Skip to main content
NCBI home page
F1000Research logoLink to F1000Research
. 2026 Aug 8;15:1339. [Version 1] doi: 10.12688/f1000research.186991.1

Conceptual Understanding and Conceptual Change Through Human AI Collaboration in Science Education: An Integrated Conceptual Framework from a Systematic Literature Review

Riski Muliyani 1, Andi Suhandi 1,a, Eka Cahya Prima 1, Muslim Muslim 1, Yudi Kurniawan 1, Dhita Rismayani Priatna 1,2
PMCID: PMC13612998  PMID: 42800973

Abstract

Human Artificial Intelligence (Human AI) Collaboration in science education is gaining momentum and more research is required on how it can facilitate students’ conceptual understanding and conceptual change. Currently, most studies on AI in science education investigate the use of one particular tool and its effects on learning. However, it is the learning mechanisms that are triggered by these tools that help develop knowledge. This systematic literature review was conducted by performing a systematic search within the Scopus database which was last updated on 15 June 2026. In line with the PRISMA 2020 guidelines, a total of 28 empirical studies research were analyzed. These studies of empirical research were published between 2023 and 2026. Six roles of AI supporting the students’ conceptual understanding as well as five mechanisms supporting the students’ conceptual change were identified. Five interconnected learning mechanisms were derived from the analyzed studies, which are presented in an integrated conceptual framework. The framework is based on the theoretical framework of constructivism, the views on learning of the sociocultural theory, the approach of inquiry based science learning as well as on the metacognitive theory. Within the framework AI functions as a learning partner. The framework illustrates how, within science learning environments that are supported by AI, students continuously construct, evaluate and refine their scientific knowledge in order to achieve deeper understanding.

Keywords: Human AI Collaboration; Conceptual understanding; Conceptual change; Misconception; Systematic reviews.

1. Introduction

Advances in Generative Artificial Intelligence (Gen AI) and Large Language Models (LLMs) have recently taken center stage in Intelligent Technologies for Education. While past Intelligent Technologies for Education supported learners to retrieve information, to generate content, or to engage in automated tutoring, today’s sophisticated Artificial Intelligence (AI) systems support learners’ cognition, metacognition, and epistemic processes to construct advanced scientific knowledge ( Kim et al., 2022; Kong et al., 2025). As a result, Human AI Collaboration, or learning in which learners and AI systems engage in a cyclical and continuous process of constructing, evaluating, and improving knowledge, is becoming a new learning paradigm ( Cress & Kimmerle, 2023). That is, AI systems are evolving from being simple instructional aids to being learners’ important epistemic partners who support learners’ reasoning, reflection, explanation, and evidence evaluation at all points in the learning process.

Human AI Collaboration is not merely a new technology to support learning, it transforms our understanding of how scientific knowledge is generated and improved. Rather than just making education more efficient by providing information, giving feedback or supporting learning to solve problems, new technologies support processes of knowledge construction of learners. They generate scientific explanations, visualizations, adaptive scaffolding, dialog for reflection and feedback ( Yan et al., 2024; Zhai et al., 2024). Thus, learning with AI changes our perspective on learning from just transferring scientific knowledge contained in instructional resources to learners, to a collaborative process in which human beings and AI systems jointly and progressively construct knowledge by means of interactions between human thinking and AI supported reasoning ( Tsakeni et al., 2025; Yan et al., 2024).

For the last few decades technology has been developing at a rapid pace and this is also apparent within the area of education. New ways of learning are arising and thus also new ways of teaching within science education. Learning science is more than just learning information; it also requires more than just being able to apply algorithms or to memorize facts. It requires understanding of the world in a scientific way. This process requires learners to construct a coherent conceptual representation of the scientific phenomena that they encounter, to recognize inconsistencies in their prior knowledge, to evaluate competing scientific explanations of natural phenomena, and to reorganize their conceptual structures into scientifically accepted models of natural phenomena ( El Fathi et al., 2025; Pacaci et al., 2024). Hence, the processes of achieving conceptual understanding and that of conceptual change are not two separate ends which learners strive to reach. Rather, these are ongoing processes that learners go through in trying to construct scientific knowledge of the world ( Addido et al., 2022). Research into Human AI Collaboration is an emerging field. Evidence from this emerging field is beginning to reveal the educational value of Human AI Collaboration. It goes beyond increasing learners’ test scores in learning scientific concepts. Rather, it has the potential to transform the ways in which learners can construct, evaluate and transform scientific knowledge in order to make sense of the world. It can help them to identify their own misconceptions, to evaluate evidence, to generate alternative representations and models of scientific phenomena, to engage in reflective dialogue, and to construct mental models of scientific phenomena and then to reconceptualize them through multiple interactions with the AI system ( El Fathi et al., 2025; Zhai et al., 2024).

Although considerable progress has been made in the field of Human AI Collaboration and learning with AI, there is much yet to be understood about learning. Current learning theories of conceptual understanding and conceptual change were developed within the context of human-centered educational settings and thus do not provide enough explanations of learning processes that take place within environments that are mediated by AI. Such environments for Human AI Collaboration are not merely for the transmission of scientific knowledge; they support learners to negotiate meanings, to evaluate AI generated explanations, to compare different representations of identical content, and most importantly, to continuously reorganize their conceptual knowledge by way of their interaction with the AI system. While current learning theories can explain how humans attain conceptual understanding of scientific facts within a human learning environment, they are not sufficient to explain the more complex processes of knowledge construction that take place within Human AI Collaboration environments. A more comprehensive theory of learning is required ( Ding et al., 2024; D. Lee et al., 2021; Pacaci et al., 2024).

A rapidly growing body of research has been conducted on how AI can support education within learning environments by means of various functions ( Almasri, 2024; S. S. Lee & Moore, 2024). Many studies investigate the use of Intelligent Tutoring Systems, conversational agents, adaptive learning, automatically generated learning contents and visualizations as well as formative feedback and diagnostics by means of AI. Most of these technologies are used in isolation so that even though individual technologies have the potential to support learning processes, little is known about the learning processes supported by Human AI Collaboration. Also, it is unclear how different educational mechanisms and technologies interact and how they together can support students in constructing scientific knowledge and in achieving a deeper understanding and conceptual change by means of Educational Mechanisms and Technologies ( Kong et al., 2025; Yan et al., 2024). Thus, currently knowledge on technologies and learning mechanisms is fragmented and mostly lacks a theoretical foundation that explains how single educational mechanisms and technologies interact in order to support students’ conceptual understanding.

Although systematic literature reviews on AI in education have been conducted, they still have limitations. Most studies that focus on different aspects of the use of AI in education, on the instructional value of AI, on the different types of AI technologies that are used in education, on learners’ perspectives on and uses of AI in education, on the ethical issues of the use of AI in education, and so on ( Kamalov et al., 2023; Létourneau et al., 2025), have rarely investigated Human AI Collaboration. In particular, while a number of studies have explained how learning with AI affects learners’ cognitive, metacognitive, and epistemic processes, no systematic literature review has yet developed an integrated theoretical framework that explains how Human AI Collaboration can support learners’ construction of scientific knowledge through the processes of attaining conceptual understanding, achieving conceptual change, and constructing scientific knowledge. Most studies on using AI to support science learning have thus treated the different functions that AI can perform in isolation of each other and thus cannot provide a comprehensive integrated understanding of learning with AI. This review aims at several objectives. These objectives are to be answered by the research questions for this review.

RQ1: How does Human AI Collaboration support students’ conceptual understanding of scientific concepts in science education?

RQ2: How do Human AI Collaborations support students’ conceptual change in science learning?

RQ3: What are the learning mechanisms that connect students’ conceptual understanding and their conceptual change when they learn science by collaborating with AI?

RQ4: What integrated conceptual framework can be developed to describe how Human–AI Collaboration can support students to develop their scientific knowledge through their conceptual understanding and their conceptual change in science education?

2. Methods

2.1 Study design

The purpose of this study, conducted in the form of a Systematic Literature Review (SLR), was to identify, evaluate and organize empirical studies regarding Human AI Collaboration with the objective of fostering students’ conceptual understanding and supporting their conceptual change in science education. By following a transparent, consistent and replicable procedure, in accordance with the guidelines of the PRISMA model for systematic reviews, the present study aimed at not only providing a mapping of studies, but also at examining the underlying mechanisms that connect conceptual understanding and conceptual change in the context of Human AI Collaboration for science education and to develop a comprehensive conceptual framework ( Page et al., 2021).

2.2 Eligibility criteria

To filter the results, criteria for the inclusion and exclusion of studies were defined and applied to the title and abstract as well as the full text screening step based on the PICOS framework (Population, Intervention, Context, Outcomes, and Study Characteristics) ( Brignardello-Petersen et al., 2025). Only studies with students or teachers in a science education setting in which AI was integrated in the learning process were included in the review. The studies had to report on the following learning outcomes: students’ conceptual understanding, students’ conceptual change, students’ misconceptions, and students’ misconceptions remediation. These inclusion and exclusion criteria are listed in detail in Table 1.

Table 1. Inclusion and exclusion criteria.

PICOS Criteria Inclusion Exclusion
Study Characteristics (S)
  • •
    Experimental/quasi experimental research (quantitative, qualitative, mixed methods or design based).
  • •
    Peer reviewed journal articles and peer reviewed conference papers.
  • •
    Full text available in English.
  • •
    Published between 2023 and 2026.
  • •
    Literature reviews, systematic reviews, meta analyses, bibliometric studies, editorials, book reviews, conference abstracts, dissertations and theses, and non peer reviewed publications.
  • •
    Purely theoretical papers without empirical data.
Population (P)
  • •
    Any level of learner (primary, secondary, tertiary or informal science education).
  • •
    Pre service teachers and in service teachers.
  • •
    General public samples without educational objectives.
  • •
    Participants involved in non science learning contexts.
Intervention (I)
  • •
    Studies on human AI collaboration in science learning environments where learners interact with AI during their science learning activities.
  • •
    The AI functions as tutor, coach, mentor, discussion partner, feedback giver, reflection partner, misconception detector etc. as a cognitive scaffold for learning.
  • •
    AI system development studies and/or studies that report on AI system’s technical performance.
  • •
    The use of an AI system solely for administrative tasks (e.g. grading, scheduling, attendance).
  • •
    AI operating without meaningful learner AI interaction.
Comparator/Context (C)
  • •
    Studies on Human AI collaboration are compared to traditional instruction or to non AI learning environments or even without a comparison group within different perspectives on learning processes.
  • •
    Physics, chemistry, biology, earth science, environmental science, integrated science and/or STEM education.
  • •
    Only different AI models, algorithms or system architectures are compared, but not implemented for educational purposes.
  • •
    Non science educational contexts.
Outcomes (O)
  • •
    Evidence of conceptual understanding.
  • •
    Evidence of conceptual change.
  • •
    Misconception identification, diagnosis, or reduction.
  • •
    Knowledge restructuring or conceptual reconstruction.
  • •
    Degree to which learners develop their understanding of scientific explanations or of scientific reasoning in Human AI collaboration settings.
  • •
    Measured outcomes such as users’ satisfaction, technology acceptance, system’s usability, users’ engagement, users’ motivation, and users’ attitudes toward using AI in learning environments without measuring conceptual learning.
  • •
    The study only presents results regarding the technical accuracy or performance of the AI system developed in the study.
Open in a new tab

