 (moderate)

Post-test results. Source: Katona & Katonáné Gyönyörű (2025), “TechTrends”, DOI 10.1007/s11528-025-01088-8.

The authors wrote that “The students in the EG learned much better (p < 0.001) with a large effect size of 1.40, while the CG learned moderately better.” The experimental group also reported higher engagement across behavioural (d = 1.10), emotional (d = 1.00), and cognitive (d = 1.40) measures, each at p < 0.001. An analysis of covariance found the AI advantage held independent of students’ socio-economic background, and the engagement questionnaire returned a Cronbach’s alpha of 0.89.

Effect sizes (Cohen’s d) for AI adaptive learning versus traditional teaching Effect sizes (Cohen’s d): AI adaptive vs traditional teaching 0 0.5 1.0 1.5 Learning — AI adaptive (EG) 1.40 Learning — traditional (CG) 0.74 Engagement — behavioural 1.10 Engagement — emotional 1.00 Engagement — cognitive 1.40
Cohen’s d by measure. Green and grey: within-group learning gains; teal: between-group engagement (experimental vs control). Source: Katona & Katonáné Gyönyörű (2025), “TechTrends”.

Stated limitations

The authors note the 13-week duration limits any conclusion about long-term retention and call for multi-semester studies. They also flag potential selection bias, because participants were recruited through survey responses rather than full random sampling of the student population.

Background

Adaptive learning systems use data on each learner’s responses to adjust pace, difficulty, and content in real time, an approach often paired with techniques such as Bayesian Knowledge Tracing. Interest has grown as schools and universities test whether such tools can narrow gaps for students with fewer resources rather than widen them. Winss Solutions has tracked this question through coverage of AI + Learning Differences, a blueprint for inclusive learning without boundaries and of policies for the digital transformation of school education set out by the OECD, which examines the digital divide in schools.

Commercial tools have moved in the same direction, including Google’s “Learn Your Way,” which generates AI-augmented textbooks for personalized learning. The Hungarian experiment, published in the journal “TechTrends”, adds a controlled measurement to that record by testing the approach specifically with socially disadvantaged students and reporting the size of the effect.


Source: TechTrends (Springer)

What do you feel about this?