The independent variable was categorical. Students in the experimental condition routinely wrote self-explanations and submitted them through the LMS to receive elaborative feedback from ChatGPT, whereas students in the control condition revised their explanations based solely on the teacher’s written rubric. The primary dependent variable was conceptual understanding of the wave–particle duality of light, measured with a 24-item two-tier multiple-choice test; scores were converted to a 0–100 scale and showed high internal reliability (KR-20 = .82). The quality of students’ self-explanations was treated as the second dependent variable and was assessed with an expert-validated five-level SOLO-based rubric that achieved an inter-rater consistency index of .88.

Note. Δ Mean = adjusted posttest mean minus pretest mean.

3.2 Conceptual Understanding

The ANCOVA revealed a significant effect of the instructional model on conceptual understanding after controlling for prior knowledge, F(1, 59) = 24.56, p < .001, partial η² = .29. The adjusted mean difference of 8.5 points (95% CI [4.8, 12.2]) corresponded to Cohen’s d = 1.25, a large effect.

3.3 Self-Explanation Quality

The median SOLO rubric score in the experimental group increased from 2 to 4 (multistructural → relational/extended abstract), whereas that of the control group rose to 3. The Mann–Whitney test yielded U = 228.5, p < .001, r = 0.55, indicating a large effect.

3.4 Learning Engagement

On the learning engagement scale, the experimental group (M = 4.21, SD = 0.39) scored significantly higher than the control group (M = 3.68, SD = 0.45), t(60) = 3.34, p = .001, d = 0.86.

3.5 ChatGPT Interaction Patterns

Of the 446 ChatGPT interactions analyzed, 57% were elaborative (explaining core concepts and providing examples), 25% were metacognitive (prompting students to reflect on their strategies), and 18% were surface-level descriptive. Self-explanations were revised twice on average after feedback, and student excerpts showed improved clarity in linking the photoelectric equation with the interference pattern.

4. Discussion

The quantitative findings show that integrating ChatGPT feedback raised students’ mean posttest score by 8.5 points after covariate adjustment, equivalent to Cohen’s d = 1.25. This effect exceeds the average improvement (g ≈ 0.87) summarized in a meta-analysis of 51 cross-disciplinary ChatGPT studies, suggesting that modern physics—with abstract phenomena such as interference and the photoelectric effect—benefits substantially from generative-AI scaffolding (Wang & Fan, 2025). The consistency of this pattern is also in line with meta-analytic evidence identifying engagement as a key mediator of the benefits of ChatGPT (Heung & Chiu, 2025; Rabiu & Nuhu, 2024).

From the perspective of self-explanation theory, the rise in SOLO levels from multistructural to relational/extended abstract signals deep conceptual restructuring. The SOLO rubric used in this study assessed the extent to which students “unified” the wave and particle models; the improvement underscores the role of elaborative feedback in facilitating the integrative stage that is most difficult to reach through conventional instruction alone (Brion, 2024). In other words, ChatGPT functioned as a “dialogue partner” that pushed students to revise their mental representations until the two models became coherent.

The qualitative findings revealed that 57% of ChatGPT responses were elaborative (conceptual expansion), 25% metacognitive (inviting strategic reflection), and 18% descriptive. This pattern indicates that the quality of feedback—not merely its speed—is the key factor. Tsing and Liu (2025), studying automated feedback in physics problem solving, reported a similar profile in which only deep elaboration was associated with gains in transfer scores. Teachers therefore need to design prompt templates that elicit elaborative and metacognitive output rather than shallow answers.

Through the lens of Cognitive Load Theory (Sweller, 1988), AI feedback can balance the intrinsic load of quantum topics with productive germane load. Yan et al. (2025) distinguished between momentary performance gains and genuine learning; in this study, although immediate scores increased, students’ revision traces and interviews indicated internalization of the concepts, countering concerns about “performance inflation” unaccompanied by real learning. Research on AI-based cognitive load likewise shows that adaptive systems are effective when they provide just-in-time scaffolding, as observed in the elaborative logs of this study (Gkintoni et al., 2025).

The affective dimension is equally important: the engagement score of the experimental group (M = 4.21) was significantly higher (d = 0.86). A recent systematic review on ChatGPT and engagement attributes this to the combination of immediacy and interactive dialogue, which strengthens students’ sense of control over their learning (Heung & Chiu, 2025). Interviews in this study reinforced this finding: students reported “feeling attended to” because the feedback was personalized, and they were also motivated to “prove” that human thinking remains crucial.

