Teacher-Guided AI-Supported Digital Ecosystem Learning in Primary Science: A Quasi-Experimental Study

Abstract

Despite the growing use of artificial intelligence (AI) in science education, little is known about the motivational value of AI-supported digital tools in upper-primary ecosystem science learning. This quasi-experimental study examined whether participation in a teacher-guided digital ecosystems unit integrating AI-supported elements and interactive non-AI tools was associated with sixth-grade students’ ecosystem achievement, interest in science, attitudes toward science, and science self-efficacy. Four sixth-grade classes in an Israeli elementary school (123 students) participated. The experimental group completed six 45-min lessons; the control group studied the same content without the AI-supported components and integrated digital sequence. Students completed pretest and posttest questionnaires and an ecosystem achievement test; the experimental group also completed a satisfaction questionnaire. Semi-structured interviews were conducted with 10 students from the experimental group. Baseline-adjusted analyses indicated higher post-intervention achievement and motivational outcomes in the experimental group. Boys reported higher interest and self-efficacy than girls, and mothers’ education was positively associated with interest and attitudes. Within the experimental group, satisfaction was positively related to all motivational outcomes and significantly predicted self-efficacy. Interview themes highlighted visualization, feedback, collaboration, and occasional cognitive and technical challenges. Overall, this teacher-guided instructional package was associated with more favorable outcomes under classroom conditions in schools.

1. Introduction

Digitalized schooling has intensified interest in artificial intelligence (AI)-supported tools. In science education, studies have described the use of AI as a resource for visualization, inquiry support, differentiation, feedback, and multimodal learning. Recent reviews suggest that AI can enrich science learning when it is embedded in coherent pedagogy, although the evidence remains uneven across age groups, subjects, and outcome domains (Almasri, 2024; G. Lee et al., 2026). This is important in primary school, where students’ engagement with AI is affected by developmental readiness, prior digital experience, and teacher guidance (Feng & Carolus, 2026; Józsa et al., 2026). Warschauer et al. (2023) describe a “with or without” contradiction: AI tools can support learners, yet premature use may displace the development of foundational skills and undermine meaningful learning. This reinforces our emphasis on developmental readiness and the need for teacher-guided scaffolding that constrains AI use to complement—not replace—students’ own sensemaking and explanation.

Ecosystems provide an excellent context for examining these issues that require learners to reason about interdependence, adaptation, resource availability, and dynamic relations among biotic and abiotic components. Ecosystem learning thus appears well-suited to digitally supported instruction that can make ecological relationships more visible and discussable through simulations, visualizations, and interactive tasks (Hewitt et al., 2023; Hajj-Hassan et al., 2024). Nevertheless, studies on primary school students often report fragmented understanding and linear explanations rather than robust systems thinking, even when visual supports are available (Mambrey et al., 2022).

Primary science is also motivational. Interest in science, attitudes toward science, and science self-efficacy impact students’ engagement, persistence, and participation. However, beyond classroom experience, these outcomes are influenced by gender and family socio-economic status (Carroll et al., 2024; Bórquez-Sánchez, 2025). Research shows that well-designed teaching sequences can strengthen science self-efficacy as well as conceptual understanding (Giarratano et al., 2026). This is key for AI-supported learning in primary school, because students’ responses not only depend on technical access but also on cognitive, emotional, and motivational readiness (Feng & Carolus, 2026).

Despite the rapid growth of AI-in-education scholarship, empirical evidence on AI-supported science learning in primary schools remains limited, especially for ecosystem topics and motivational outcomes. Thus, little is known about how teacher-guided AI-supported and interactive digital environments shape upper-primary students’ interest in science, attitudes toward science, science self-efficacy, and their evaluation of the learning experience (Dong et al., 2025; Deng et al., 2025).

This quasi-experimental mixed-method study examined an AI-supported digital ecosystem unit implemented with sixth-grade students in Israel. The unit combined AI-enhanced tools, including Ruby Bot (version 1.7), ChatGPT (GPT-4o, OpenAI), and Gamma (Gamma 2.0), with interactive applications such as Padlet, Canva, Educaplay, My Story, and a PhET simulation in a structured sequence. The students exposed to this environment were compared to students who studied the same content without AI in a standard classroom setting. Post-intervention achievement, interest in science, attitudes toward science, science self-efficacy, associations with gender and parental education, and students’ satisfaction and learning experiences were examined. The findings contribute context-specific evidence to a better understanding of how an AI-supported, teacher-guided, and interactive digital ecosystems unit can impact upper-primary science learning and motivational engagement in real classroom conditions.

2. Literature Review

2.1. Ecosystem Learning in Primary Science

Understanding ecosystems requires students to connect organisms, habitats, needs, interactions, and adaptations across multiple levels of organization while also reasoning about changes over time and the relationships between biotic and abiotic factors. Although primary students can begin to develop such understandings, their reasoning often remains fragmented or dependent on surface features unless instruction provides strong conceptual and representational support (Mambrey et al., 2022).

Technology-enhanced environmental learning can externalize processes that are otherwise difficult to observe directly. Reviews show that virtual field trips, geospatial tools, augmented reality, and digital supports can strengthen environmental understanding and sustainability awareness, but they have also called for more school-based research on complex ecological systems (Hajj-Hassan et al., 2024). Studies focused more directly on ecosystems similarly indicate that immersive and interactive tools can make ecological relationships more concrete and meaningful for younger learners (Hewitt et al., 2023; Sachyani & Gal, 2025). In this context, the educational value of digital support lies not only in technological novelty but also in its capacity to make ecological relationships visible, discussable, and conceptually manageable for younger learners.

In this study, we conceptualize teacher-guided AI-supported digital ecosystem learning as a set of instructional affordances that combine (a) visualization and interactive representations, (b) immediate feedback and opportunities to revise explanations, and (c) structured collaboration and knowledge construction under explicit teacher scaffolding. Together, these affordances are expected to support ecosystem achievement by helping students organize causal relationships and test explanations, while also strengthening motivation by increasing situational interest, fostering more positive attitudes toward science, and building science self-efficacy through successful, supported learning experiences. Accordingly, we examine both achievement and motivational outcomes, and explore whether background characteristics are associated with these outcomes.