2.3 Search strategy

For the literature study a search strategy was developed. The test searches that were conducted with the already known keywords and terms related to the topic of artificial intelligence and conceptual understanding of the search strategy for literature study. A systematic search was then conducted on the database Scopus, a very large and very reliable database of peer reviewed journals, within the scientific area of (experimental and social) education, (natural and social) science and technology. The search string consists of terms that are often used in research on artificial intelligence and the learning outcomes that are relevant to conceptual understanding. The search query that was actually used is: TITLE-ABS-KEY ((“artificial intelligence” OR “generative AI” OR “ChatGPT” OR “large language model” OR “AI chatbot”) AND (“conceptual understanding” OR “concept learning” OR “conceptual change” OR “misconception”)). The literature search was conducted on 15 June 2026. Search results are limited to English language studies published from 2023 onwards in order to identify the most recent relevant evidence for the review. Importantly, studies were selected for this review, in part, because they addressed aspects of Human AI Collaboration, even when that term was not explicitly written. Studies that detailed human AI interactions that supported students’ conceptual understanding and/or addressed issues of students’ conceptual change, and their existing and investigated misconceptions and ways of remediation were selected for this review.

2.4 Study selection

Initially the articles, retrieved through the Scopus search, were downloaded from the Scopus website as a CSV format. Then these articles went through a three step process of screening: title screening, abstract screening and full text review. Articles were excluded on the basis of title screening if they dealt with topics which were clearly not relevant to education and/or did not contain any aspects of AI supported learning. The abstracts of the articles which had not been excluded on the basis of title screening were then screened against the inclusion and exclusion criteria for relevant studies on Human AI Collaboration and their implications for learning and teaching as well as for studies dealing with topics related to conceptual understanding, conceptual change and misconceptions and their remediation. In case of doubt potentially relevant articles were read in their entirety as part of the full text review in order to check whether all inclusion and exclusion criteria were met. Due note was taken of three aspects, namely (1) the type of human AI interaction studied in a particular article, (2) the educational setting in which an AI system was implemented, and (3) the learning outcomes in terms of the learners’ conceptual understanding, their experiences of conceptual change, their identification of and their remediation of misconceptions. The study selection process is presented in the PRISMA flow diagram ( Figure 1).

Figure 1. PRISMA flow diagram of the study selection process.

Figure 1.

Open in a new tab

2.5 Data extraction

After selecting the relevant studies, we extracted data from the studies and organized them in an Excel file with four main parts of information. The first part contains information about the studies, i.e., level of participants, scientific domain, type of study and type of AI used in the study. The second part contains information about Human AI Collaboration in the studies, i.e., the role of AI in education, type of human AI interactions and learning activities in which human and AI collaborate to facilitate learning. The third part contains information about the specific conceptual learning outcomes that were measured in the studies, i.e., students’ understanding of specific concepts, level of conceptual change, students’ ability to recognize their own and peers’ misconceptions and students’ ability to remediate these misconceptions. Finally, the fourth part of information that we extracted contains information about the theoretical foundations of the studies and about the learning mechanisms that are used in order to support students’ learning, i.e., feedback, scaffolding, coaching and scripting.

2.6 Data analysis and synthesis

This analysis was carried out by means of thematic coding by the author, based on the above stated framework. The selected articles were analyzed with respect to four criteria: (1) form of Human AI Collaboration, (2) learning outcomes in terms of conceptual content, (3) learning mechanisms which promote the development of conceptual content, and (4) theoretical frameworks which are used in order to implement AI in science education. In an iterative process, the findings from the single studies were analyzed using the aforementioned thematic coding. The emerging categories and themes were compared within and across the studies. In addition, the results from the single studies were summarized using descriptive statistics regarding the characteristics of the included studies as well as the applied AI technologies and the forms of Human AI Collaboration. A narrative synthesis of the coding results was conducted to identify recurring patterns, interthematic relationships, and research deficits. The results of the analysis were used to develop a conceptual framework outlining the relations of Human AI Collaboration, learning mechanisms, conceptual understanding, and conceptual change of conceptual understanding in science education.

3. Result

3.1 Characteristics of the review studies

The present review includes 28 studies and in detail describes the current research on Human AI Collaboration in science education. A synthesis of studies was conducted and a study characteristics table was set up and analyzed with respect to the following criteria: year of publication, students’ educational level, science domain, AI technology used and research design. The trends of Human AI Collaboration in order to support students’ conceptual understanding, conceptual change and to identify and to remediate students’ misconceptions in science education are described in the present review. A summary of the characteristics of the reviewed studies is given in Table 2.

Table 2. Characteristics of the included studies.

Characteristics Categories n %
Publication Year 2023 3 10.71
2024 7 25
2025 11 39.29
2026 7 25
Education Level Primary Education 4 14.29
Secondary Education 5 17.86
Higher Education 10 35.71
Teacher Education 7 25
Mixed/Professional 2 7.14
Science Domain Physics 9 32.14
Chemistry 3 10.72
Biology 2 7.14
Environmental and Earth Science 7 25
Integrated Science/STEM 7 25
AI Technology LLMs & Generative Chatbots 14 50
Machine Learning & AIoT 5 17.86
Intelligent Tutoring Systems & Adaptive AI 4 14.29
Visualization, Audio & Metaverse 5 17.85
Research Design Mixed Methods Research 8 28.57
Design and Development Research 6 21.43
Experimental Research 5 17.86
Qualitative Research 4 14.28
Other Designs 5 17.86
Open in a new tab

3.2. Human AI collaboration for supporting conceptual understanding

We reviewed studies where different AI technologies were used to support Human AI Collaboration and students’ conceptual understanding. Our synthesis showed that Human AI Collaboration supports students’ scientific understanding by constructing, refining, evaluating, and reorganizing scientific knowledge with the help of AI tools that function as a cognitive partner. Human AI Collaboration more than just providing information to students, supports their conceptual understanding by mediating their cognitive and metacognitive processes in developing their conceptual understanding. Our synthesis of the reviewed studies led to the identification of six different but complementary roles of Human AI Collaboration ( Table 3).

Table 3. Human AI collaboration for supporting conceptual understanding.

Human AI Role Function AI Technologies Used Conceptual Understanding Outcome Representative Studies
AI Tutor and Cognitive Scaffolding Provides adaptive guidance and problem solving support Rule based chatbot, GPT-4, MyBotBuddy, GPAiM (GenAI agents) Enhanced conceptual understanding and higher order thinking skills Khor,  E. T., & Chan, L, 2025; Lee,  et al., 2023; Lizano-Sánchez et al., 2025; Wu et al., 2026
Inquiry Facilitator and Problem Solving Partner Exploring data, conducting experiments and modeling of natural phenomena Teachable Machine, Node RED, XAI (LIME), IoT platforms, Python (Astroquery) Concept understanding is transferred to computational models in order to deepen practical as well as conceptual understanding Chookaew et al., 2024; Marín Díaz, 2025; Serik et al., 2025; Zha et al., 2025
Feedback and Assessment Assistant Provides immediate feedback and formative assessment ITAS (hybrid scoring system), AI adaptive systems, LLM based evaluation More accurate identification of students’ conceptual difficulties and improvement in their conceptual performance Brückner et al., 2026; Kurniawan et al., 2024
Visualization and Explanation Assistant Presents abstract concepts through multimodal representations Bing AI, Ideogram, ChatGPT, Claude, TTS tools Improved comprehension of complex scientific concepts Clark, 2023; De Souza et al., 2024; Dmitrenko, 2025
Dialogue and Reflection Partner Facilitates dialogue and metacognitive reflection ChatGPT, DeepSeek, Teachable Machine (AMAIL framework) Development of deeper mental models and conceptual reasoning Kaharu et al., 2026; Relmasira & Donaldson, 2025
Co Creator and Instructional Design Partner Supports collaborative instructional design and knowledge production ChatGPT, Gemini, Jupyter AI Deepened conceptual understanding and improved learning organization Elshall & Badir, 2025; Gonçalves Costa et al., 2024; Kim,  W. J. 2026.
Open in a new tab

Several functions were assigned to AI in the reviewed studies and as before they are of a pedagogical nature. They are not linked to any single technology and include functions of a tutor, supporting learners in their constructive processes and providing them with adaptive scaffolds. Also discussed were the function of an inquiry facilitator allowing students to deal with scientific issues in a problem solving manner, functions of feedback and assessment as well as of visualization and functions for visualization and explanation of given contents and for enabling dialogue and reflection. Finally there is the function of a co creator: as well as supporting single learners in their knowledge production AI also supports learning in teams. Its use supports the production of knowledge and the design of learning as well as of teaching. The roles for AI presented above thus support the learners’ knowledge construction by backing the various processes instead of single learning tasks.

Most importantly for us, however, the effectiveness of Human AI Collaboration is determined more dependent on the quality of learners’ interactions with AI than on the degree of technical sophistication of the employed technology. While studies have reported generally positive learning outcomes with a variety of forms of AI, learners have been enabled to use intelligent tutoring systems, generative AI, adaptive learning environments as well as a variety of different visualizations to build scientific knowledge and to complete school tasks ( Nguyen et al., 2024; Vaccaro et al., 2024). Thus, various technologies can be used to support essentially the same pedagogical functions, thus functioning primarily as learning partners for Human AI Collaboration. AI should therefore not be confounded with information ( Jose et al., 2025; Kim et al., 2022; Yan et al., 2024).

A key finding from the synthesized material is that the focus of current research on use of AI for science learning has shifted from an assessment of short-term acquisition of science concepts to a more continuous construction and refinement of a learner’s scientific model of the world for science concepts over time ( Hudson et al., 2023; Zhu et al., 2022). Most current research utilizing AI in order to support learning of science concepts involves the use of adaptive tutoring by the AI system. However, a growing body of research is now supporting the use of AI for a number of other functions in support of learning of science concepts. For example: (a) support for scientific inquiry; (b) support for learners’ reflection on learning science; (c) support for learners’ argumentation regarding science concepts; and (d) support for learners’ collaborative meaning making regarding science learning ( Katsenou et al., 2025; Yusuf et al., 2025).

