The development of Artificial Intelligence (AI) has transformed the digital education landscape through adaptive learning systems, learning analytics, and automated data-driven feedback ( Fombona et al., 2025; Hariyanto et al., 2025). In higher education, AI is being utilized to analyze student learning patterns, predict academic achievement, and recommend personalized learning resources ( Long et al., 2026). In the context of mathematics education, AI expands the learning experience through interactive visualizations, spatial transformation simulations, and real-time feedback that supports conceptual exploration ( Lin & Jiang, 2025; Yoon et al., 2024). However, optimism about AI in education is not entirely without criticism. Several studies have shown that technology integration that is not designed on sound pedagogical principles can increase cognitive load ( Gkintoni et al., 2025), visual distraction, and procedural dependency, thereby not always contributing to deep conceptual understanding. Furthermore, most research still focuses on the general effectiveness of AI on engagement or academic performance ( Singh et al., 2026), while studies specifically linking AI to the development of higher-order cognitive skills—such as spatial literacy and understanding of geometric concepts—are still limited. Thus, an approach is needed that not only integrates AI as a technological tool but also systematically orchestrates the interactions among technology, pedagogy, and content to support meaningful knowledge construction.

Spatial literacy refers to an individual’s capacity to visualize, manipulate, reason, and communicate spatial relationships between objects and their representations ( Newcombe & Shipley, 2015; Sorby et al., 2018). In geometry learning, spatial literacy serves as a cognitive foundation for understanding the structure, properties, and transformations of shapes relationally ( Pujawan et al., 2020). Conceptually, spatial literacy encompasses three main dimensions: (1) visualization, the ability to perform mental rotations and transformations; (2) reasoning, the ability to establish logical relationships related to position and dimension; and (3) communication, the ability to represent ideas through symbols, diagrams, or concrete models ( Uttal & Cohen, 2012). Empirical research shows a positive correlation between spatial skills and academic achievement in geometry ( Atit et al., 2022; Young et al., 2018). However, spatial literacy is not synonymous with understanding geometric concepts. Conceptual understanding involves the ability to coherently link definitions, properties, representations, and relationships between concepts, rather than simply performing visual manipulations. Some student teachers demonstrate adequate procedural skills but still struggle to explain the mathematical reasoning behind certain geometric transformations or relationships. It indicates a gap between visual-spatial abilities and in-depth conceptual understanding. For student teachers, mastery of spatial literacy and understanding of geometric concepts are crucial, as both influence not only their personal academic performance but also the quality of the representations and explanations they will provide to students in the future. Therefore, learning interventions at the teacher education level require designs that simultaneously strengthen both the visual-spatial and conceptual dimensions.

Although studies on spatial aspects in mathematics education continue to grow, research focused on empirical interventions remains relatively limited. Of the 652 Scopus-indexed documents related to spatial aspects, only a small fraction explicitly highlight interventions to improve spatial reasoning ( Palupi et al., 2023). This dominance of conceptual approaches indicates a gap between theoretical elaboration and pedagogical implementation. The development of AI-based digital textbooks offers an opportunity to bridge this gap. Unlike conventional e-books, AI-based systems can provide dynamic visualizations, spatial transformation simulations, and adaptive feedback based on student responses. The integration of abstraction theory with the Technological Pedagogical Content Knowledge (TPACK) framework ( Sahrudin et al., 2025) provides a conceptual foundation for designing modules that not only present geometric content but also manage pedagogical and technological interactions in an integrated manner. The effectiveness of AI-based digital textbooks in improving spatial literacy and understanding of geometric concepts has not been widely tested empirically, particularly among student teachers. Some studies have focused more on its impact on engagement, self-efficacy, and learning satisfaction ( Anierobi et al., 2025; Gerlich, 2025; M. Zhou & Peng, 2025), while its direct impact on conceptual understanding and spatial reasoning still requires further investigation. Furthermore, while dynamic visualizations can help reduce cognitive load in students with low spatial abilities ( Sitompul et al., 2026), excessive adaptive features may actually reduce the need for independent mental elaboration. Therefore, the design of AI-based digital textbooks must strike a balance between scaffolding and independent exploration to encourage the internalization of concepts truly.

This study used a paired-samples design (pretest-posttest), in which each respondent served as their own control. Each participant underwent measurements before and after the intervention, which involved using an AI-based digital textbook for geometry learning. This approach enables direct analysis of individual performance changes and the assessment of learning improvements resulting from the intervention under identical teaching conditions, regardless of class size.

Respondents completed a seven-week intervention with an AI-based digital textbook designed to enhance spatial literacy. The textbook featured interactive 3D visualizations, adaptive feedback, and personalized task recommendations, focusing on spatial figures such as cubes, cuboids, prisms, and pyramids. Before being used in the learning process, the digital textbook was developed and validated by experts to ensure content suitability and pedagogical feasibility. Once validated, participants were given a pre-test to assess their initial spatial intelligence abilities.