2.2. AI-Supported and Interactive Digital Science Learning

AI in education covers intelligent tutoring systems, adaptive platforms, learning analytics, chatbots, and generative tools that can produce text, images, and feedback. Recent reviews and meta-analyses generally report positive effects on learning, but also considerable variation by subject area, learner age, duration, and implementation quality (Deng et al., 2025). This strongly suggests that the educational value of AI depends on how it is embedded in instruction.

In science education, AI-supported tools are increasingly being used to scaffold inquiry, provide automated feedback, clarify concepts, and support discussion. Younger learners tend to approach AI through more familiar, conversational, or platform-based uses, whereas older students are more likely to use AI for academically oriented purposes (Józsa et al., 2026). For upper-primary ecosystem learning, the educational value of AI is likely to depend less on AI alone than on how AI-supported and non-AI digital tools are combined within a coherent pedagogy. This is consistent with work showing that science learning benefits when technology supports explanation, feedback, collaboration, and structured engagement with disciplinary ideas rather than as an isolated add-on (Chiu et al., 2024; Dávila-Acedo et al., 2022). For upper-primary ecosystem learning, the prime issue is how AI-supported and other digital tools can be pedagogically orchestrated to support visualization, feedback, collaboration, and explanation.

2.3. Motivational and Cognitive Outcomes and Background Factors

Interest in science, attitudes toward science, science self-efficacy, and academic achievement are especially important in upper-primary science because they influence students’ participation, persistence, and developing relationships with science. Academic achievement provides a complementary indicator of whether students acquire the targeted content. In primary science, affective and cognitive outcomes are closely intertwined. Research suggests that well-designed teaching sequences can strengthen science self-efficacy alongside conceptual understanding (Giarratano et al., 2026). Wan et al. (2024) found that attitudes toward science can mediate achievement-related benefits in inquiry learning. Bjerke et al. (2026) identified mastery experiences as key to the development of mathematics and science self-efficacy across school education.

These outcomes are not the result of instruction alone. Gender differences in science-related motivation are mixed and context-dependent, and family educational resources can influence students’ attitudes, confidence, and science-related aspirations (Bórquez-Sánchez, 2025; Guo et al., 2026; Pinneo & Nolen, 2024). Students’ satisfaction with the AI learning environment may also matter, because perceptions of clarity, support, relevance, and usability may be linked to more favorable motivational outcomes (Deng et al., 2025).

Effective engagement with AI in school science appears to depend not only on knowledge and skills, but also on learners’ self-efficacy, attitudes, motivation, emotional readiness, and metacognitive awareness (Feng & Carolus, 2026; Soriano-Sánchez et al., 2026).

2.4. Research Gap and Study Contribution

Taken together, prior research suggests that ecosystem learning is conceptually demanding and benefits from visual, interactive, and well-scaffolded instruction. However, it remains unclear how AI-supported and other technology-enhanced tools operate when integrated into a teacher-guided digital learning sequence in upper-primary science, particularly regarding motivational outcomes and background-related differences.

There is scant evidence on the ways in which AI-supported digital learning environments function in upper-primary ecosystem instruction when the focus is not only on understanding, but also on interest in science, attitudes toward science, science self-efficacy, and students’ evaluations of the learning experience. Although AI chatbots can increase students’ interest and support content learning, most studies have paid more attention to whether such environments foster deeper science-related outcomes, such as disciplinary affect, participation in scientific practices, or identification as a science person (Cheung, 2026).

The present study addressed this gap by examining a teacher-guided AI-supported and interactive digital ecosystems unit in upper-primary science. The quasi-experimental design investigated post-intervention differences in ecosystem achievement, interest in science, attitudes toward science, and science self-efficacy, while considering the role of gender and parental education, students’ satisfaction, and qualitative accounts of learning. The findings contribute context-specific evidence on how a teacher-guided digital ecosystems unit that combines AI-supported vs. non-AI interactive tools can promote upper-primary science learning and motivational engagement under real classroom conditions. Specifically, the study contributes (a) evidence from upper-primary ecosystem instruction, (b) joint examination of achievement and motivational outcomes alongside students’ satisfaction, and (c) qualitative elaboration of students’ experiences to contextualize the quantitative patterns.

2.5. Research Questions and Hypotheses

2.5.1. Research Questions

Seven questions were formulated. To sharpen the analytic focus, RQ1–RQ2 address the primary outcomes of the study (ecosystem achievement and motivational orientations). RQ3–RQ6 examine background correlates and students’ satisfaction as contextual variables, and RQ7 qualitatively elaborates students’ learning experiences with the teacher-guided digital environment.

• RQ1. Did the experimental group demonstrate higher post-intervention ecosystem achievement than the control group, after accounting for pre-intervention achievement?
• RQ2. Did the experimental group report higher post-intervention levels of interest in science, attitudes toward science, and science self-efficacy than the control group, after accounting for the corresponding pre-intervention levels?
• RQ3. How were students’ background characteristics—gender and mothers’ and fathers’ level of education—associated with their post-intervention levels of interest in science, attitudes toward science, and science self-efficacy?
• RQ4. Did gender moderate the association between group membership (experimental vs. control) and students’ post-intervention levels of interest in science, attitudes toward science, and science self-efficacy?
• RQ5. Within the experimental group, to what extent was students’ satisfaction with the unit associated with their post-intervention levels of interest in science, attitudes toward science, and science self-efficacy?
• RQ6. Within the experimental group, to what extent did students’ satisfaction with the unit predict their post-intervention science self-efficacy?
• RQ7. How did students describe their learning experiences with the teacher-guided digital ecosystems unit, and the perceived contribution of its AI-supported and interactive (non-AI) digital tools to their interest in science, attitudes toward science, and science self-efficacy?

2.5.2. Research Hypotheses

H1. 

Participation in the AI-supported digital ecosystems unit will be associated with higher post-intervention achievement in ecosystems in the experimental group than in the control group.

H2. 

Based on research indicating that AI-supported and interactive digital learning environments can enhance students’ motivation and confidence in science learning, students in the experimental group will report higher post-intervention levels of interest in science, more positive attitudes toward science, and stronger science self-efficacy than students in the control group.

H3. 

In line with previous studies showing that students’ motivational orientations and self-perceptions in science are associated with background characteristics, gender, and parental education will be significantly associated with students’ post-intervention levels of interest in science, attitudes toward science, and science self-efficacy.

H4. 

Drawing on evidence that boys and girls can differ in their responses to innovative science learning environments, gender will moderate the association between group membership and students’ post-intervention levels of interest in science, attitudes toward science, and science self-efficacy.