3.3. Human AI Collaboration for Supporting Conceptual Change

An analysis of the studies, investigating students’ conceptual change of their understanding when collaborating with AI, revealed five strategies that effectively support students to identify, evaluate and change their own conceptions. These five strategies are presented in Table 4.

Table 4. Human AI collaboration for supporting conceptual change.

Human AI mechanism Function AI Technologies used Conceptual change outcome Representative studies
Misconception Diagnosis and Pattern Detection Identifies learners’ misconceptions and cognitive structures Machine Learning Ensemble (Multilayer Perceptron, Support Vector Machine, Random Forest), Natural Language Processing, GPT-4, GPT-5 More accurate diagnosis of misconceptions Braune et al., 2026; Conradty & Bogner, 2026; Kökver et al., 2025
Adaptive Scaffolding and Feedback Driven Revision Provides adaptive feedback and corrective explanations for erroneous concepts Rule-based Intelligent Tutoring System (Aristarchus), Learning Analytics Dashboard Knowledge restructuring and enhanced scores in student understanding Georgiou et al., 2026; Pektaş et al., 2026
Reflective Dialogue and Mental Model Externalization Helps students explore and externalize their mental models Bing AI/DALL.E, ChatGPT Better scientific conceptual coherence and deeper reflection De Souza et al., 2024; Ding,  et all., 2023; Nendauni, 2025
Cognitive Conflict and Affective Monitoring Leverages cognitive conflicts experienced by students to trigger conceptual change Affectiva Facial Emotion Recognition, iMotions Conceptual shift from misconceptions to scientific concepts, accompanied by students’ “aha” or “oh” moments Ezquerra et al., 2024
Fact Checking and Information Validation Encourages critical evaluation and analysis of AI generated responses ChatGPT Improved scientific literacy and information validation skills Mennella & Quadros-Mennella, 2024
Open in a new tab

The reviewed studies examined in what ways Human AI Collaboration can support students’ conceptual change. On the basis of five different types of AI technologies five different mechanisms for fostering the five mentioned steps for giving given conceptions a more accurate form are identified and described in Table 4. These five mechanisms supporting students in identifying, evaluating or even changing their given conceptions step by step for developing more accurate scientific explanations for a given problem of science learning in general support a process of knowledge restructuring. Thus, in addition to identifying their given misconceptions the five different types of AI tools support learners in developing more accurate scientific explanations for a given problem of science learning instead of being used in isolation as a variety of separated educational tools.

In addition to identifying various forms of technology used to support various functions in order to bring about conceptual change through learners’ active processing of learning processes (e.g. diagnose misconceptions and support feedback, adaptive feedback and reflection, as well as information evaluation), the results reveal that the greatest value of AI in the context of education pertains to its pedagogical potential. The current state of research in Human AI Collaboration can be synopsized by the so-called corrective perspective. First, the AI system recognizes mistakes by learners and then the AI system supports learners in their learning process by means of adaptive feedback that replaces learners’ misconceptions with corresponding scientific conceptions ( Ding et al., 2024; El Fathi et al., 2025; Pacaci et al., 2024). Such an AI system supports learners to attain higher levels of conceptual understanding. However, from a theoretical point of view, current approaches to Human AI Collaboration only explain parts of processes of conceptual change because they do not sufficiently account for learners’ active construction of scientific knowledge while they are engaged in learning with an AI system, whereas error correction by the AI system is predominantly in the focus of current approaches to Human AI Collaboration.

New research on AI is increasingly starting to consider it from a broader knowledge construction perspective. Thus, as opposed to focusing on its use for diagnosis or for correction, AI is increasingly being viewed as a learning partner with whom students can engage in reflection, dialogue, argument, meaning making and knowledge construction ( Sánchez Muñoz et al., 2025). Conceptual change as a result of such AI mediated learning is a gradual process of examination, of justification, of refinement and of restructuring of scientific knowledge, a process that takes place while students are interacting with the technology that supports their learning.

While there are an increasing number of studies that investigate how humans and AIs can support each other in learning, an in depth analysis of such collaboration is still rare in the literature. While misconception diagnosis and feedback are still the focus of most studies, the potential of AIs to support students in their reflection, in their collaborative reasoning and in their knowledge construction based on evidence, is not yet explored enough. Thus, in the future, research on Human AI Collaboration in learning has to move from technology based interventions to design of environments that support conceptual change as a process of scientific knowledge construction over time.

3.4. Mechanisms linking conceptual understanding and conceptual change

In contrast to the two distinct learning objectives of conceptual understanding and conceptual change within Human AI Collaboration, these two objectives are part of a continuum of learners’ processes of developing their scientific understanding when collaborating with AI. Combining the results from Tables 3 and 4, five interconnected learning mechanisms of learners’ processes of developing, examining and restructuring their scientific concepts were identified (see Table 5). The processes of developing, examining and restructuring of learners’ scientific concepts form a continuum leading from their first conceptual understanding to more refined or even completely replaced concepts by more adequate scientific explanations during Human AI Collaboration over time.

Table 5. Learning mechanisms of human AI collaboration supporting conceptual understanding and conceptual change.

Learning Mechanism Contribution to Conceptual Understanding and Conceptual Change Limitations in the Literature
Adaptive Support (Scaffolding and Feedback) Helps learners to first come up with ideas and to correct mistakes by means of learning support which is given in the form of hints and feedback. While the majority of studies on formative feedback on test results concentrate on error correction and in the long run on test scores, only a few studies investigate the matter of long term conceptual change.
Knowledge Exploration and Construction Supports learners to build up mental models of their knowledge and to understand concepts by problem solving, visualization or different representations of the knowledge to be learned by the learner. Most of the studies that have been mentioned above above follow learners in order to investigate the final results of their understanding of scientific concepts, whereas an in depth study of the developmental processes of their understanding is rarely carried out.
Reflection and Dialogue Students can also explain, evaluate or even change their own ideas or understanding by having a dialogue with the AI or by reflecting on its output. Even though nowadays so called reflective dialogues are more and more often used in educational processes, they are mostly treated as additional support in AI based instructional design processes.
Cognitive Conflict Cognitive conflict and its role in bringing about learners’ knowledge restructuring by recognizing dissonance between their initial understanding and scientific explanations. Cognitive conflict is mostly used in isolation from other strategies such as reflection and argumentation in educational design.
Information Evaluation and Validation Aids students to verify facts and answers generated by the AI system and more importantly teaches them to evaluate information from any source. Most studies on using AI in science education treat AI as mainly giving answers to questions posed by the student and only secondarily as learning partner which the student can use to verify information and together with the AI to co construct scientific meaning and evaluate sources of information for accuracy.
Open in a new tab

These five functions are part of a learning process. Firstly, learners get to know new things with the help of adaptive support and feedback. Then they work more on the knowledge they already have acquired with the help of knowledge exploration, visualization, problem solving and various ways of presenting scientific content. Later on in the learning process, learners can reflect on their own learning process by means of discussion and the explanation of their own ideas. They question their own ideas and improve them by discussing with others. In order to make learners aware of differences between their existing conceptions and the scientific explanations, the cognitive conflict that occurs is a crucial function. The functions of information evaluation and validation are the last functions in the learning process. In these functions learners check the AI generated responses and other sources of evidence that they have found by themselves. After that they integrate the new scientific knowledge that they have acquired into their existing understanding of the corresponding scientific content. In this way, conceptual change is a continuous process of learning that takes place step by step by means of various learning activities. Several forms of technology support the same processes of learning. Thus, Intelligent Tutoring Systems, generativeAI, machine learning, learning analytics, data visualization, and conversational AI Systems all support learning processes that remain consistent across technologies. So, supporting Human AI Collaboration efficiently for learning depends more on how these several forms of technology can support learning than on the technology itself.

The current research is one sided, as the majority of studies and research on adaptive technologies for learning, on misconception diagnosis, and on feedback, are rooted in the traditional corrective view of learning for conceptual change ( Katsenou et al., 2025; Kim et al., 2022). These studies and other similar studies offer insights on how students can be supported in order to achieve higher conceptual accuracy in learning scientific concepts, however, they only outline a part of the learning process for achieving such concepts, as students are not supported in constructing their own scientific knowledge by means of inquiry, reflection, dialogue and collaborative reasoning.

A further finding from this review is the contrast between the current research on conceptual change on the one hand, and recent Human AI Collaboration studies on the other hand. Theories on conceptual change describe a slow process in which learners and their learning environments incrementally add to the learners’ existing knowledge by means of reflection, discussion, argumentation, etc. Instead of this, recent studies on Human AI Collaboration still view the AI system mainly as a diagnostic or corrective tool that supports the learners in case they do not arrive at a correct solution. In the findings from this review, however, the AI system is viewed as a learning partner that supports the learners in their processes of exploring ideas, in comparing their own explanations of phenomena, in evaluating evidence for claims, and in developing a scientific understanding of phenomena in a step by step manner. Future research on Human AI Collaboration therefore should not be restricted to the analysis of single mechanisms that are used by AI systems to support human learning in science classrooms and to study their effects on human learning in given contexts. Rather, future research should also be aimed at analyzing how several of these mechanisms can work together in order to support learners in a long term process of knowledge construction and of conceptual change.

3.5. Integrated framework of human AI collaboration for conceptual understanding and conceptual change

The functions for Human AI Collaboration, which are supported by a variety of AI functions, support both the conceptual understanding and the conceptual change of students. For the support of Human AI Collaboration for joint knowledge construction, a large number of studies on tutoring systems, adaptive feedback systems, dialogue systems, AI based visualization systems, and so on, were reviewed. The effects of the functions for Human AI Collaboration were studied in a number of cases, but in each case, the functions were studied in isolation. Therefore, knowledge about the effects of the functions for Human AI Collaboration is currently fragmented and insufficient to explain the interaction of the different learning processes that support the long term development of students’ scientific knowledge in detail. An integrated conceptual model was developed on the basis of Tables 3-5 to explain how the different AI supported learning mechanisms support the transition of students from their conceptual understanding to their conceptual change ( Figure 2).

Figure 2. Framework of Human AI Collaboration.

Figure 2.

Open in a new tab

Our framework is based on four main learning theories. The first of these is Constructivism. In AI supported science learning children can construct a scientific understanding of the world by using their prior knowledge and learning new information ( Nguyen et al., 2024). In doing so they go through different phases of cognitive conflict as their current conceptions are challenged by real world data. The second learning theory is Sociocultural Theory. This theory explains how children learn by interacting with their environment, through dialogue and the shared meaning making of learners, teachers, peers and now technology ( Wheaton et al., 2024). The third learning theory that we used is Inquiry Based Learning. Children develop a scientific understanding of the world by carrying out an investigation into a question or problem. In doing so they use evidence, modeling and solve problems in order to arrive at a conclusion ( Morris, 2025). The last learning theory that we used is Metacognitive Theory. This theory explains how children can monitor, evaluate and control their own learning while using AI generated explanations ( Tsakeni et al., 2025). Together these four learning theories provide a coherent explanation of the different processes that take place during AI supported science learning. Children’s conceptual understanding develops progressively over time through a number of different processes instead of just one single event.

