Changes in Pre-Service Physics Teachers’ TPACK and Collaborative Problem Solving Associated with an AI-Supported CTD-PBL Module: A Quasi-Experimental Study

Abstract

Generative artificial intelligence (AI) is increasingly entering teacher education, yet evidence remains limited on its responsible integration into discipline-specific pedagogical preparation. This study examined whether an AI-supported Collaborative TPACK Competency Development module based on problem-based learning (CTD-PBL) was associated with greater pre–post gains in pre-service physics teachers’ self-reported technological pedagogical content knowledge (TPACK) and perceived collaborative problem-solving (CPS) processes. Informed by ADDIE, the 8-week module used DeepSeek as a bounded scaffold for collaborative lesson design, feedback, verification, and reflective revision while preserving teacher judgment. An intact-class quasi-experimental pre-test/post-test design involved 130 third-year pre-service physics teachers at a public university in western China. Two existing classes were randomly allocated at the class level to CTD-PBL or conventional instruction. Compared with the conventional group, the CTD-PBL group reported higher post-test TPACK scores (M = 4.04 vs. M = 3.40, p < 0.001, d = 1.02) and higher perceived CPS process scores (M = 3.62 vs. M = 3.05, p < 0.001, d = 0.88), with stronger pre–post gains in both outcomes. The findings provide a bounded curriculum design case showing how generative AI can be embedded in physics teacher education through problem-based tasks, collaborative scaffolding, and human verification procedures.

1. Introduction

Generative artificial intelligence is increasingly being integrated into teacher education as a resource for lesson planning, instructional decision-making, feedback generation, and reflective improvement [1,2,3,4]. Its educational significance does not lie simply in faster access to information, but in the possibility of reorganising how pre-service teachers engage with curriculum materials, analyse learners’ difficulties, design classroom tasks, and revise pedagogical representations [5,6,7]. In this sense, AI integration in teacher education should be understood as a curriculum and pedagogy issue, not merely as tool adoption [8,9]. At present, however, the empirical base remains uneven [10,11,12]. Many studies discuss teachers’ perceptions of AI, attitudes toward digital tools, or the affordances of generative systems, but fewer provide intervention-based evidence showing how AI can be embedded in teacher education in a discipline-specific, replicable, and competence-oriented manner [13]. This limitation is important because teacher education cannot treat AI as a neutral productivity device [14]. If AI is introduced without clear instructional purposes, task structures, review procedures, and ethical boundaries, it may encourage superficial lesson design, unexamined dependence on generated outputs, or the replacement of professional judgement by automated suggestions [15]. What is needed, therefore, is evidence on AI-supported curriculum modules that position generative AI as a pedagogical scaffold: a resource that helps pre-service teachers search, compare, design, receive feedback, revise, and reflect, while leaving pedagogical responsibility with human participants.

Physics teacher education provides a demanding context for examining this issue [16]. Teaching physics requires more than the accurate transmission of disciplinary content [17,18]. Pre-service physics teachers must learn to transform abstract concepts such as force, field, energy, motion, electric circuits, and wave phenomena into teachable representations that can be understood by learners with diverse prior conceptions [19]. They also need to connect conceptual explanation with experimentation, modelling, visualisation, simulation, classroom demonstration, and inquiry-based task design [20]. These requirements make physics teaching highly dependent on the coordination of content knowledge, pedagogical reasoning, and technological representation [21,22,23]. For example, a simulation of electromagnetic induction is not pedagogically valuable simply because it is digital; it becomes valuable only when the teacher can align it with learners’ misconceptions, experimental evidence, representational transitions, classroom questioning, and assessment of understanding [24,25,26]. For this reason, technological pedagogical content knowledge, or TPACK, is not a generic digital competence in physics teacher education. It represents the capacity to integrate physics content, teaching strategy, and technological tools into coherent instructional decisions. At the same time, physics teaching often involves collaborative professional work, including designing laboratory tasks, evaluating model-based explanations, planning demonstrations, revising worksheets, and coordinating group inquiry activities. Collaborative problem solving, or CPS, is therefore not an additional soft skill in this context [27]. It is closely connected to how pre-service physics teachers negotiate shared understanding, take appropriate instructional action, and organise team-based design work [28,29]. In physics teacher education, TPACK and CPS should thus be regarded as highly coupled professional competences: one concerns the quality of technology-mediated instructional design, while the other concerns the collaborative processes through which such design is generated, justified, and improved [27,30].