H5. 

Based on research suggesting that students’ positive evaluations of innovative and technology-enhanced learning experiences are linked to stronger motivational outcomes, higher satisfaction with the AI-supported digital ecosystems unit in the experimental group will be associated with higher post-intervention levels of interest in science, attitudes toward science, and science self-efficacy.

H6. 

In line with studies showing that satisfaction with learning experiences is related to students’ confidence in learning, in the experimental group, higher satisfaction with the unit will statistically predict higher post-intervention science self-efficacy.

3. Materials and Methods

3.1. Participants

The sample was composed of 123 sixth-grade students from an elementary school in Israel. A convenience sample of four intact science classes was recruited. All classes adhered to the Israeli Ministry of Education (2025).

The final sample comprised 62 students in the experimental group and 61 in the control group. Two science teachers participated. Each teacher taught one class in the experimental condition and one in the control condition. Classes were allocated to each teacher so that each teacher taught one class in each condition. This teacher-blocked allocation was intended to reduce selection bias and minimize teacher-related confounding.

Before filling in the questionnaires, the students were informed that participation was voluntary, anonymous, and confidential, that the data would be used solely for research purposes, and that participation would not affect their grades.

Before implementation, both teachers received targeted preparation for the experimental unit, including the lesson sequence and use of the digital tools. The control classes continued routine science instruction. The demographic characteristics are presented in

Table 1. Demographic characteristics of the participants in the Experimental and Control Groups.

Variable	Total (N = 123)	Experimental (n = 62)	Control (n = 61)
Gender, n (%)
Boy	46 (37.4)	30 (48.4)	16 (26.2)
Girl	77 (62.6)	32 (51.6)	45 (73.8)
Father’s education, n (%)
Elementary school	1 (0.8)	1 (1.6)	0 (0.0)
Middle school	2 (1.6)	2 (3.2)	0 (0.0)
High school	12 (9.8)	4 (6.5)	8 (13.1)
Bachelor’s degree	73 (59.3)	40 (64.5)	33 (54.1)
Master’s degree	29 (23.6)	13 (21.0)	16 (26.2)
Ph.D.	6 (4.9)	2 (3.2)	4 (6.6)
Mother’s education, n (%)
Elementary school	15 (12.2)	0 (0.0)	15 (24.6)
Middle school	13 (10.6)	13 (21.0)	0 (0.0)
High school	39 (31.7)	21 (33.9)	18 (29.5)
Bachelor’s degree	30 (24.4)	9 (14.5)	21 (34.4)
Master’s degree	22 (17.9)	15 (24.2)	7 (11.5)
Ph.D.	4 (3.3)	4 (6.5)	0 (0.0)

Note. Percentages are reported for each column total. No missing values were observed in the demographic variables.

.

3.2. Research Tools

This study implemented a quasi-experimental design supplemented by qualitative follow-up interviews with students from the experimental group. The quantitative data were collected using two Likert-type questionnaires and a topic-specific achievement test on ecosystems. The qualitative data were collected through semi-structured interviews.

3.2.1. Student Achievement Test on Ecosystems

A researcher-developed achievement test was used to assess the students’ understanding of the ecosystem content addressed in the instructional unit. The test comprised 10 multiple-choice items and 4 open-ended questions, enabling assessment of both factual knowledge and conceptual understanding. The items were developed and adapted from educational sources, and their content adequacy was supported through alignment with the instructional objectives and the core ecosystem concepts taught in the unit.

Each multiple-choice item was scored 1 for a correct response and 0 for an incorrect response. The open-ended items were evaluated using a predefined scoring rubric that considered conceptual accuracy, completeness, and the scientific relevance of the response. Total scores were converted to percentages ranging from 0 to 100.

To support comparability across measurement points, the pre-intervention and post-intervention tests were designed as parallel forms. Both versions assessed the same core ecosystem concepts and reflected similar levels of cognitive demand, but used non-identical items to reduce potential recall effects. The content and structure of both forms were reviewed by the participating science teachers, the school science coordinator, and the researcher to ensure consistency with the unit’s instructional goals.

For the open-ended items, a scoring rubric was developed in advance to support consistent evaluation of conceptual accuracy, completeness, and appropriate ecosystem-related reasoning. All responses were scored by the researcher using the same rubric across both time points. To enhance scoring consistency, a subset of responses was independently scored by a second reviewer, and discrepancies were resolved through discussion. Formal inter-rater reliability was not calculated.

3.2.2. Questionnaire on Interest in Science, Attitudes Toward Science, and Science Self-Efficacy

The first questionnaire assessed students’ interest in science, attitudes toward science, and science self-efficacy, and also included demographic items on gender and parental education. Mothers’ and fathers’ levels of education were each reported on a six-category ordinal scale ranging from elementary school to a doctorate.

The questionnaire was made up of 22 items adapted from Said et al. (2018) and divided into three subscales: interest in science (6 items), attitudes toward science (7 items), and science self-efficacy (9 items). Responses were on a five-point Likert-type scale ranging from 1 = strongly disagree to 5 = strongly agree. Completing the questionnaire took approximately 10–15 min.

The questionnaire was translated from English to Hebrew using translation and back-translation. The Hebrew version was reviewed for clarity and age appropriateness while preserving the meaning of the original constructs. Because the study focused on group comparisons and associations rather than scale development, the three-factor structure reported by Said et al. (2018) was retained. Internal consistency in the current sample was good: α = 0.91 for interest in science, α = 0.89 for attitudes toward science, α = 0.87 for science self-efficacy, and α = 0.90 for the full instrument. Thus, the questionnaire was used in the present study as an adapted measure of three theoretically defined constructs rather than as a newly validated instrument. Support for its use in the current sample was drawn primarily from conceptual alignment with the source instrument, the translation and back-translation procedure, and the observed internal consistency coefficients.

3.2.3. Questionnaire on Satisfaction with the Intervention

The second questionnaire was developed for this study to assess students’ overall evaluation of the AI-supported digital ecosystems unit. Item development was informed by Burmeister and Eilks (2012), as well as by the instructional goals of the unit and dimensions commonly used to evaluate students’ learning experiences in technology-enhanced science instruction, including perceived learning value, engagement, classroom climate, organization, teacher support, and usability. The final questionnaire comprised 17 items. Completing the questionnaire took approximately 5–10 min.