In order to support learning in a sustainable way, five mechanisms for learning interact in the process of knowledge construction. At first adaptive support for learning is generated in order to establish a first understanding of the subject matter under study. This basic understanding is then in turn consolidated during the learning process by means of knowledge exploration and knowledge construction by means of inquiry, visualization, modeling and the use of multiple representations for one and the same content. In the process of learning about his or her own learning processes by means of reflection and dialogue, the learner can take a closer look at his or her own thinking. Therefore, he or she is more likely to experience a cognitive conflict when scientific evidence confronts his or her existing conceptions ( Pacaci et al., 2024; Pektaş et al., 2026). However, instead of merely changing single misconceptions, the learner is able to reorganize his or her entire existing knowledge structure as a result of this conflict ( Shih et al., 2026). In the process of learning, information evaluation and validation are crucial. Learners check their AI generated responses as well as other information in a critical manner ( Walter, 2024). They integrate the resulting scientific explanations that have been generated in this process step by step into their already existing knowledge structures. This stepwise development of the conceptual understanding of a learner leads to so called conceptual change after several cycles of the mentioned mechanisms for learning.

The proposed framework for Human AI Collaboration builds upon current research by identifying a number of as yet underdeveloped aspects. In contrast to previous studies on Human AI Collaboration, which have investigated adaptive tutoring, feedback, visualization, dialogue, or even misconception diagnosis as separate instructional functions and their short and long term effects on learning outcomes, our framework describes how these functions can function as a number of interrelated learning mechanisms that support the learners’ ongoing construction of scientific knowledge ( Ji et al., 2023; Thomas et al., 2024). While previous studies treat AI primarily as a teaching aid for correcting students’ misconceptions, our framework views the AI system as a learning partner. In contrast to the traditional view of conceptual change as a process of replacing students’ incorrect conceptions with correct scientific conceptions, which is refined and restructured in detail through their prolonged interactions in AI supported learning environments, the learners’ scientific understanding is in the long run refined and restructured in detail through their prolonged interactions in AI supported learning environments.

This framework also points out several avenues for future research. As already is the case with adaptive feedback and with correcting misconceptions, research on reflection, reflection, dialog, and cognitive conflict as processes of learning that are interwoven and as yet too not fully researched when using AI for learning, will be very important. Most studies on the use of AI in learning so far investigate one function of the AI system in isolation and then follow its use in one learning session or in several learning sessions ( Bauer et al., 2025). Instead, it is more important to investigate how in the long run several functions support learning when used in combination. The integrated framework that we presented here can be used as a starting point for the design of learning environments in which AI supports learning of science and in which learners construct, evaluate, and change scientific explanations for real world phenomena while acquiring knowledge of science.

4. Conclusion

Human AI Collaborations in science education is a field of research that is rapidly evolving and not yet fully understood. While there are many excellent studies that focus on how AI can support students in achieving their learning goals for scientific learning, there is a severe lack of in depth studies on how Human AI Collaborations can support students’ conceptual understanding and their process of conceptual change in science. Our review aimed to answer the question of how the AI system functions as a collaborative learning partner in order to support the continuous construction, evaluation, and refinement of students’ scientific knowledge about science. As a result of our review, it seems to be more appropriate to treat the processes of students’ conceptual understanding and their conceptual change as a matter of fact. Within the given framework of the learning environment, these processes are brought about by a variety of learning mechanisms rather than by a number of separate teaching methods that are intended to achieve a number of specific goals.

An integrated conceptual framework for studying Human AI Collaborative Learning in Science Education is proposed here. The framework outlines relationships between several learning theories, a set of mechanisms that can be supported by AI in order to foster Human AI Collaborative Learning, and ways by which knowledge is constructed during Science learning. This framework can serve as a groundwork for future research on Human AI Collaborative Learning in Science, for the design and implementation of environments where Humans and Artificially Intelligent agents collaborate in learning, and for the study and design of pedagogical approaches where AI functions as a learning partner for the processes of inquiry, reflection, critical evaluation and the continued construction and development of scientific knowledge during the scientific learning process.

Use of AI assisted technology

The authors of this work used Gemini during the preparation of this work for the translation and improvement of the academic language from Indonesian to English.

Acknowledgements

The authors would like to express their sincere gratitude to the Indonesian Education Scholarship, Center for Higher Education Funding and Assessment, and Indonesian Endowment Fund for Education for their support of this research.

Funding Statement

This research was supported by the Indonesian Education Scholarship, Center for Higher Education Funding and Assessment, and Indonesian Endowment Fund for Education.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

[version 1; peer review: 1 approved

Data availability statement

The completed PRISMA 2020 Checklist and Flow Diagram for Human AI Collaboration for Supporting Conceptual Understanding and Conceptual Change in Science Education: A Systematic Literature Review has been uploaded to Zenodo and published under Creative Commons Zero v1.0 Universal (CC0) license. Checklist and Flow Diagram can be downloaded online at https://doi.org/10.5281/zenodo.21430817 ( Muliyani et al., 2026).