During the seven-week intervention, students learned using the digital textbook according to a predetermined schedule. This learning process allowed them to interact with 3D models, receive real-time adaptive feedback, and complete tasks tailored to their individual abilities. Following the completion of the learning process, students were given a post-test to assess the effectiveness of the digital textbook in improving spatial literacy. Pre-test and post-test results were compared, and learning gains were analyzed using N-gain to quantitatively evaluate the intervention’s impact on students’ spatial abilities.

The respondents in the study consisted of 45 prospective teacher students (M ± SD = 20.0 ± 0.9) enrolled in the Geometry and Measurement course in the second semester of the Primary Teacher Education Study Program, Faculty of Teacher Training and Education, Universitas Muhammadiyah Prof. Dr. Hamka, Special Capital Region of Jakarta, Indonesia. There were 5 men (11.1%) and 40 women (88.9%). All students (respondents) provided written consent after being informed of the study’s objectives and procedures.

The main instrument used in this study was a spatial literacy test comprising three dimensions, adapted from established frameworks and aligned with the course syllabus. The first dimension is visualization (cognitive levels of understanding, applying, and analyzing), which measures students’ ability to represent objects spatially. For example, students are asked: “Draw a three-dimensional sketch of a cube with a side length of 5 cm and label each side.”

The second dimension is reasoning (cognitive level of evaluating), which assesses students’ ability to interpret and evaluate geometric relationships. One question uses a triangular prism: students are asked to create a spatial model of the prism, explain the relationship between the base and the perpendicular edge, relate the prism’s height to its volume and everyday applications, calculate the volume, and summarize their findings.

The third dimension is communication (cognitive level of creating), which emphasizes students’ ability to convey mathematical arguments logically and visually. For example, students respond to the statement that “The volume of a pyramid is always half the volume of a cube because it appears smaller” by presenting mathematical calculations, sketching a supporting spatial model, and explaining the rationale for their choice of visual representation (see Table 1).

Data analysis began with descriptive analysis, including calculating the mean and standard deviation of pre-test and post-test scores, as well as N-gain, to obtain an initial overview of the data distribution and variability. Normality test results indicated that the pre-test and post-test scores did not follow a normal distribution, as indicated by the Kolmogorov-Smirnov and Shapiro-Wilk tests (p < 0.05; see Table 2). Therefore, the non-parametric Wilcoxon signed-rank test was applied to evaluate significant differences in students’ spatial literacy scores before and after the intervention.

Next, N-gain analysis was conducted using the formula of pre-test score-pre-test score divided by ideal score-pre-test score with the following criteria: a score ≥ 0.70 is categorized as high, 0.30–0.69 as moderate, and < 0.30 as low ( Blegur, Ma’mun, Berliana, Mahendra, Bakhri, et al., 2024b; Hake, 1999), to evaluate the effectiveness of geometric learning using AI-based digital books to improve students’ spatial literacy. This approach integrates descriptive, inferential, and relative perspectives, allowing a comprehensive interpretation of learning improvements.

Descriptive analysis

Figure 1 visualizes the analysis results of 15 questions. All post-test scores were higher than pre-test scores, indicating an increase in student achievement after the intervention. The most significant increases occurred in question 12 (Δ = 124), which measures evaluating solutions based on logical and systematic procedures; question 11 (Δ = 83), which measures calculating the volume of spatial objects—both within the reasoning dimension; and question 15 (Δ = 107), which measures providing spatial arguments based on the analysis of geometric problems within the communication dimension. This pattern indicates that the intervention had a stronger impact on higher-order cognitive aspects, particularly on items with relatively low initial scores.

Conversely, the smallest increases were found in questions 2 (Δ = 10) and 1 (Δ = 25), which measure the ability to visualize three-dimensional objects and identify geometric shapes in various orientations, respectively, within the visualization dimension. Both items had relatively high pre-test scores, thus limiting the scope for improvement (a ceiling effect). In aggregate, the average score increased from 92.9 to 164.2, with an average difference of 71.3 points per question. These results empirically support the effectiveness of AI-based digital textbooks in improving students’ spatial literacy in geometry learning, particularly in previously more challenging aspects of mathematical reasoning and communication.

A more detailed analysis of the visualization ability dimension revealed that all indicators improved from pre-test to post-test. The greatest improvement occurred in converting two-dimensional images to three-dimensional models and vice versa (ΔM = 0.89), followed by using mathematical representations to present spatial information (ΔM = 0.68), identifying geometric shapes in various orientations (ΔM = 0.54), and visualizing objects in three dimensions (ΔM = 0.22). The decrease in standard deviation across all indicators indicates that the intervention not only increased the average score but also more evenly distributed students’ abilities.