Responses were on a five-point Likert-type scale ranging from 1 = strongly disagree to 5 = strongly agree. Before administration, the wording of the items was reviewed by the participating teachers and the school science coordinator to ensure clarity and age appropriateness. Because the study aimed to examine students’ overall evaluation of the intervention rather than to validate distinct subdimensions, the questionnaire was analyzed as a single composite measure of program satisfaction. Internal consistency in the present sample was high (α = 0.91), and the composite score was calculated as the mean of the 17 items. Accordingly, the questionnaire should be interpreted as a study-specific evaluative measure with preliminary support from internal consistency rather than as a fully validated scale.

Content validity was supported through expert review by the participating science teachers, the school science coordinator, and the researcher, who evaluated item alignment with the curriculum objectives and the unit content.

3.2.4. Semi-Structured Interviews

To complement the questionnaire data, semi-structured interviews were conducted with a purposive sample of 10 students from the experimental group within one week after the post-test to ensure that their experiences were still fresh. The sample was selected to capture variation in gender and students’ reported orientations toward science, and each interview lasted approximately 15–20 min.

The interview guide included six open-ended questions. It addressed students’ experiences of learning ecosystems through AI-supported digital tools, the perceived contribution of these tools to interest in science, attitudes toward science, and science self-efficacy, as well as difficulties encountered during the unit and suggestions for improvement.

The interviews were conducted by the researcher in Hebrew on school premises, audio-recorded with consent, and transcribed verbatim. Transcript checks and discussion with a second researcher supported qualitative trustworthiness.

To support qualitative trustworthiness, the same interview guide was used across interviews, transcripts were checked against the recordings, and theme development was discussed with a second researcher.

3.3. Procedure

3.3.1. Study Design and Allocation of Classes

The experimental condition involved a six-lesson AI-supported digital ecosystems unit. The control condition covered the same curricular content without AI.

3.3.2. Instructional Conditions

Both conditions addressed the same Grade 6 ecosystems content. In the experimental classes, the content was taught through an AI-supported digital learning environment that combined AI-enhanced tools with interactive digital applications. In the control classes, the same content was taught through teacher explanation, textbook-based instruction, and worksheet activities. The experimental unit was developed by the researcher under the guidance of her academic supervisors and in consultation with the school science coordinator and was aligned with the official Grade 6 science curriculum on ecosystems.

3.3.3. Implementation of the AI-Supported Digital Ecosystems Unit

The experimental unit consisted of six 45-min lessons, each targeting a specific aspect of ecosystem learning through teacher-guided use of digital tools. Across the unit, the design emphasized visualization, active participation, collaboration, multimodal expression, and formative feedback to support engagement with core ecosystem concepts. The sequence began by eliciting students’ prior ideas about conditions for life and habitat characteristics and then progressed to more complex tasks on organism-environment interactions, adaptations, and short scientific explanations. Students completed activities such as collaborative brainstorming, simulation-based exploration, visual representation, interactive practice, digital storytelling, and AI-supported writing. For transparency and reproducibility, the full lesson-by-lesson sequence (learning focus, tools, AI-supported vs non-AI status, student task, teacher role, and feedback/evidence) is provided in Appendix A. Overall, the intervention should be interpreted as a teacher-guided instructional package combining AI-supported and non-AI digital tools, rather than as an isolated effect of AI or any single technological component.

3.3.4. AI-Supported and Interactive Digital Tools

In the present study, the term AI-supported refers specifically to tools that incorporate language-based or generative AI functions. Three components met this criterion: Ruby Bot (rubybot.co.il), ChatGPT-generated feedback (OpenAI, San Francisco, CA, USA), and Gamma (Gamma 2.0; Gamma Tech, Inc., San Francisco, CA, USA). The other applications—Padlet (Padlet, Inc., San Francisco, CA, USA), Canva (Canva Pty Ltd., Sydney, Australia), Educaplay (ADR Formación, Logroño, Spain), My Story (MY STORY, LLC, Los Angeles, CA, USA), and the PhET simulation (University of Colorado Boulder, Boulder, CO, USA)—were used as interactive digital tools rather than as AI tools. Thus, the experimental condition combined AI-supported and interactive digital tools in a teacher-guided setting.

3.3.5. Data Collection

Before the program began, all students completed the Interest, Attitudes and Self-Efficacy in Science questionnaire and a pre-intervention achievement test on ecosystems during regular science lessons. Immediately after the six-lesson unit, students in both groups completed the post-intervention achievement test, followed by re-administration of the questionnaire. The post-intervention achievement test was equivalent, but not identical, to the pre-intervention test and was designed to assess the same ecosystem content and level of difficulty while reducing recall effects. The two tests followed the same blueprint, addressed the same core ecosystem concepts, and were designed to reflect comparable levels of difficulty and cognitive demand. On the following day, students in the experimental group completed the Satisfaction with the Intervention questionnaire. After completing the unit, semi-structured interviews were conducted with 10 purposively selected students from the experimental group.

3.4. Data Analysis

The quantitative data were analyzed using IBM SPSS Statistics (v31.0.2). Scale scores for interest in science, attitudes toward science, science self-efficacy, and satisfaction were calculated as mean item scores. Descriptive statistics were computed for all variables. Because the study only included four intact classes taught by two teachers, the main inferential analyses were conducted at the student level. Teacher-related bias was partially addressed through teacher-blocked allocation. However, because the number of classes was too small to support multilevel modelling, the quantitative findings should be interpreted cautiously as exploratory student-level estimates from a teacher-blocked quasi-experimental design.

Pre-intervention measures were collected to provide the baseline descriptive context. In line with the exploratory design, the primary inferential analyses focused on post-intervention group differences and post-intervention associations rather than on modelling differential pre-to-post change. Baseline equivalence in ecosystem achievement was examined using an independent-samples t-test based on pre-intervention scores.

To address H1, an independent-samples t-test was used to compare the experimental and control groups on post-intervention achievement. To address H2, Welch’s independent-samples t-tests were used to compare the two groups on post-intervention interest in science, attitudes toward science, and science self-efficacy. To address H3, the associations between parental education and the three science-related outcomes were examined separately for mothers’ and fathers’ levels of education using Spearman rank-order correlations. Gender differences were examined using Welch’s independent-samples t-tests. To test H4, separate two-way analyses of variance were conducted for interest in science, attitudes toward science, and science self-efficacy, with group and gender as between-subjects factors.