References

  1. Addido J, Burrows A, Slater T: The Effect of the Conceptual Change Model on Conceptual Understanding of Electrostatics. Educ. Sci. 2022;12(10):696. 10.3390/educsci12100696 [DOI] [Google Scholar]
  2. Almasri F: Exploring the Impact of Artificial Intelligence in Teaching and Learning of Science: A Systematic Review of Empirical Research. Res. Sci. Educ. 2024;54(5):977–997. 10.1007/s11165-024-10176-3 [DOI] [Google Scholar]
  3. Bauer E, Greiff S, Graesser AC, et al. : Looking Beyond the Hype: Understanding the Effects of AI on Learning. Educ. Psychol. Rev. 2025;37(2):45. 10.1007/s10648-025-10020-8 [DOI] [Google Scholar]
  4. Braune J, Conradty C, Bogner FX, Paul J: A game-changer for qualitative research: Artificial intelligence as an efficient tool for analyzing student conceptions about microplastics. Frontiers in Education ,2026;11,1770878. 10.3389/feduc.2026.1770878 [DOI] [Google Scholar]
  5. Brignardello-Petersen R, Santesso N, Guyatt GH: Systematic reviews of the literature: An introduction to current methods. Am. J. Epidemiol. 2025;194(2):536–542. 10.1093/aje/kwae232 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Brückner M, Thyssen C, Huwer J: AI for adaptive science teaching: Strengthening teacher self-efficacy and perceived usefulness. Frontiers in Education. 2026;11:1759274. 10.3389/feduc.2026.1759274 [DOI] [Google Scholar]
  7. Chookaew S, Kitcharoen P, Howimanporn S, et al. : Fostering student competencies and perceptions through artificial intelligence of things educational platform. Computers and Education: Artificial Intelligence. 2024;7:100308. 10.1016/j.caeai.2024.100308 [DOI] [Google Scholar]
  8. Clark TM: Investigating the Use of an Artificial Intelligence Chatbot with General Chemistry Exam Questions. Journal of Chemical Education. 2023;100(5):1905–1916. 10.1021/acs.jchemed.3c00027 [DOI] [Google Scholar]
  9. Conradty C, Bogner FX: Strategic Competence in Sustainability Education: Conceptual Patterns Identified Through AI-Assisted Qualitative Analysis. Sustainability ,2026;18(7),3643. 10.3390/su18073643 [DOI] [Google Scholar]
  10. Cress U, Kimmerle J: Co-constructing knowledge with generative AI tools: Reflections from a CSCL perspective. Int. J. Comput.-Support. Collab. Learn. 2023;18(4):607–614. 10.1007/s11412-023-09409-w [DOI] [Google Scholar]
  11. De Souza MG, Won M, Treagust D, et al. : Visualising relativity: Assessing high school students’ understanding of complex physics concepts through AI-generated images. Physics Education. 2024;59(2):025018. 10.1088/1361-6552/ad1e71 [DOI] [Google Scholar]
  12. Ding Y, Zhu G, Bian Q, et al. : Analysis of students’ conceptual change in learning Newton’s third law with an integrated framework of model analysis and knowledge integration. Physical Review Physics Education Research. 2024;20(2):020141. 10.1103/PhysRevPhysEducRes.20.020141 [DOI] [Google Scholar]
  13. Ding L, Li T, Jiang S, Gapud A: Students’ perceptions of using ChatGPT in a physics class as a virtual tutor. International Journal of Educational Technology in Higher Education ,2023;20(1),63. 10.1186/s41239-023-00434-1 [DOI] [Google Scholar]
  14. Dmitrenko N: (2025). AI-POWERED TOOLS FOR TEACHING ESP TO PRE- SERVICE PHYSICS TEACHERS.
  15. El Fathi T, Saad A, Larhzil H, et al. : Integrating generative AI into STEM education: Enhancing conceptual understanding, addressing misconceptions, and assessing student acceptance. Discip. Interdiscip. Sci. Educ. Res. 2025;7(1):6. 10.1186/s43031-025-00125-z [DOI] [Google Scholar]
  16. Elshall AS, Badir A: Balancing AI-assisted learning and traditional assessment : The FACT assessment in environmental data science education Frontiers in Education;2025;10,1596462. 10.3389/feduc.2025.1596462 [DOI] [Google Scholar]
  17. Ezquerra Á, Pamplona S, Casas-Mas A, Nieto-Gómez I: Estudio de emociones en un proceso de ruptura cognitiva a través del reconocimiento facial. Enseñanza de las Ciencias. Revista de investigación y experiencias didácticas ,2024;42(1),23–42. 10.5565/rev/ensciencias.5987 [DOI] [Google Scholar]
  18. Georgiou A, Koutouveli C, Gazeas K, Kasapakis V, Galani A, Lia: Misconceptions in Planetary Geography: A Formative Expert Evaluation of a Serious Game Integrating Augmented Reality and Artificial Intelligence. European Journal of Geography ,2026;17(2),98–113. 10.48088/ejg.a.geo.17.2.098.113 [DOI] [Google Scholar]
  19. Gonçalves Costa GJD, Nascimento Júnior W, Mombelli MN, et al. : Revisiting a Teaching Sequence on the Topic of Electrolysis: A Comparative Study with the Use of Artificial Intelligence. Journal of Chemical Education. 2024;101(8):3255–3263. 10.1021/acs.jchemed.4c00247 [DOI] [Google Scholar]
  20. Hudson B, Gericke N, Olin-Scheller C, et al. : Trajectories of powerful knowledge and epistemic quality: Analysing the transformations from disciplines across school subjects. J. Curric. Stud. 2023;55(2):119–137. 10.1080/00220272.2023.2182164 [DOI] [Google Scholar]
  21. Ji H, Han I, Ko Y: A systematic review of conversational AI in language education: Focusing on the collaboration with human teachers. J. Res. Technol. Educ. 2023;55(1):48–63. 10.1080/15391523.2022.2142873 [DOI] [Google Scholar]
  22. Jose B, Cleetus A, Joseph B, et al. : Epistemic authority and generative AI in learning spaces: Rethinking knowledge in the algorithmic age. Frontiers in Education. 2025;10:1647687. 10.3389/feduc.2025.1647687 [DOI] [Google Scholar]
  23. Kaharu SN, Mansyur J, Werdhiana IK, Tule R: A Comparative Analysis of Mental Models in Floating, Suspending, and Sinking Phenomena: Evidence from Students and Large Language Models (ChatGPT and DeepSeek). European Journal of Educational Research ,2026;15(3),941–962. 10.12973/eu-jer.15.3.941 [DOI] [Google Scholar]
  24. Kamalov F, Santandreu Calonge D, Gurrib I: New Era of Artificial Intelligence in Education: Towards a Sustainable Multifaceted Revolution. Sustainability. 2023;15(16):12451. 10.3390/su151612451 [DOI] [Google Scholar]
  25. Katsenou R, Kotsidis K, Papadopoulou A, et al. : Beyond Assistance: Embracing AI as a Collaborative Co-Agent in Education. Educ. Sci. 2025;15(8):1006. 10.3390/educsci15081006 [DOI] [Google Scholar]
  26. Khor ET, Chan L. Exploring the Effect of Scaffolding Strategies in GenAI Chatbot on Student Engagement and. 2025.
  27. Kim WJ: Teachers’ Use of Generative Artificial Intelligence for Designing Science Lessons in Support of Environmental Science Agency. Research in Science Education. 2026;56(1):203–222. 10.1007/s11165-025-10262-0 [DOI] [Google Scholar]
  28. Kim J, Lee H, Cho YH: Learning design to support student-AI collaboration: Perspectives of leading teachers for AI in education. Educ. Inf. Technol. 2022;27(5):6069–6104. 10.1007/s10639-021-10831-6 [DOI] [Google Scholar]
  29. Kökver Y, Pektaş HM, Çelik H: Artificial intelligence applications in education: Natural language processing in detecting misconceptions. Education and Information Technologies ,2025;30(3),3035–3066. 10.1007/s10639-024-12919-1 [DOI] [Google Scholar]
  30. Kong X, Fang H, Chen W, et al. : Examining human–AI collaboration in hybrid intelligence learning environments: Insight from the Synergy Degree Model. Humanities and Social Sciences Communications. 2025;12(1):821. 10.1057/s41599-025-05097-z [DOI] [Google Scholar]
  31. Kurniawan W, Riantoni C, Lestari N, Ropawandi D: A Hybrid Automatic Scoring System: Artificial Intelligence- Based Evaluation of Physics Concept Comprehension Essay Test. International Journal of Information and Education Technology ,2024;14(6),876–882. 10.18178/ijiet.2024.14.6.2113 [DOI] [Google Scholar]
  32. Lee D, Wright M, Faber C, et al. : Participation in Biology Education Research Influences Students’ Epistemic Development. CBE—Life Sciences Education. 2021;20(4):ar58. 10.1187/cbe.20-11-0255 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Lee J, An T, Chu H-E, Hong H-G, Martin SN: Improving Science Conceptual Understanding and Attitudes in Elementary Science Classes through the Development and Application of a Rule-Based AI Chatbot. Asia-Pacific Science Education ,2023;9(2),365–412. 10.1163/23641177-bja10070 [DOI] [Google Scholar]
  34. Lee SS, Moore RL: Harnessing Generative AI (GenAI) for Automated Feedback in Higher Education: A Systematic Review. Online Learning. 2024;28(3). 10.24059/olj.v28i3.4593 [DOI] [Google Scholar]
  35. Létourneau A, Deslandes Martineau M, Charland P, et al. : A systematic review of AI-driven intelligent tutoring systems (ITS) in K-12 education. Npj Science of Learning. 2025;10(1):29. 10.1038/s41539-025-00320-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Lizano-Sánchez F, Idoyaga IJ, Orduna P, Rodriguez-Gil L, Arguedas-Matarrita C: Students’ interactions with an artificial intelligence assistant in a remote chemistry laboratory. Frontiers in Education ,2025;10,1712743. 10.3389/feduc.2025.1712743 [DOI] [Google Scholar]
  37. Mennella T, Quadros-Mennella P: Student Use, Performance and Perceptions of ChatGPT on College Writing Assignments. Journal of University Teaching and Learning Practice. 2024;21(1). 10.53761/pgwk1a93 [DOI] [Google Scholar]
  38. Morris DL: Rethinking Science Education Practices: Shifting from Investigation-Centric to Comprehensive Inquiry-Based Instruction. Educ. Sci. 2025;15(1):73. 10.3390/educsci15010073 [DOI] [Google Scholar]
  39. Muliyani R, Suhandi A, Prima EC, et al. : PRISMA 2020 Checklist and Flow Diagram for Human–AI Collaboration for Supporting Conceptual Understanding and Conceptual Change in Science Education: A Systematic Literature Review (Version 1.0). Zenodo. 2026. 10.5281/ZENODO.21430817 [DOI]
  40. Nendauni L: Integrating Artificial Intelligence (AI) literacy into writing centre pedagogy: A reflective practitioner’s perspective. ;2025. vol.9:146–156. 10.36615/4c3ydf26 [DOI] [Google Scholar]
  41. Nguyen A, Hong Y, Dang B, et al. : Human-AI collaboration patterns in AI-assisted academic writing. Stud. High. Educ. 2024;49(5):847–864. 10.1080/03075079.2024.2323593 [DOI] [Google Scholar]
  42. Pacaci C, Ustun U, Ozdemir OF: Effectiveness of conceptual change strategies in science education: A meta-analysis. J. Res. Sci. Teach. 2024;61(6):1263–1325. 10.1002/tea.21887 [DOI] [Google Scholar]
  43. Page MJ, McKenzie JE, Bossuyt PM, et al. : The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. J. Clin. Epidemiol. 2021;134:178–189. 10.1016/j.jclinepi.2021.03.001 [DOI] [PubMed] [Google Scholar]
  44. Pektaş HM, Karamustafaoğlu O, Çelik H: The Role of Educational Data Mining and Artificial Intelligence Supported Learning Analytics on Conceptual Change: New Approaches to Differentiated Instruction. J. Sci. Educ. Technol. 2026;35(1):1–26. 10.1007/s10956-025-10214-3 [DOI] [Google Scholar]
  45. Relmasira SC, Donaldson JP: Deep Learning Pedagogies Enhance AI Literacy in Elementary Students: A Five-Cycle Implementation Study. Online Learning In Educational Research (OLER) ,2025;5(1),143–154. 10.58524/oler.v5i1.720 [DOI] [Google Scholar]
  46. Sánchez Muñoz JA, Flores-Eraña G, Silva-Campos JM, et al. : GenAI as a cognitive mediator: A critical-constructivist inquiry into computational thinking in pre-university education. Frontiers in Education. 2025;10:1597249. 10.3389/feduc.2025.1597249 [DOI] [Google Scholar]
  47. Shih C-C, She H-C, Huang L-Y, et al. : Electroencephalography (EEG) Predicts the Success of Scientific Conceptual Changes Involving Weak and Radical Restructuring. J. Sci. Educ. Technol. 2026;35(1):160–174. 10.1007/s10956-025-10243-y [DOI] [Google Scholar]
  48. Thomas DR, Lin J, Gatz E, et al. : Improving Student Learning with Hybrid Human-AI Tutoring: A Three-Study Quasi-Experimental Investigation. Proceedings of the 14th Learning Analytics and Knowledge Conference. 2024;404–415. 10.1145/3636555.3636896 [DOI]
  49. Tsakeni M, Nwafor SC, Mosia M, et al. : Mapping the Scaffolding of Metacognition and Learning by AI Tools in STEM Classrooms: A Bibliometric–Systematic Review Approach (2005–2025). J. Intelligence. 2025;13(11):148. 10.3390/jintelligence13110148 [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Vaccaro M, Almaatouq A, Malone T: When combinations of humans and AI are useful: A systematic review and meta-analysis. Nat. Hum. Behav. 2024;8(12):2293–2303. 10.1038/s41562-024-02024-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Walter Y: Embracing the future of Artificial Intelligence in the classroom: The relevance of AI literacy, prompt engineering, and critical thinking in modern education. Int. J. Educ. Technol. High. Educ. 2024;21(1):15. 10.1186/s41239-024-00448-3 [DOI] [Google Scholar]
  52. Wheaton M, Ardoin NM, Bowers AW, et al. : Sociocultural learning theories for social-ecological change. Environ. Educ. Res. 2024;30(8):1193–1210. 10.1080/13504622.2024.2347888 [DOI] [Google Scholar]
  53. Wu T, Zhai X, Song Y: The effects of GAI-enhanced pedagogical agents in the metaverse (GPAiM) on elementary school students’ conceptual understanding and cognitive engagement patterns. Computers & Education ,2026;245,105555. 10.1016/j.compedu.2025.105555 [DOI] [Google Scholar]
  54. Yan L, Greiff S, Teuber Z, et al. : Promises and challenges of generative artificial intelligence for human learning. Nat. Hum. Behav. 2024;8(10):1839–1850. 10.1038/s41562-024-02004-5 [DOI] [PubMed] [Google Scholar]
  55. Yusuf H, Money A, Daylamani-Zad D: Pedagogical AI conversational agents in higher education: A conceptual framework and survey of the state of the art. Educ. Technol. Res. Dev. 2025;73(2):815–874. 10.1007/s11423-025-10447-4 [DOI] [Google Scholar]
  56. Zhai C, Wibowo S, Li LD: The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: A systematic review. Smart Learning Environments. 2024;11(1):28. 10.1186/s40561-024-00316-7 [DOI] [Google Scholar]
  57. Zhu G, Scardamalia M, Moreno M, et al. : Discourse Moves and Emotion in Knowledge Building Discourse and Metadiscourse. Frontiers in Education. 2022;7:900440. 10.3389/feduc.2022.900440 [DOI] [Google Scholar]
  58. Marín Díaz G: Integrating Exploratory Data Analysis and Explainable AI into Astronomy Education: A Fuzzy Approach to Data-Literate Learning. Education Sciences. 2025;15(12):1688. 10.3390/educsci15121688 [DOI] [Google Scholar]
  59. Serik M, Tleuzhanova K, Nurgaliyeva S: Enhancing Master’s-Level STEM Education through AI-Driven IoT Projects: A Kazakhstan Experiment. International Journal of Information and Education Technology ,2025;15(10),2086–2094. 10.18178/ijiet.2025.15.10.2407 [DOI] [Google Scholar]
  60. Zha S, Bragdon MM, Gong N, Wang J, Leavesley S, Eaton R, Bosarge E: A Case Study of Integrating AI Literacy Education in a Biology Class. International Journal of Artificial Intelligence in Education ,2025;35(4),2453–2477. 10.1007/s40593-025-00476-8 [DOI] [Google Scholar]
F1000Res. 2026 Sep 24. doi: 10.5256/f1000research.206380.r520683

Reviewer response for version 1

Prima Aswirna 1

1. Absence of an Operational Definition for "Human-AI Collaboration"

The manuscript positions "Human-AI Collaboration" as its central paradigm but fails to explicitly define what constitutes genuine collaboration versus routine AI usage (e.g., intelligent tutoring, automated feedback, or simple prompt-response interactions). Without distinct inclusion criteria for collaborative dynamics, standard AI-assisted learning tools are conflated with collaborative partnerships.

