Generative AI-supported instruction as a cognitive scaffold: effects on senior high school students' geometric reasoning and proof construction

Synthesis: This quasi-experimental pre-test/post-test study with 86 senior high school students in Ghana tested generative AI as a cognitive and representational scaffold for geometry, built to prompt and question rather than solve. The experimental group outscored the control on a geometry reasoning and proof test, and the advantage held after adjusting for baseline performance. Visualization alone does not build proof skill, so the paper positions AI as the bridge from visual intuition to deductive argument, reading the result through Van Hiele's levels and Cognitive Load Theory. Perceptions of Feedback and explanation quality were favorable, but only the experimental group was surveyed and the paper never names the GenAI model or the data-collection window, so read it as evidence that a well-scaffolded mathematics intervention worked, not that a particular tool did.

Key Findings

1. Quasi-experimental pre-test/post-test nonequivalent control group design: 86 students from two intact Ghanaian senior high school classes, 43 per group, from a population of 558.
2. The experimental group rose from 12.37 (SD = 3.04) to 21.74 (SD = 4.42) on the Geometry Reasoning and Proof Test, t(42) = −18.410, p < .001; the control group rose from 13.21 (SD = 2.82) to 16.79 (SD = 4.76), t(42) = −5.843, p < .001.
3. Post-test scores differed significantly between groups, t(84) = −4.999, p < .001, d ≈ 1.05.
4. The ANCOVA controlling for pre-test scores showed a significant effect of instructional group, F(1, 83) = 50.30, p < .001, partial η² = .377, with the model explaining approximately 50.3% of post-test variance (R² = .503, Adjusted R² = .491).
5. The groups did not differ significantly at pre-test, t(84) = 1.325, p = .189, supporting baseline comparability of the intact classes.
6. Experimental-group students reported favorable perceptions (overall mean 4.10 on a 5-point scale, items 3.81–4.23), highest for feedback ("The feedback provided by generative AI helps me correct my mistakes," M = 4.23, SD = .812), lowest for explaining answers (M = 3.81, SD = 1.180).
7. Attitudes were similarly positive (overall mean 4.13, items 3.91–4.47); strongest was that using generative AI during geometry lessons improves understanding (M = 4.47, SD = .631), lowest was confidence in using technological tools (M = 3.91, SD = .840).

Design and the instructional intervention

Sample size came from a G*Power analysis (two-tailed d = 0.80, α = 0.05, power 0.80, 1:1 allocation), yielding 86 students selected by convenience sampling. Measurement used the Geometry Reasoning and Proof Test (GRPT), with multiple-choice and open-ended items, plus a five-point Likert perception and attitude questionnaire; both groups shared one teacher, the same content, and equivalent instructional time. A pilot with 30 excluded senior high school students established instrument quality: Spearman-Brown split-half reliability 0.859 for the GRPT and Cronbach's alpha of 0.878 (perception) and 0.914 (attitude).

Generative AI was positioned as a cognitive and representational scaffold, explicitly not a problem solver: it generated dynamic representations and guided prompts that pushed students to justify the geometric relationships they saw, identify relevant theorems, and order logical statements; it would not supply complete proofs or final answers, and its outputs drove teacher-led discussion aimed at independent thinking. This maps onto the study's two anchors: Van Hiele's progression from visual recognition toward formal deduction, and Cognitive Load Theory's argument that offloading diagram interpretation frees working memory for theorem selection and proof sequencing.

Results: performance, then perceptions

Levene's test satisfied equality of variances (pre-test p = .640, post-test p = .769), permitting equal-variance t-tests, and normality held at pre-test; the control group's failed Shapiro–Wilk post-test (p = .025) was judged tolerable given roughly equal groups above 30 and a consistent bias-corrected bootstrap. Analysis used IBM SPSS Statistics Version 27.

The performance gap is the study's core learning gain: after adjustment for prior performance in the ANCOVA, group membership remained a strong predictor of post-test score (F(1, 83) = 50.30, p < .001), and pre-test score was itself a significant covariate (F(1, 83) = 45.72, p < .001, partial η² = .355).

Perceptions and attitudes came from descriptive statistics and only from the experimental group: Engagement-adjacent items scored highest — useful Feedback, clear explanations, and step-by-step guidance.

Reading the result against the wider literature

The study joins work showing generative AI improves problem-solving outcomes when it supports reasoning rather than supplying answers; its contribution is the proof-construction target. The paper's review argues that technology-supported visualization tools — dynamic geometry environments, augmented reality, 3D printing — build spatial and conceptual understanding but do not, on their own, establish proof; visualization advances reasoning only when paired with purposeful instructional scaffolding. Positioning dialogic, natural-language AI as that scaffolding is what makes this study more specific than "AI improves math achievement."

What this means for practice

• Instructors. Treat the AI as a prompter, not an answer key: the intervention withheld complete proofs, leaving students to supply the theorem choice and logical order.
• Instructional designers. Design for the transition, not just the diagram: build prompts that demand justification of what a representation shows.
• Curriculum designers. Proof-oriented geometry objectives, not only spatial or conceptual ones, are a reasonable place to embed AI-supported activities.
• Teacher educators. Train on pedagogically appropriate use of generative AI as a scaffold for mathematical thinking, not a source of answers.
• Researchers. The study measured performance, not the theorized visualization and reasoning processes; direct measures of those are the next step.

Limitations

• Stated limits: a single-school quasi-experimental design constraining generalizability; the cognitive processes behind the improvement, especially visualization, went unmeasured; and perceptions and attitudes came only from the experimental group.
• The generative AI tool and version are never named and no data-collection window is given, so results cannot be tied to a particular model generation — a reproducibility gap.
• Reported pre-test means are inconsistent: the independent-samples table and text give the experimental group 13.21 (SD = 2.82) and the control group 12.37 (SD = 3.04), while the paired-samples table and text swap those values. The equivalence conclusion (p = .189) is unaffected, but the baselines should be read with caution.
• The design cannot rule out teacher, novelty, or selection effects, and the paper frames the findings as instructional effectiveness, not causal evidence.

Citation

Davor, I. (2026). Generative AI-supported instruction as a cognitive scaffold: Effects on senior high school students' geometric reasoning and proof construction. Frontiers in Education, 11, 1869961.