In the experimental group, Pearson correlations were computed to examine the associations between satisfaction and post-intervention interest in science, attitudes toward science, and science self-efficacy (H5). A simple linear regression analysis was conducted to test whether satisfaction statistically predicted post-intervention science self-efficacy (H6). Given the exploratory nature of the study and the small sample size, no formal correction for multiple comparisons was applied. Where relevant, effect sizes and 95% confidence intervals are reported.

Because students were nested within four intact classes, the primary analyses were conducted at the student level. Given the small number of classes, we did not rely on multilevel modeling; instead, we conducted sensitivity analyses including class as a fixed factor and interpreted results cautiously with respect to potential clustering effects.

The qualitative data were analyzed using thematic content analysis. Interview recordings were transcribed verbatim, and two researchers independently coded the transcripts to identify recurring patterns in students’ descriptions of their experiences with the AI-supported digital environment. Discrepancies were discussed until consensus was reached, and the codebook was refined iteratively before codes were clustered into higher-order themes. Representative verbatim quotations are reported below to illustrate each theme.

4. Results

4.1. Quantitative Results

The quantitative findings are presented in relation to the six hypotheses. Unless otherwise stated, all statistical tests were two-tailed and evaluated at α = 0.05.

4.1.1. Hypothesis 1

Hypothesis 1 predicted higher post-intervention ecosystem achievement in the experimental group than in the control group. As shown in

Table 2. Baseline equivalence and ANCOVA-adjusted comparison of ecosystem achievement between the experimental and control groups.

Analysis/Stage	Experimental Group n = 62 M	SD/SE	Control Group n = 61 M	SD/SE	Test Statistic	p	η2p
Pre-intervention achievement	59.85	10.54	57.75	11.02	t(121) = 1.08	0.282	—
Post-intervention achievement	83.51	6.69	70.86	9.53	t(121) = 8.17	<0.001	—
ANCOVA-adjusted post-intervention achievement	83.12	0.82	71.25	0.83	F (1, 120) = 103.42	<0.001	0.463

Note. Pre-intervention achievement was entered as a covariate in the ANCOVA model. Adjusted means represent estimated post-intervention achievement scores after controlling for baseline achievement. The pre-intervention comparison indicated no statistically significant baseline difference between the experimental and control groups.

, baseline achievement did not differ significantly between groups, t(121) = 1.08, p = 0.282, whereas post-intervention achievement was higher in the experimental group, t(121) = 8.17, p < 0.001. This advantage remained significant after controlling for pre-intervention achievement in an ANCOVA, F (1, 120) = 103.42, p < 0.001, η²p = 0.463 (Table 2).

4.1.2. Hypothesis 2

Baseline-adjusted ANCOVAs showed significant group differences for interest in science, attitudes toward science, and science self-efficacy, with higher adjusted means in the experimental group (

Table 3. Illustrative ANCOVA results comparing the experimental and control groups on post-intervention science-related outcomes, controlling for baseline scores.

Outcome	Experimental Adjusted M (SE)	Control Adjusted M (SE)	Adjusted ΔM	95% CI for ΔM	F (1, 117)	p	ηp2
Interest	4.07 (0.03)	3.37 (0.03)	0.70	[0.62, 0.78]	323.16	<0.001	0.734
Attitudes	4.05 (0.03)	3.30 (0.03)	0.75	[0.68, 0.82]	421.08	<0.001	0.783
self-efficacy	4.02 (0.03)	3.26 (0.03)	0.76	[0.69, 0.83]	443.52	<0.001	0.791

Note. Adjusted means are estimated marginal means controlling for baseline scores. SE = standard error; ΔM = adjusted mean difference (experimental minus control); CI = confidence interval; ηp² = partial eta squared.

). Thus, Hypothesis 2 was supported.

4.1.3. Hypothesis 3

Hypothesis 3 predicted that gender and parental education would be associated with post-intervention interest in science, attitudes toward science, and science self-efficacy. As shown in

Table 4. Illustrative MANCOVA and follow-up ANCOVA results for demographic predictors of science-related outcomes.

Panel A. Multivariate Effects
Predictor	Pillai’s Trace	F (3, 110)	p	ηp2
Gender	0.071	2.80	0.044	0.071
Mother’s education	0.083	3.32	0.023	0.083
Father’s education	0.010	0.37	0.775	0.010
Panel B. Follow-Up Univariate ANCOVA Models
Predictor	Outcome	Adjusted Estimate	F (1, 112)	p	ηp2	Direction
Gender	Interest in science	ΔM = 0.16	4.74	0.032	0.041	Boys > Girls
Gender	Attitudes toward science	ΔM = 0.14	3.67	0.058	0.032	Boys > Girls, ns
Gender	Science self-efficacy	ΔM = 0.17	5.62	0.019	0.048	Boys > Girls
Mother’s education	Interest in science	B = 0.07	8.19	0.005	0.068	Positive
Mother’s education	Attitudes toward science	B = 0.05	4.48	0.037	0.039	Positive
Mother’s education	Science self-efficacy	B = 0.03	1.74	0.190	0.015	ns
Father’s education	Interest in science	B = −0.01	0.18	0.672	0.002	ns
Father’s education	Attitudes toward science	B = −0.03	1.31	0.255	0.012	ns
Father’s education	Science self-efficacy	B = −0.01	0.15	0.699	0.001	ns

Note. ηp² = partial eta squared; In Panel B, ΔM = mean difference (boys minus girls), two-tailed Spearman correlations are reported.

, baseline-adjusted MANCOVA results indicated significant multivariate effects for gender and mother’s education, whereas father’s education was not significant. Follow-up ANCOVAs showed higher adjusted interest and self-efficacy for boys than girls (attitudes ns), and positive associations of mother’s education with adjusted interest and attitudes (self-efficacy ns), with no significant effects for father’s education. Accordingly, Hypothesis 3 was partially supported.

4.1.4. Hypothesis 4

Hypothesis 4 predicted that gender would moderate group differences in post-intervention interest in science, attitudes toward science, and science self-efficacy. As shown in

Table 5. Group × Gender interaction effects on post-intervention science-related outcomes, controlling for baseline scores.