2. Blurred Distinction Between Conceptual Understanding and Conceptual Change

The text treats conceptual understanding and conceptual change almost interchangeably at times, despite their fundamental theoretical differences. Conceptual understanding involves constructing and enriching accurate scientific representations, whereas conceptual change entails restructuring, revising, or replacing entrenched misconceptions. The review does not clearly state whether included studies explicitly measured conceptual change or if change was merely inferred from higher understanding scores.

3. Single-Database Search Bias (Scopus Only)

Limiting the systematic literature search exclusively to Scopus introduces significant selection bias. Given the interdisciplinary nature of AI in science education—spanning computer science, educational technology, and cognitive science—key databases such as Web of Science, ERIC, IEEE Xplore, Education Research Complete, and PsycINFO should have been included.

4. Unjustified Temporal Cutoff (2023–2026)

Restricting the review to studies published from 2023 onwards excludes foundational pre-2023 empirical research on intelligent tutoring systems, automated scaffolding, and conversational agents in science education. The authors provide no methodological rationale explaining why earlier empirical work on Human-AI collaboration was omitted.

5. Lack of Protocol Preregistration

The review protocol was not preregistered in public registries such as PROSPERO or Open Science Framework (OSF). Preregistration is a core standard of PRISMA guidelines to prevent post-hoc changes to inclusion criteria and selective outcome reporting.

6. Absence of Quality Appraisal / Risk of Bias Assessment

The review does not conduct a formal critical appraisal or risk of bias evaluation of the 28 included empirical studies. Without evaluating study quality (e.g., sample size validity, control group rigor, measurement instrument reliability), the strength of the synthesized evidence remains unverified.

7. Omission of Inter-Rater Reliability

The authors state that screening and selection were conducted, but fail to report whether independent dual screening occurred or what inter-rater agreement metrics (e.g., Cohen’s $\kappa$ or percentage agreement) were achieved. This leaves the screening process vulnerable to individual selection bias.

8. Single-Coder Bias in Thematic Analysis

Section 2.6 indicates that thematic coding was performed "by the author" (singular). Conducting qualitative coding without a second independent coder or inter-coder reliability checks undermines the trustworthiness and objectivity of the derived themes and mechanisms.

9. Overly Causal Claims Based on Non-Causal Evidence

The review claims that "these processes are brought about by a variety of learning mechanisms" (p. 13). Asserting direct causality is methodologically inappropriate because the included studies consist of heterogeneous qualitative, design-based, and quasi-experimental research rather than pooled randomized controlled trials or meta-analyses.

10. Ambiguous and Unsubstantiated Assertions ("As a Matter of Fact")

In Section 4, the authors state that "it seems to be more appropriate to treat the processes of students' conceptual understanding and their conceptual change as a matter of fact" (p. 13). This phrasing is scientifically ambiguous and overstates what can be inferred from a synthesis of 28 studies.

11. Unsubstantiated Introductory Claims

The opening statement that Human-AI Collaboration "is gaining momentum" is presented as a premise without supporting bibliometric data, publication growth rates, or citation trend analyses.

12. Theoretical Eclecticism Without Integration Rationale

The proposed framework combines four major learning theories: Constructivism, Sociocultural Theory, Inquiry-Based Learning, and Metacognitive Theory. However, the paper does not adequately explain how these distinct traditions—which hold differing assumptions about individual cognition versus social interaction—conceptually align or why alternative learning models were excluded.

13. Undefined Interrelationships Among Framework Components

Figure 2 presents five learning mechanisms as separate pillars beneath learning outcomes. The manuscript fails to specify whether these mechanisms operate sequentially, iteratively, conditionally, or hierarchically during student-AI interactions.

14. Undifferentiated Aggregation of AI Technologies

The synthesis groups widely divergent AI technologies—ranging from rule-based chatbots and intelligent tutoring systems to Large Language Models (LLMs), machine learning models, and IoT platforms—under the single banner of "AI". Because LLMs possess vastly different affordances and hallucination risks compared to rule-based tutors, treating them as functionally equivalent obscures tool-specific impacts.

15. Anthropomorphic Framing of AI as a "Learning Partner"

The manuscript repeatedly characterizes AI as a "learning partner" possessing epistemic authority. Framing non-sentient software as an active agent risks misrepresenting computational prompt-response loops as genuine social/collaborative reciprocity.

16. Numerical and Structural Inconsistencies

The paper asserts that it identified "six roles of AI" in supporting conceptual understanding and "five mechanisms" supporting conceptual change, but elsewhere presents five learning mechanisms connecting the two. The structural mapping between the 6 roles (Table 3) and the 5 mechanisms (Tables 4 & 5) is left visually and conceptually disconnected.

17. PRISMA Reporting Discrepancies

Although the text claims alignment with PRISMA 2020 guidelines, key details (such as detailed exclusion breakdown numbers per screening stage) are omitted from the main text and relegated to an external Zenodo link rather than being fully articulated within the body of the article.

18. Narrow and Vulnerable Search Query Strategy

The search string TITLE-ABS-KEY (("artificial intelligence" OR "generative AI" OR "ChatGPT" OR "large language model" OR "AI chatbot") AND ("conceptual understanding" OR "concept learning" OR "conceptual change" OR "misconception")) omits broader terms like "cognitive conflict," "schema restructuring," "adaptive scaffolding," or specific domain terms. Furthermore, searching without domain-specific keywords forced reliance on post-hoc manual filtering to isolate science education studies.

19. Search Date Ambiguity

The manuscript states that the Scopus search was "last updated on 15 June 2026". It remains unclear whether this date reflects the initial search or a re-run, leaving the exact timeline and reproducibility of record retrieval ambiguous.

20. Overrepresentation and Domain Bias Toward Physics

According to Table 2, Physics accounts for 32.14% of the studies, while Chemistry (10.72%) and Biology (7.14%) are severely underrepresented. The authors do not address how this domain skew impacts the generalizability of their framework across other scientific disciplines with differing conceptual structures.

21. Methodological Vagueness in Narrative Synthesis

Section 2.6 mentions that a "narrative synthesis of the coding results was conducted", but fails to define the specific synthesis protocol used (e.g., Popay et al.'s framework for narrative synthesis). This leaves the step-by-step translation from empirical findings to framework components unverified.

22. Incomplete Categorization in Descriptive Statistics

In Table 2, under "Research Design," categories such as "Other Designs" (17.86%) are left undefined. Furthermore, criteria for distinguishing overlapping categories (e.g., "Mixed Methods Research" vs. "Experimental Research") are omitted.

23. Failure to Assess Publication Bias

The review does not discuss or account for publication bias. Given that studies publishing positive learning gains from novel AI implementations are more likely to be accepted than null results, the review likely overestimates the efficacy of AI in fostering conceptual change.

24. Lack of Practical and Instructional Translation

While the article presents a high-level conceptual framework (Figure 2), it provides minimal guidance on how educators or instructional designers can operationalize these mechanisms in actual classroom settings. It lacks practical prompt structures, activity designs, or curriculum integration protocols.

25. Grammatical Errors and Structural Redundancies

The text contains several grammatical and stylistic flaws that detract from its academic precision, such as "a total of 28 empirical studies research were analyzed" (p. 1) and repetitive phrasing like "These studies of empirical research." Minor typographical errors in headings (e.g., "Knowledge Exprolation & Construction" in Figure 2) also remain uncorrected.

Are the rationale for, and objectives of, the Systematic Review clearly stated?

Partly

Is the statistical analysis and its interpretation appropriate?

Partly

If this is a Living Systematic Review, is the ‘living’ method appropriate and is the search schedule clearly defined and justified? (‘Living Systematic Review’ or a variation of this term should be included in the title.)

Partly

Are sufficient details of the methods and analysis provided to allow replication by others?

Partly

Are the conclusions drawn adequately supported by the results presented in the review?

Yes

Reviewer Expertise:

Educational Science and AI in Education Science.

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.

F1000Res. 2026 Sep 15. doi: 10.5256/f1000research.206380.r520675

Reviewer response for version 1

Torang Siregar 1

Peer Review Report

The manuscript presents a systematic literature review (SLR) investigating how Human-AI Collaboration supports students' conceptual understanding and conceptual change in science education. The authors analyzed 28 empirical studies published between 2023 and 2026, identified through a Scopus database search. The review identifies six roles of AI supporting conceptual understanding (AI tutor, inquiry facilitator, feedback assistant, visualization assistant, dialogue partner, and co-creator) and five mechanisms supporting conceptual change (misconception diagnosis, adaptive scaffolding, reflective dialogue, cognitive conflict, and fact-checking). The authors propose an integrated conceptual framework grounded in constructivism, sociocultural theory, inquiry-based learning, and metacognitive theory, positioning AI as a "learning partner" rather than merely an instructional tool. The framework illustrates how students continuously construct, evaluate, and refine scientific knowledge through interconnected learning mechanisms.

Answers to Review Questions

1. Are the rationale for, and objectives of, the Systematic Review clearly stated?

Strengths:

  1. The introduction provides a reasonable overview of the growing role of AI in education and the shift from AI as a tool to AI as an epistemic partner.

  2. The four research questions (RQ1–RQ4) are explicitly stated and map onto the review's objectives.

  3. The authors identify a genuine gap: most existing reviews focus on AI in education broadly, but few examine Human-AI Collaboration specifically in relation to conceptual understanding and conceptual change.