The Wilcoxon signed-rank test showed a significant difference between pre-test and post-test scores on three visualization indicators: the ability to identify geometric shapes (Z = −2.807, p = 0.005), the ability to convert 2D-3D (Z = −4.729, p < 0.001), and the use of spatial representations (Z = −4.205, p < 0.001). However, the improvement in three-dimensional object visualization was not statistically significant (Z = −1.768, p = 0.077). The effect sizes showed a large effect on 2D-3D conversion ability (r = 0.70) and use of spatial representation (r = 0.63), a medium effect on geometric shape identification (r = 0.42), and a small effect on visualization of three-dimensional objects (r = 0.26) (see Table 3). These findings indicate that the intervention was more effective at improving transformation and spatial representation abilities than at improving direct 3D object visualization abilities.

Furthermore, in the reasoning dimension, all indicators showed substantial improvement from pre-test to post-test. The greatest improvement occurred in evaluating solutions based on logical and systematic procedures (ΔM = 2.75), followed by developing problem-solving strategies in everyday life (ΔM = 2.05), calculating the volume of spatial objects (ΔM = 1.85), and understanding the relationship between shape, size, position, and orientation in space (ΔM = 1.60). The decrease in standard deviation across all indicators indicates that students’ abilities became more evenly distributed after the intervention, thereby increasing average achievement and reducing inter-individual variation.

The Wilcoxon signed-rank test results showed that all mathematical reasoning indicators improved significantly between the pre-test and post-test (p < 0.001). The effect size was in the very large category (r = 0.84–0.87), with the highest Z value for the logical and systematic solution evaluation indicator (Z = −5.849, r = 0.87) (see Table 4). Improvements were also seen in the development of problem-solving strategies, the calculation of spatial object volumes, and the understanding of spatial relationships. Descriptively, the decrease in standard deviations across all indicators indicates a more even distribution of abilities after the intervention. These findings indicate that the applied learning is associated with substantial improvements in mathematical reasoning, especially in higher-order thinking skills.

Furthermore, in the communication dimension, all indicators of mathematical communication skills showed a clear improvement from pre-test to post-test. The greatest improvement occurred in providing spatial arguments based on the results of geometric problem analysis (ΔM = 2.38), followed by using mathematical language to explain spatial problem solutions (ΔM = 1.85), and presenting spatial understanding through drawings, sketches, or written descriptions (ΔM = 1.82). The low pre-test mean score, particularly for spatial argumentation (M = 0.60), indicates that, before the intervention, students were still weak at formally communicating geometric reasoning results. The decrease in the post-test standard deviation indicates that students’ communication skills improved more evenly after the intervention.

The Wilcoxon signed-rank test results showed that all indicators of mathematical communication improved significantly between the pre-test and post-test (p < 0.001). The effect size was in the very large category (r = 0.76–0.82), with the highest effect on the ability to provide spatial arguments (Z = −5.493, r = 0.82) (see Table 5). Substantial improvements were also seen in the use of mathematical language and the ability to present spatial understanding. These findings indicate that implementing AI-based digital textbook learning is associated with significant strengthening of aspects of mathematical communication, particularly in argumentative ability, a key component of higher-order mathematical thinking.

Wilcoxon signed-rank

Based on an analysis of 45 students, all dimensions of spatial literacy increased significantly from pre-test to post-test. The visualization dimension increased from 3.12 ± 1.33 to 3.77 ± 0.76 (Z = −4.969; p < 0.001; r = 0.74), reasoning from 1.68 ± 1.22 to 3.67 ± 0.79 (Z = −5.419; p < 0.001; r = 0.81), and communication from 1.21 ± 1.28 to 3.22 ± 1.03 (Z = −5.756; p < 0.001; r = 0.86). The total spatial literacy score increased from 2.26 ± 1.52 to 3.63 ± 0.85 (Z = −5.780; p < 0.001; r = 0.86) (see Table 6), confirming that the intervention had a large effect both statistically and practically. These findings indicate that the intervention was effective in improving students’ multidimensional spatial literacy, especially in spatial reasoning and communication, which represent higher-order thinking skills.

N-gain analysis

The N-gain analysis results demonstrated that all spatial literacy dimensions were in the high category (g ≥ 0.70), with an average N-gain of 0.78 (77.77%), indicating effectiveness. The largest increase occurred in the reasoning dimension (N-gain = 0.86; 85.82%), followed by visualization (0.74; 74.37%) and communication (0.72; 72.15%). In aggregate, the total spatial literacy score increased from 1524 to 2450, from an ideal score of 2700, resulting in an N-gain of 0.79 (78.74%) (see Table 7). These results indicate that the intervention was not only statistically significant but also highly effective in practice, as measured by the normalized gain.

Scientifically, an N-gain value in the range of 0.70–0.86 indicates that most of the potential improvement in the score was achieved after the intervention. The dominant improvement in the reasoning dimension confirms that the intervention was very effective in developing higher-order thinking skills, such as spatial analysis and problem-solving. Meanwhile, the communication dimension, although slightly lower, remained in the high category, indicating significant progress in students’ ability to articulate spatial ideas. Thus, the intervention is very effective in comprehensively improving students’ spatial literacy.