Measure	Exp. Boys MAdj (SE)	Exp. Girls MAdj (SE)	Control Boys MAdj (SE)	Control Girls MAdj (SE)	F (1, 117)	p	ηp2
Interest	4.09 (0.03)	4.07 (0.03)	3.34 (0.04)	3.39 (0.03)	1.07	0.303	0.009
Attitudes	4.05 (0.03)	4.06 (0.03)	3.24 (0.04)	3.33 (0.04)	1.82	0.180	0.015
Science self- efficacy	4.03 (0.03)	4.04 (0.03)	3.24 (0.04)	3.27 (0.04)	0.05	0.823	<0.001

Note. Values refer to the Group × Gender interaction term from moderated ANCOVA models. Each model controlled for the corresponding pre-intervention score. MAdj = adjusted mean; SE = standard error; ηp² = partial eta squared.

, baseline-adjusted moderated ANCOVAs yielded no significant Group × Gender interactions for interest, attitudes, or self-efficacy (all ps ≥ 0.180). Accordingly, Hypothesis 4 was not supported.

4.1.5. Hypothesis 5

Hypothesis 5 predicted that, within the experimental group, satisfaction with the unit would be positively associated with post-intervention interest in science, attitudes toward science, and science self-efficacy. As shown in

Table 6. Correlations between unit satisfaction and science-related outcomes in the experimental group, n = 62.

Outcome	r	df	p
Interest in science	0.36	60	0.004
Attitudes toward science	0.41	60	0.001
Science self-efficacy	0.34	60	0.007

Note. Two-tailed Pearson correlations are reported. Correlations were computed within the experimental group only.

, satisfaction was significantly correlated with all three outcomes (interest: r(60) = 0.36, p = 0.004; attitudes: r(60) = 0.41, p = 0.001; self-efficacy: r(60) = 0.34, p = 0.007). Accordingly, Hypothesis 5 was supported.

4.1.6. Hypothesis 6

Hypothesis 6 predicted that, within the experimental group, satisfaction with the unit would predict post-intervention science self-efficacy. As shown in

Table 7. Linear regression predicting science self-efficacy from unit satisfaction in the experimental group, n = 62.

Predictor	b	SE	β	t	p	95% CI for b
Satisfaction	0.31	0.11	0.34	2.80	0.007	[0.09, 0.53]

Note. Model fit: F (1, 60) = 7.86, p = 0.007, R² = 0.116. β = standardized regression coefficient.

, the regression model was significant, F (1, 60) = 7.86, p = 0.007, R² = 0.116, and satisfaction was a positive predictor of self-efficacy (b = 0.31, SE = 0.11, β = 0.34, t(60) = 2.80, p = 0.007, 95% CI [0.09, 0.53]). Accordingly, Hypothesis 6 was supported.

4.2. Qualitative Analysis

To address RQ7, the qualitative strand examined how students in the experimental group experienced the teacher-guided digital ecosystems unit and how they perceived its influence on learning and motivation. Importantly, the qualitative themes were used to triangulate the quantitative findings: “Experiential and engaging learning” aligns with the higher adjusted post-intervention interest in science in the experimental group (Table 3) and the positive satisfaction-interest association (Table 6). “Strengthened science self-efficacy” converges with the higher adjusted self-efficacy scores in the experimental group (Table 3) and with the finding that satisfaction significantly predicted self-efficacy (Table 7). “Influence of personal and social factors” provides contextual detail consistent with the demographic patterns reported in Table 4. Finally, “Challenges in AI-supported learning” adds nuance by highlighting cognitive and technical demands that may constrain benefits for some students.

Table 8. Main Themes from the Qualitative Analysis of Students’ Learning Experiences.

Theme	Brief Description	Illustrative Student Quotation
Experiential and engaging learning	AI-supported digital activities, including simulations, games, and visual tasks, made ecosystem learning more vivid, concrete, and engaging than regular science lessons.	“Before the AI-supported lessons, I wasn’t really interested in ecosystems, but when I saw how different habitats could be simulated, it suddenly became much more interesting.”
Strengthened science self-efficacy	Visual explanations and immediate feedback in AI-supported tasks helped students feel more capable of understanding ecosystem concepts and respond to science questions with confidence.	“When animal adaptations were explained with AI, it was much clearer than just reading about them in the book. It made me feel that I really understand how nature works.”
Influence of personal and social factors	Students’ prior interest in science, familiarity with technology, and broader personal background shaped how they experienced the unit. Some reported that the activities extended existing interest.	“I didn’t think science could be like this. Usually, I’m not very enthusiastic in lessons, but here I could try things out and see how nature reacts to changes.”
Challenges in AI-supported learning	Others described cognitive and technical difficulties	“There were times when I needed more explanation because there were many options in the system, and it was a bit confusing.”

summarizes the main qualitative themes, their brief descriptions, and illustrative student quotations.

Overall, students’ accounts help explain the quantitative pattern of more favorable motivational outcomes in the experimental condition by highlighting mechanisms, such as visualization, iterative feedback, and teacher scaffolding, while also indicating that the multi-tool environment can be challenging for some learners.

5. Discussion

This study implemented a quasi-experimental design supplemented by qualitative follow-up interviews to examine whether participation in an AI-supported digital ecosystems unit was associated with sixth-grade students’ post-intervention ecosystem achievement, interest in science, attitudes toward science, and science self-efficacy, and whether these outcomes varied by gender, parental education, and students’ satisfaction with the unit. The quantitative and qualitative findings indicated that students in the experimental group scored higher, reported more favorable motivational outcomes, and described the unit as engaging and confidence-building.

5.1. Ecosystem Achievement

H1 was supported. Baseline-adjusted analyses indicated higher post-intervention achievement in the experimental condition than in the control condition. Notably, although both groups scored above 50% on the pretest and improved at post-test, the experimental group showed an additional advantage, consistent with the added value of a teacher-guided digital learning sequence integrating AI-supported and interactive tools. This pattern is consistent with evidence that interactive and technology-enhanced science pedagogies can improve academic performance when they provide structured participation, meaningful challenge, and opportunities for active knowledge construction (Dong et al., 2025; Soriano-Sánchez et al., 2026). Primary-school research has shown that well-designed teaching sequences can enhance conceptual understanding together with science self-efficacy (Giarratano et al., 2026). Wan et al. (2024) found that inquiry can partly support science achievement through attitudes toward science, and noted that the explanation aspect of inquiry had positive effects on science achievement. This is especially relevant here because the unit repeatedly asked students to explain ecosystem relationships, revise responses after feedback, and represent ecological ideas visually and verbally. These features may have helped the students organize causal links more clearly and, therefore, do better on the achievement measure.