Weaknesses and Required Revisions:

  1. The rationale for focusing exclusively on "Human-AI Collaboration" is underdeveloped. The introduction does not adequately distinguish Human-AI Collaboration from related constructs such as AI-assisted learning, AI-supported learning, or intelligent tutoring. The authors assert that Human-AI Collaboration represents "a new learning paradigm" (p. 3) but do not provide a clear operational definition or criteria for what constitutes "collaboration" versus mere "support" or "assistance." This is a critical conceptual gap that undermines the review's coherence.

  2. The claim that Human-AI Collaboration "is gaining momentum" lacks empirical support. This is a plausible assertion, but it is presented without citation to bibliometric data, trend analyses, or prior reviews that would substantiate it. The authors should provide evidence (e.g., publication trends, citation patterns) to support this claim.

  3. The gap statement is somewhat overstated. The authors claim that "no systematic literature review has yet developed an integrated theoretical framework that explains how Human AI Collaboration can support learners' construction of scientific knowledge" (p. 4). While this may be true for the specific intersection of Human-AI Collaboration, conceptual understanding, and conceptual change in science education, the authors should acknowledge related frameworks in the learning sciences and AI in education literature (e.g., knowledge building, computer-supported collaborative learning) that may offer partial integration.

  4. The rationale for excluding studies before 2023 is not justified. Given the long history of intelligent tutoring systems, adaptive learning, and conversational agents in science education, the decision to restrict the review to 2023–2026 requires explicit justification. The authors should explain whether earlier research was excluded because it did not meet the Human-AI Collaboration criteria or because of a deliberate decision to focus on recent developments.

  5. The objectives of the review could be more precisely articulated. While the research questions are clear, the introduction does not explicitly state what the review aims to contribute beyond the framework itself—for example, whether it aims to inform instructional design, guide future research, or provide a theoretical synthesis.

Recommendations:

  1. Provide an explicit operational definition of Human-AI Collaboration and distinguish it from related constructs.

  2. Support the "gaining momentum" claim with bibliometric or trend evidence.

  3. Justify the 2023–2026 publication window.

  4. Clarify the review's intended contribution and audience.

  5. Acknowledge related frameworks and explain how the proposed framework extends or differs from them.

2. Are sufficient details of the methods and analysis provided to allow replication by others?

Strengths:

  1. The authors report following PRISMA 2020 guidelines and provide inclusion/exclusion criteria based on the PICOS framework.

  2. The search string is explicitly stated.

  3. The study selection process is described in three stages (title, abstract, full-text screening).

  4. A PRISMA flow diagram is referenced (Figure 1).

Weaknesses and Required Revisions:

  1. The review protocol was not preregistered. The authors do not report whether the review protocol was registered (e.g., PROSPERO, OSF, or another registry). Preregistration is a hallmark of rigorous systematic reviews and should be reported. If the review was not preregistered, this limitation should be acknowledged.

  2. Database selection is unjustifiably narrow. The review relies exclusively on Scopus. While Scopus is a large and reputable database, a single-database search introduces database-selection bias, particularly in an interdisciplinary area spanning education, learning sciences, computer science, and AI. The authors should either (a) expand the search to include Web of Science, ERIC, Education Research Complete, PsycINFO, IEEE Xplore, or other relevant databases, or (b) provide a strong justification for using Scopus alone and acknowledge this as a limitation.

  3. Inter-rater reliability is not reported. The authors state that screening was conducted but do not report whether multiple reviewers independently screened studies, how disagreements were resolved, or what the inter-rater reliability was. This is a significant methodological omission. At minimum, the authors should report Cohen's kappa or percentage agreement for screening decisions.

  4. The PRISMA flow diagram is referenced but not provided in the manuscript text. The authors state that the PRISMA checklist and flow diagram are available on Zenodo, but the flow diagram should be included in the manuscript (or at least summarized in text) to show the number of records identified, screened, excluded, and included, along with reasons for exclusion. The current manuscript does not report these numbers.

  5. Search date ambiguity. The authors state the search was "last updated on 15 June 2026" (p. 4). It is unclear whether this was the original search date or an update. If it was an update, the original search date should be reported. This distinction is important for reproducibility and for understanding the currency of the evidence base.

  6. Data extraction process lacks detail. The authors state that data were extracted into an Excel file with four main parts, but do not report:
    1. Who performed the extraction (one author or multiple)?
    2. Whether extraction was piloted or calibrated?
    3. How discrepancies were resolved?
    4. Whether authors of primary studies were contacted for missing data?
  7. Thematic coding process is underdescribed. The authors state that thematic coding was conducted "by the author" (singular) (p. 7). This raises concerns about bias and reliability. The authors should report:
    1. Whether coding was inductive, deductive, or abductive.
    2. How the coding framework was developed and refined.
    3. Whether a second coder was involved and, if so, how reliability was established.
    4. How themes were derived and validated.

  8. Descriptive statistics are reported but not fully explained. Table 2 presents frequencies and percentages for publication year, education level, science domain, AI technology, and research design. However, the authors do not explain how studies were categorized (e.g., how "Mixed Methods Research" was distinguished from "Experimental Research") or how many studies fell into each category for certain variables (e.g., the table shows percentages but not always clear n values).

  9. Quality appraisal of included studies is absent. The authors do not report conducting a risk of bias or quality appraisal of the included studies. While quality appraisal is not always required in systematic reviews of this type, it is strongly recommended and would strengthen the review's credibility. If quality appraisal was not conducted, this limitation should be acknowledged.

  10. Synthesis methods are insufficiently described. The authors state that "a narrative synthesis of the coding results was conducted" (p. 7), but do not describe how the synthesis was performed, how themes were compared across studies, or how the integrated framework was derived from the synthesis. A more detailed account of the analytical process is needed.

Recommendations:

  1. Report whether the review protocol was preregistered and provide a registration number or URL if available.

  2. Expand the database search or justify the use of Scopus alone and acknowledge the limitation.

  3. Report inter-rater reliability for screening and coding.

  4. Include the PRISMA flow diagram in the manuscript and report the number of records at each stage, with reasons for exclusion.

  5. Clarify the search date(s) and whether the search was updated.

  6. Provide more detail on data extraction and thematic coding, including who performed these tasks and how reliability was ensured.

  7. Report quality appraisal of included studies or acknowledge its absence as a limitation.

  8. Describe the synthesis process in greater detail, including how the integrated framework was derived.

3. Is the statistical analysis and its interpretation appropriate?

This is a systematic literature review that employs narrative synthesis and thematic coding. No primary statistical analysis was conducted. The descriptive statistics presented in Table 2 (frequencies and percentages) are appropriate for summarizing study characteristics, but they are not inferential statistics and should not be interpreted as such. The authors correctly avoid causal or inferential claims based on these descriptive data.

Recommendation:

  1. Clarify in the methods section that the review employs narrative synthesis and descriptive statistics only, and that no meta-analysis was conducted (and why).

4. Are the conclusions drawn adequately supported by the results presented in the review?

Strengths:

  1. The conclusions are broadly consistent with the findings presented in Tables 3–5 and Figure 2.

  2. The authors appropriately call for future research to move beyond isolated AI functions toward integrated, long-term studies of Human-AI Collaboration.

  3. The proposed framework is clearly presented and grounded in the reviewed literature.

Weaknesses and Required Revisions:

  1. Causal language is used without sufficient evidence. The authors state that "these processes are brought about by a variety of learning mechanisms" (p. 13). This is causal language that implies a deterministic relationship between mechanisms and outcomes. Given that the review synthesizes primarily correlational, quasi-experimental, and design-based studies—and does not conduct a meta-analysis or assess causal inference—this language is too strong. The authors should use more cautious phrasing such as "may be supported by," "appear to be facilitated by," or "are associated with."

  2. The claim that the framework "explains" how Human-AI Collaboration supports learning is overstated. The framework is a conceptual synthesis, not an empirically validated causal model. The authors should frame it as a "heuristic" or "proposed" framework that requires empirical testing, rather than as an explanatory model.

  3. The statement "It seems to be more appropriate to treat the processes of students' conceptual understanding and their conceptual change as a matter of fact" (p. 13) is unclear and problematic. The phrase "as a matter of fact" is ambiguous and appears to make a stronger claim than the evidence warrants. The authors should revise this passage to clearly state their intended argument—likely that conceptual understanding and conceptual change should be conceptualized as emergent, iterative processes rather than discrete instructional outcomes—and support this with evidence from the review.

  4. The conclusion does not adequately acknowledge the limitations of the review. The authors should discuss:
    1. The narrow database coverage (Scopus only).
    2. The exclusion of pre-2023 studies.
    3. The lack of quality appraisal.
    4. The absence of preregistration.
    5. The potential for publication bias.
    6. The limited generalizability of the framework given the small number of studies (n = 28) and their concentration in certain science domains (e.g., physics = 32.14%).

  5. The framework's theoretical foundations are asserted rather than justified. The authors state that the framework is based on constructivism, sociocultural theory, inquiry-based learning, and metacognitive theory (p. 12). However, they do not explain why these particular theories were chosen, how they were identified from the reviewed studies, or how they integrate with one another. A stronger theoretical justification is needed.

  6. The framework's components are not clearly distinguished. The framework (Figure 2) presents five learning mechanisms (Adaptive Support, Knowledge Exploration and Construction, Reflection and Dialogue, Cognitive Conflict, Information Evaluation and Validation) and links them to learning outcomes. However, the relationships between these mechanisms—whether they are sequential, iterative, or interdependent—are not clearly explained. The authors should clarify the nature of these relationships and provide a more detailed narrative description of how the framework operates.

  7. The role of AI in the framework is underspecified. The framework positions AI as a "learning partner," but does not clearly distinguish between different types of AI (e.g., LLMs, intelligent tutoring systems, machine learning) or explain how the framework applies across different AI technologies. Given the diversity of AI tools in the reviewed studies, the framework should account for this heterogeneity.

Recommendations:

  1. Replace causal language with more cautious, associational phrasing.

  2. Frame the framework as a proposed heuristic requiring empirical validation, not as an explanatory model.

  3. Revise the "matter of fact" passage for clarity and accuracy.

  4. Add a limitations section that acknowledges the review's methodological constraints.

  5. Provide stronger justification for the theoretical foundations and clarify how the theories integrate.

  6. Clarify the relationships between framework components and the role of different AI technologies.

  7. Ensure that conclusions are explicitly tied to the evidence presented in Tables 3–5 and Figure 2.

5. If this is a Living Systematic Review, is the 'living' method appropriate and is the search schedule clearly defined and justified?

The manuscript does not describe itself as a "Living Systematic Review," nor does it use this term in the title or text. Therefore, this question is not applicable. However, the authors may wish to consider whether a living systematic review approach would be appropriate given the rapidly evolving nature of AI in education. If they choose to adopt this approach, they should include "Living Systematic Review" in the title and clearly define the search schedule and update criteria.