These results arrive at an important policy moment in Indonesia. Beginning in the 2025 academic year, the government has been preparing an AI curriculum across all levels of schooling; this study provides evidence of good practice that could be adopted as a showcase of how AI supports the scientific literacy targets of the Kurikulum Merdeka. The implications include the need for teacher training in prompt literacy, quality assurance of AI-generated content, and standardized mechanisms for auditing student data.

Nevertheless, the education ecosystem must remain alert to the risk of over-reliance. An ongoing MIT study reported weakened neural activity when students repeatedly relied on ChatGPT to write essays, indicating potential skill atrophy when AI is used without critical reflection (Kosmyna & D’Ausilio, 2025). The present data did not indicate such a decline within the four-week intervention, but follow-up surveys are needed to monitor long-term effects, particularly on independent reasoning.

4.1 Limitations and Future Research

The limitations of this study include a single-site sample, a short intervention period, and the absence of retention measures several months after the intervention. Multi-site longitudinal designs examining other topics—for example, the quantum mechanics of the hydrogen atom—would test the generalizability of the findings. In the future, triangulation with eye-tracking or EEG data could separate performance gains attributable to AI “copy-editing” from genuine changes in cognitive representation.

Overall, this discussion shows that ChatGPT can act as adaptive scaffolding that reduces cognitive load, deepens conceptual elaboration, and increases engagement, provided that implementation is supported by strong AI literacy and ethical-pedagogical oversight. Further research will determine whether these benefits are sustainable and can be scaled nationally to improve Indonesia’s standing in future PISA surveys.

5. Conclusion

This study demonstrates that integrating ChatGPT-based feedback into a self-explanation strategy significantly improves understanding of the wave–particle duality of light, deepens the quality of students’ self-explanations, and simultaneously raises their learning engagement. Using a cluster-randomized design, the ANCOVA showed an adjusted posttest mean difference of 8.5 points, corresponding to a large effect size (Cohen’s d = 1.25). The rise in SOLO levels from multistructural to relational/extended abstract indicates more mature conceptual restructuring, while the distribution of ChatGPT feedback—dominated by elaborative and metacognitive types—confirms the importance of high-quality generative-AI scaffolding.

These findings contribute to the literature by showing that, in cognitively demanding modern physics topics, AI can act as an adaptive tutor that balances the intrinsic load of the material with productive germane load, thereby fostering meaningful learning. In practical terms, the results support the adoption of AI in classrooms as a partner to teachers, provided that it is accompanied by prompt literacy, content-audit mechanisms, and critical reflection to prevent over-reliance.

However, the generalizability of the findings is limited by the single-school context and the four-week intervention period. Multi-site longitudinal research, long-term retention analyses, and application to other modern physics topics are needed to establish the sustainability and breadth of the impact. Overall, this study provides empirical evidence that using ChatGPT as a real-time feedback provider can be an effective strategy for improving the scientific literacy of senior high school students while supporting the digital transformation of education in Indonesia.

5.1 Recommendations

1. Schools should provide dedicated training on designing prompts that elicit elaborative and metacognitive output from ChatGPT, while equipping students to critically evaluate and revise AI responses rather than copying them verbatim.
2. Education authorities could incorporate an “AI as a Learning Partner” module into the Kurikulum Merdeka, beginning with modern physics and expanding to other subjects after successful piloting. The module should address ethical use, data privacy, and techniques for auditing AI-generated content.
3. Although ChatGPT provides real-time responses, teachers remain responsible for selecting, supplementing, and, where necessary, correcting AI feedback so that it aligns with competency targets and the classroom context; this double-check mechanism minimizes the risk of misinformation.
4. Beyond academic achievement, it is important to monitor changes in students’ learning motivation, self-confidence, and potential dependence on AI. Instruments such as motivation surveys, eye-tracking, or neurocognitive studies could form part of a holistic evaluation.

By implementing these recommendations, the use of ChatGPT as an adaptive feedback provider in modern physics classrooms is expected not only to improve scientific literacy but also to build a critical and ethical learning culture in the era of AI-based learning.

Declarations

Ethics statement. This study was conducted with the permission of SMA Taruna Bakti. Informed consent was obtained from all participating students and their parents or guardians. [Add ethics approval body and reference number if available.]

Funding. This research received no external funding.

Conflict of interest. The author declares no conflict of interest.

Data availability. The data supporting the findings of this study are available from the corresponding author upon reasonable request.

Use of generative AI. ChatGPT was used as part of the research intervention as described in the Method section. [Declare any AI tools used for language editing or translation of this manuscript, as required by the journal.