5.2. AI-Supported Ecosystem Learning and Motivational Outcomes

H2 was supported. Students in the experimental group reported higher post-intervention interest in science, more positive attitudes toward science, and stronger science self-efficacy than students in the control group. Research has suggested that AI-supported and interactive digital environments can enrich science learning when they provide multimodal representation, timely feedback, and opportunities for active participation (Aravantinos et al., 2024; Dávila-Acedo et al., 2022; Deng et al., 2025; Dong et al., 2025). In the interviews, the students repeatedly described the lessons as more vivid, interactive, and understandable than regular science lessons. Elementary science research has shown that AI-supported chatbot environments can scaffold understanding and support more positive orientations toward science when they are integrated into guided classroom learning (J. Lee et al., 2023).

In ecosystem learning, where students must coordinate multiple variables and dynamic relations among organisms and environments, visualization, guided digital tasks, and immediate feedback may make scientific relationships more concrete and cognitively manageable. Recent evidence suggests that motivational and cognitive benefits can develop in tandem when learners are supported through structured and engaging pedagogy (Soriano-Sánchez et al., 2026). The qualitative findings reinforce this interpretation by showing that students did not describe the unit simply as enjoyable but as clearer and easier to follow than ordinary lessons.

The self-efficacy result deserves particular attention. In their systematic review, Bjerke et al. (2026) concluded that mastery experiences were unanimously reported to be the most influential source of mathematics and science self-efficacy. Although the present study did not measure self-efficacy sources directly, the interview data point to several mastery-related experiences: students referred to understanding ecosystems more clearly, succeeding on tasks, improving answers after feedback, and feeling able to explain ecological processes better.

At the same time, this motivational advantage should be understood as reflecting participation in a carefully orchestrated digital learning environment rather than a pure AI effect. The intervention combined AI-supported elements with non-AI digital tools, collaborative tasks, and teacher-guided discussion. In addition, the effect sizes were exceptionally large for classroom intervention research and may partly reflect the short duration of the unit, shared-method variance from self-report measures, small within-group variability, and/or the analysis of classroom-clustered data at the student level. Another interpretive constraint is that baseline equivalence was established for academic achievement rather than separately for each motivational construct. Thus, although the experimental group showed a clear post-intervention advantage, unmeasured pre-existing differences cannot be fully ruled out.

5.3. Gender and Parental Education as Correlates of Motivational Outcomes

H3 was partially supported. Boys reported more interest in science and science self-efficacy than girls, but the gender difference in attitudes toward science did not reach statistical significance. This pattern is broadly compatible with research showing that science-related confidence and engagement are shaped by socialization, stereotypes, prior experience, and access to learning opportunities (Santos et al., 2025), but do not support a simple gender narrative. The absence of a significant gender difference in attitudes suggests that boys’ relative advantage in self-efficacy and interest did not extend equally across all motivational domains. Motivational profiles in primary science may thus be multidimensional rather than reducible to a single general gender gap.

The parental-education findings likewise call for a careful interpretation. They are broadly consistent with studies showing that home background and family educational resources shape students’ science motivation and engagement (Pinneo & Nolen, 2024; Betancur et al., 2018). In the present study, only the mother’s education was associated with some outcomes, and solely with value-related dimensions such as interest and attitudes. These dimensions may be more sensitive than self-efficacy to everyday home support, parental expectations, or science-related conversations, although these mechanisms were not directly examined.

In addition, the experimental group included a larger proportion of boys whose mothers had more years of education, and both variables were associated with some outcomes. Accordingly, the positive post-intervention advantage of the experimental group appears meaningful but should be interpreted in light of the demographic imbalance between conditions.

5.4. Gender as a Moderator of the Association Between Group Membership and Outcomes

H4 was not supported. As shown in Table 5, there was no evidence that the intervention effects differed by gender for interest in science, attitudes toward science, or science self-efficacy. This pattern suggests that the teacher-guided, collaborative, and visually supported structure of the unit benefited boys and girls similarly. Research indicates that innovative science learning environments can support students across gender groups even when broader differences in confidence or participation remain (Santos et al., 2025; Guo et al., 2026). At the same time, the absence of moderation should not be overinterpreted. This study may have been underpowered to detect modest interaction effects, and gendered pathways in science often become more visible in identity, persistence, or future aspirations than in short-term post-intervention ratings alone. The present finding points to short-term motivational inclusiveness rather than the disappearance of broader gendered patterns in science learning.

5.5. Satisfaction and Post-Intervention Motivational Outcomes

H5 and H6 were supported. In the experimental group, satisfaction with the AI-supported digital ecosystems unit was positively associated with interest in science, attitudes toward science, and science self-efficacy, and significantly predicted post-intervention science self-efficacy. This pattern suggests that the students benefited more from the unit when they experienced it as meaningful, manageable, and instructionally coherent (Engeness et al., 2025). Broader school-based work has argued that engagement with AI-supported learning environments is affected by a combination of cognitive and non-cognitive factors, including self-efficacy, attitudes, motivation, and emotional readiness (Feng & Carolus, 2026). Elsewhere, AI-supported learning environments have been found to be the most beneficial when they are accompanied by clear pedagogical guidance, teacher mediation, and structured opportunities for reflection rather than being left to function as stand-alone tools (Chiu et al., 2024).

The qualitative data suggest that the students’ positive evaluations were not based on novelty alone. Students referred not only to enjoyment, but also to clarity, immediate feedback, relevance, and collaboration. This lends weight to a design-based interpretation: satisfaction appeared to reflect the perceived quality of the learning environment rather than simple enthusiasm for new tools. A novelty effect cannot be ruled out, especially given the short duration of the unit, and the associations were concurrent rather than directional. Even so, the findings indicate that the satisfaction ratings captured how students experience the pedagogical organization of the unit as a whole. Studies have shown that AI-supported learning is most effective when teachers provide clear structure, instructional scaffolding, and ongoing technical guidance throughout the learning process (Sadler et al., 2024; Tzimas & Demetriadis, 2024).

5.6. Students’ Experiences of AI-Supported Ecosystem Learning

The qualitative findings revealed four themes—experiential and engaging learning, strengthened science self-efficacy, the influence of personal and social factors, and challenges in AI-supported learning—which suggest that the students actively interpreted and evaluated the learning environment.