Summary of Required Revisions

Must Be Addressed (Scientific Soundness):

  1. Provide an explicit operational definition of Human-AI Collaboration and distinguish it from related constructs.

  2. Report inter-rater reliability for screening and coding, or acknowledge its absence as a limitation.

  3. Include the PRISMA flow diagram in the manuscript and report the number of records at each stage, with reasons for exclusion.

  4. Expand the database search or justify the use of Scopus alone and acknowledge the limitation.

  5. Report whether the review protocol was preregistered and provide registration details if available.

  6. Clarify the search date(s) and whether the search was updated.

  7. Replace causal language with cautious, associational phrasing in the conclusions.

  8. Revise the "matter of fact" passage for clarity and accuracy.

  9. Add a limitations section that acknowledges the review's methodological constraints.

  10. Provide a more detailed description of the thematic coding and synthesis process, including who performed these tasks and how reliability was ensured.

Should Be Addressed (Strengthening the Manuscript):

  1. Justify the 2023–2026 publication window.

  2. Support the "gaining momentum" claim with bibliometric or trend evidence.

  3. Acknowledge related frameworks and explain how the proposed framework extends or differs from them.

  4. Report quality appraisal of included studies or acknowledge its absence as a limitation.

  5. Clarify the theoretical foundations of the framework and how the theories integrate.

  6. Clarify the relationships between framework components and the role of different AI technologies.

  7. Improve the grammatical and stylistic quality of the manuscript (e.g., "a total of 28 empirical studies research were analyzed" should be corrected to "a total of 28 empirical studies were included in the analysis").

  8. Consider whether a living systematic review approach would be appropriate given the rapidly evolving field.

Conclusion

The manuscript addresses a timely and important topic, and the proposed integrated conceptual framework represents a valuable contribution to the literature. However, the review has several methodological limitations that must be addressed to ensure scientific soundness. The most critical issues are the lack of inter-rater reliability reporting, the absence of the PRISMA flow diagram in the manuscript, the narrow database coverage, and the use of causal language that exceeds the evidence. Addressing these issues will significantly strengthen the manuscript and enhance its contribution to the field.

Are the rationale for, and objectives of, the Systematic Review clearly stated?

Partly

Is the statistical analysis and its interpretation appropriate?

Not applicable

If this is a Living Systematic Review, is the ‘living’ method appropriate and is the search schedule clearly defined and justified? (‘Living Systematic Review’ or a variation of this term should be included in the title.)

Not applicable

Are sufficient details of the methods and analysis provided to allow replication by others?

No

Are the conclusions drawn adequately supported by the results presented in the review?

Partly

Reviewer Expertise:

Educational Technology; Artificial Intelligence in Education; Science Education; Systematic Review Methodology; Learning Sciences

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.

F1000Res. 2026 Sep 2. doi: 10.5256/f1000research.206380.r515589

Reviewer response for version 1

Nikolaos Pellas 1

Overall assessment

The text addresses a timely and potentially valuable topic, and the attempt to move beyond tool-specific evaluations toward learning mechanisms underlying Human–AI collaboration is a clear strength. The proposed integrated conceptual framework could make a useful contribution. However, several aspects require clarification, particularly the definition of Human–AI collaboration, the distinction between conceptual understanding and conceptual change, the derivation and validation of the proposed mechanisms, and the methodological reporting of the systematic review.

Major comments and specific questions

1. Clarify the central concept of “Human–AI Collaboration.”

The manuscript repeatedly uses “Human AI Collaboration,” but the concept does not appear to be sufficiently operationalized.

Question:

How do the authors define Human–AI collaboration in this review, and what criteria were used to distinguish genuine collaboration from more conventional uses of AI, such as AI-assisted tutoring, automated feedback, question answering, or content generation?

Suggestion:

Provide an explicit conceptual definition and inclusion criteria. It would also be useful to explain what makes the interaction collaborative rather than simply AI-supported learning.

2. The opening claim needs stronger support.

The statement that Human–AI collaboration in science education “is gaining momentum” is plausible, but it should be supported by recent evidence.

Question:

What evidence demonstrates that Human–AI collaboration is specifically gaining momentum, rather than AI in education more generally?

3. There is a conceptual ambiguity between conceptual understanding and conceptual change.

The manuscript treats these concepts as closely related, but they are not necessarily interchangeable. Conceptual understanding may involve constructing or deepening scientifically accurate representations, whereas conceptual change often concerns restructuring or replacing prior conceptions.

Question:

How are conceptual understanding and conceptual change operationally defined and distinguished in the review? Did the included studies explicitly measure conceptual change, or did the authors infer conceptual change from improvements in conceptual understanding?

This distinction is particularly important because the manuscript later concludes that these processes are “brought about by a variety of learning mechanisms.”

4. The reported number of learning mechanisms is internally inconsistent.

The abstract/text states that “five mechanisms supporting the students’ conceptual change were identified,” followed by “Five interconnected learning mechanisms were derived.” However, the manuscript also states that six roles of AI were identified.

Question:

Are there exactly five learning mechanisms and six AI roles? What is the relationship between these two categories?

Suggestion:

A concise table showing AI role → learning mechanism → expected effect on conceptual understanding/conceptual change → supporting studies would substantially improve clarity.

5. The derivation of the framework requires more methodological transparency.

The manuscript states that an integrated conceptual framework was derived from the included studies and is based on constructivism, sociocultural theory, inquiry-based science learning, and metacognitive theory.

Questions:

How were the five mechanisms inductively or deductively identified?

Were the theoretical perspectives specified before data extraction, or were they introduced after the patterns emerged from the literature?

How did the authors avoid imposing these theories onto the findings?

This is important because the framework appears to combine several theoretical traditions that have different assumptions about learning.

6. The theoretical integration needs further justification.

The framework draws simultaneously on constructivism, sociocultural theory, inquiry-based learning, and metacognitive theory. This could be a strength, but currently the rationale for combining them is underdeveloped.

Question:

What unique explanatory contribution does each theoretical perspective make to the proposed framework, and why are these four perspectives necessary rather than alternative theoretical lenses?

7. The phrase “AI functions as a learning partner” may be too strong.

The wording suggests agency and reciprocity on the part of AI. Many AI systems used in education may provide responses or feedback without engaging in genuinely reciprocal collaborative processes.

Question:

What evidence from the included studies supports characterizing AI as a learning partner rather than as a pedagogical tool, tutor, scaffold, or conversational agent?

Suggestion:

The authors could distinguish between functional partnership and human-like agency to avoid anthropomorphic interpretations.

8. Clarify the scope of the evidence base.

The review is based exclusively on Scopus.

Question:

Why was Scopus selected as the sole database? Were Web of Science, ERIC, Education Research Complete, PsycINFO, IEEE Xplore, or other relevant databases considered?

A single-database search may introduce database-selection bias, particularly in an interdisciplinary area spanning education, learning sciences, computer science, and AI.

9. The PRISMA statement needs methodological detail.

The statement “In line with the PRISMA 2020 guidelines” is insufficient by itself.

Questions:

Was the review protocol preregistered?

Was a PRISMA 2020 flow diagram provided?

How many records were identified, screened, excluded, and included?

What were the principal exclusion reasons?

For a systematic review, these details are important for reproducibility.

10. The wording concerning the 28 studies should be corrected.

The sentence “a total of 28 empirical studies research were analyzed” is grammatically incorrect and conceptually awkward.

A clearer formulation would be:

“A total of 28 empirical studies were included in the analysis.”

The phrase “These studies of empirical research” is similarly repetitive.

11. The date of the search needs clarification.

The manuscript states that the Scopus search was “last updated on 15 June 2026.”

Question:

Does this mean the original search was conducted on that date, or was the search initially conducted earlier and subsequently updated? If it was an update, when was the original search performed?

This distinction is important for reporting the search strategy accurately.

12. The publication period raises an important question about the evidence base.

The included studies were published between 2023 and 2026.

Question:

What search and eligibility criteria resulted in no eligible empirical studies before 2023? Given the long history of intelligent tutoring systems, adaptive learning, computer-supported collaborative learning, and conversational agents, the authors should explain whether earlier research was explicitly excluded or whether it did not satisfy the Human–AI collaboration criteria.

13. “Conceptual change” appears to be treated partly as an outcome and partly as a process.

The manuscript refers to conceptual understanding and conceptual change as processes, but elsewhere discusses mechanisms “supporting” them.

Question:

Does the framework conceptualize conceptual change as an outcome, a cognitive process, or both? How is this represented in the framework?

A clear distinction between AI affordances, interaction processes, learning mechanisms, cognitive processes, and learning outcomes would strengthen the model considerably.

14. The evidence supporting causal language should be examined carefully.

The statement:

“these processes are brought about by a variety of learning mechanisms”

appears causal.

Question:

Do the included studies provide sufficient experimental or longitudinal evidence to conclude that these mechanisms bring about conceptual change, or do they primarily indicate associations and plausible explanatory mechanisms?

Given that this is a literature review, more cautious language such as “may be supported by”, “appear to be facilitated by”, or “are associated with” may be appropriate unless causal evidence is explicitly available.

15. The final interpretation appears stronger than the evidence presented.

The statement:

“it seems to be more appropriate to treat the processes of students’ conceptual understanding and their conceptual change as a matter of fact”

is unclear and potentially problematic.

Question:

What does “as a matter of fact” mean in this context? Is the intended argument that conceptual understanding and conceptual change should be conceptualized as emergent learning processes rather than predefined instructional outcomes?

The authors should revise this passage because the current wording is difficult to interpret and appears stronger than the evidence warrants.

Are the rationale for, and objectives of, the Systematic Review clearly stated?

Partly

Is the statistical analysis and its interpretation appropriate?

Not applicable

If this is a Living Systematic Review, is the ‘living’ method appropriate and is the search schedule clearly defined and justified? (‘Living Systematic Review’ or a variation of this term should be included in the title.)

Partly

Are sufficient details of the methods and analysis provided to allow replication by others?

No

Are the conclusions drawn adequately supported by the results presented in the review?

Partly

Reviewer Expertise:

Educational Technology

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Data Availability Statement

    The completed PRISMA 2020 Checklist and Flow Diagram for Human AI Collaboration for Supporting Conceptual Understanding and Conceptual Change in Science Education: A Systematic Literature Review has been uploaded to Zenodo and published under Creative Commons Zero v1.0 Universal (CC0) license. Checklist and Flow Diagram can be downloaded online at https://doi.org/10.5281/zenodo.21430817 ( Muliyani et al., 2026).

    Articles from F1000Research are provided here courtesy of F1000 Research Ltd

    RESOURCES