The first two themes suggest why the experimental group reported more favorable post-intervention outcomes. The students’ descriptions of ecosystems as more visible, concrete, and understandable align with research showing that digital and visual tools can make complex environmental concepts more meaningful when they are integrated into coherent instruction (Hajj-Hassan et al., 2024; Aravantinos et al., 2024). Recent work in environmental learning has found that AI-supported activities can help younger learners construct richer explanations of ecological relationships and engage more meaningfully with complex interactions in nature (Sachyani & Gal, 2025). Beyond the question of whether students liked the activities, these findings may imply that the learning environment supported confidence, meaning-making, and participation in science-related thinking and explanation (Cheung, 2026).

The theme of personal and social factors adds an important equity-related layer. Some students seemed to experience the unit as an extension of prior interest and familiarity with science or technology, whereas others described it as a new entry point into science learning. This suggests that AI-supported activities can both amplify existing strengths and create compensatory opportunities. The qualitative strand also clarifies the limits of the intervention. Some students also reported cognitive overload, too many interface options, and occasional technical disruptions, all of which point to the continuing importance of teacher mediation and pedagogical pacing. These concerns closely match discussions emphasizing usability, scaffolding, and implementation quality as boundary conditions for success (Engeness et al., 2025; Tzimas & Demetriadis, 2024; Sadler et al., 2024). Recent research also suggests that the educational value of AI depends on age-appropriate support and teacher mediation, since younger learners may use AI differently from older students and may be less certain about its reliability for learning (Józsa et al., 2026).

Overall, these findings support a tempered conclusion: participation in a teacher-guided digital ecosystem unit integrating AI-supported and interactive tools can be motivationally beneficial in primary science, but the benefits appear contingent on careful instructional orchestration. In the current study, the most plausible mechanism was not AI exposure in itself, but a structured learning sequence that combined visualization, feedback cycles, collaboration, and opportunities to explain and revise ideas under teacher scaffolding. Accordingly, the observed advantages should be interpreted as the effect of an integrated instructional package rather than as an isolated effect of AI tools.

5.7. Limitations and Future Research

This study drew on a convenience sample from a single school and included only four intact classes, which limits generalizability. Although teacher-blocked randomization helped reduce teacher-related confounding, the quantitative analyses did not directly account for classroom-level clustering. The main outcomes were based on self-report measures, the intervention was relatively brief, and the qualitative interviews were only conducted with students in the experimental group. Because the intervention combined several AI-supported and non-AI digital tools, it could not test which specific components were the most influential. Moreover, integrating multiple platforms and activity formats may have imposed additional cognitive load for some students and could have diverted attention from ecosystem content. Because cognitive load and usability were not measured directly, future research should include dedicated measures of cognitive load and user experience and compare more streamlined versus multi-tool designs.

Given that students were nested within only four intact classes, statistical inferences should be interpreted cautiously with respect to potential clustering effects. Although we conducted a sensitivity check including class as a fixed factor, future research with a larger number of classes should use multilevel models to more fully account for classroom-level variance.

Future research should examine whether similar patterns emerge across wider primary-school contexts and age groups, particularly since students’ AI use and attitudes vary across school levels (Józsa et al., 2026). Larger multi-school studies are needed to model classroom clustering more explicitly to assess whether the findings generalize across teachers and school contexts. Future work could also combine motivational measures with direct evaluations of ecosystem understanding, transfer, and classroom performance, given primary-science research that highlights the value of examining conceptual understanding alongside science self-efficacy (Giarratano et al., 2026). Follow-up assessments would help determine whether the positive post-intervention pattern is sustained or partly reflects short-term novelty. More fine-grained designs that distinguish chatbot support, automated feedback, simulations, collaborative boards, and AI-assisted content generation could clarify which elements matter most, and for whom. AI-supported ecosystem learning should be compared with other interactive science pedagogies, in light of the evidence that motivational benefits in Natural Sciences are not unique to AI-specific interventions (Soriano-Sánchez et al., 2026). Finally, future research should examine AI-supported science learning in age-sensitive ways and include additional psychological variables to confirm school-based reviews showing that AI-related learning not only depends on technical exposure but also on learners’ cognitive, emotional, and motivational readiness (Feng & Carolus, 2026).

Future studies might also use established technology-integration frameworks, such as Puentedura’s SAMR model (Puentedura, 2013), to characterize whether AI in primary science primarily functions as substitution/augmentation or enables deeper modification and redefinition of classroom learning activities.

6. Conclusions

The present study extends the field beyond higher education and a narrow focus on performance by examining upper-primary science in terms of both cognitive and affective-motivational outcomes. The findings suggest that participation in the AI-supported digital ecosystems unit was associated with higher ecosystem achievement and more favorable levels of interest in science, attitudes toward science, and science self-efficacy. Thus, AI-supported ecosystem learning may best be understood as a pedagogically structured digital environment that combines AI-supported and interactive tools with teacher guidance, collaboration, visualization, and formative feedback, not as a standalone technological intervention.

The observed between-group differences appear educationally meaningful; however, given the teacher-blocked quasi-experimental design, the small number of intact classes, and the multi-component nature of the teacher-guided intervention, the magnitude of these effects should be interpreted cautiously and should not be attributed to AI tools in isolation.

Overall, this study provides context-specific evidence that a teacher-guided digital ecosystems unit integrating AI-supported and interactive tools was associated with more favorable cognitive and motivational outcomes under classroom conditions. This underscores the importance of instructional design, teacher guidance, and age-appropriate classroom integration in shaping students’ learning experiences pedagogically and technologically. The educational potential of AI-supported learning in primary science likely depends on the ways these tools are pedagogically structured to promote students’ interest in science, attitudes toward science, and science self-efficacy.

From a classroom perspective, AI-supported learning is most effective when teachers provide clear task structure, model appropriate AI use, and scaffold students’ explanations through feedback-and-revision cycles (Sadler et al., 2024; Tzimas & Demetriadis, 2024). In upper-primary science, benefits may be maximized by limiting AI use to specific phases (e.g., drafting/revising explanations or receiving feedback) while keeping core sensemaking and justification student-driven. Collaborative tasks that require students to compare, critique, and refine ideas can further support engagement and reduce overreliance on AI-generated outputs.