Problem-based learning offers a suitable pedagogical architecture for cultivating these competences because it organises learning around authentic, ill-structured, and practice-oriented problems [31]. In teacher education, PBL can move pre-service teachers beyond passive reception of methods knowledge by requiring them to analyse teaching situations, identify instructional constraints, propose solutions, justify design choices, and revise their work through evidence and feedback [32,33,34,35]. This is especially relevant for physics teacher education, where instructional problems often require the integration of conceptual explanation, experimental design, digital representation, learner diagnosis, and classroom management. In the present study, PBL is not introduced simply to demonstrate its effectiveness once again; rather, it provides the instructional structure through which AI-supported scaffolds can be connected to authentic pedagogical problems and collaborative inquiry. Rather, PBL provides the instructional structure through which technology-supported resources, including AI-supported scaffolds in the present study, can be connected to authentic problems, collaborative inquiry, and pedagogical decision-making [36,37]. Without a problem-based task environment, generative AI may be used only to produce generic lesson plans or fragmented teaching materials. Within a PBL environment, however, AI can be connected to specific pedagogical purposes: helping groups compare alternative representations, formulate inquiry questions, review the coherence of learning activities, identify missing links between objectives and assessment, and generate prompts for reflection and revision. In this relationship, PBL organises authentic professional tasks, while AI provides scaffolding, feedback, and revision support. PBL gives direction to learning activity; AI expands the resources and feedback available during design work. PBL requires collaborative inquiry [38]; AI can support the quality and efficiency of that inquiry when its outputs are critically examined rather than passively accepted [36].

Despite growing interest in AI, digital pedagogy, PBL, and collaborative learning, several gaps remain in the existing literature [37,39,40,41,42,43]. First, technology integration and collaborative problem solving are often examined as separate outcomes, although authentic teacher education tasks commonly involve design, discussion, feedback, and revision within the same professional activity. Second, studies of AI in teacher education have tended to emphasise usability, acceptance, or general attitudes, while offering less intervention-based evidence on how AI-supported curriculum modules are associated with changes in specific professional competences. Third, physics teacher education remains underrepresented in research on generative AI, despite the discipline’s strong demands on abstract content representation, experimental reasoning, technological modelling, and collaborative task design. Fourth, many AI-related studies do not sufficiently clarify how AI use is governed pedagogically, including what roles AI is allowed to play, how generated outputs are reviewed, and how human judgement remains central to instructional decision-making. These gaps indicate the need for a study that treats AI not as a replacement for teachers’ reasoning, but as a regulated curriculum scaffold embedded in a problem-based teacher education module [44,45,46]. Accordingly, the present study examined whether participation in an AI-supported CTD-PBL module was associated with differential pre–post changes in pre-service physics teachers’ TPACK and CPS. The module positioned generative AI as a support for resource exploration, collaborative lesson planning, task review, instructional revision, and reflective improvement, while maintaining teacher educators’ and participants’ responsibility for evaluating and confirming all AI-generated content [47]. By focusing on both self-reported TPACK and perceived CPS processes [6,7,14,15], the study investigated whether an AI-supported problem-based teacher education module was related to stronger reported development in technology-integrated instructional knowledge and collaborative problem-solving processes in physics teacher preparation.

RQ1. To what extent was participation in the AI-supported CTD-PBL module associated with differential pre–post changes in pre-service physics teachers’ self-reported TPACK compared with conventional instruction?

RQ2. To what extent was participation in the AI-supported CTD-PBL module associated with differential pre–post changes in pre-service physics teachers’ perceived CPS processes compared with conventional instruction?

RQ3. How were the observed pre–post changes reflected across the TPACK domains and CPS dimensions targeted by the module?

2. Materials and Methods

2.1. Design and Participants

This study employed a quasi-experimental pre-test/post-test design with a non-equivalent control group to examine changes in pre-service physics teachers’ questionnaire-based TPACK and CPS outcomes and to compare the CTD-PBL and conventional groups at post-test. This design was appropriate for the teacher education context because the intervention was implemented in intact classes, where individual random assignment was not feasible. The CTD-PBL module was developed as part of a doctoral dissertation project through a systematic design, development, and validation process [48,49]. The present article focuses specifically on the intervention implementation and outcome evaluation phases of the larger project; therefore, only intervention features directly relevant to the empirical comparisons are reported. The intervention lasted 8 weeks for each group. Since the same instructor taught both the CTD-PBL and conventional groups, the intervention was scheduled across 16 calendar weeks, with the two classes taught in alternating weeks. The instructional sequence was organised around a problem-based task chain designed for pre-service physics teachers and progressively guided participants to address real-world physics teaching problems that required content comprehension, technological representation, collaborative task design, AI-supported resource use, peer feedback, and reflective revision.

The participants were 130 third-year pre-service physics teachers enrolled in an undergraduate physics teacher education programme at a public university in western China. The institution represents a regional teacher education context in which pre-service physics teachers are expected to develop technology-integrated instructional design competence and collaborative problem-solving competence in response to ongoing curriculum reform and digital education initiatives. Two intact classes participated in the study [50,51,52]. One class was assigned to the CTD-PBL group, and the other served as the conventional instruction group. The CTD-PBL group included 65 participants, comprising 44 females and 21 males. The conventional instruction group also included 65 participants, comprising 41 females and 24 males. The number of participants remained consistent across the pre-test and post-test stages, with no attrition in the reported dataset.

Table 1. Participant characteristics and baseline comparability across the two instructional groups.

Group	N	Female	Male	Pre-Test TPACK M (SD)	Pre-Test CPS M (SD)
Conventional	65	41(63.1%)	24 (36.9%)	3.16 (0.59)	2.99 (0.65)
CTD-PBL	65	44 (67.6%)	21 (32.3%)	3.01(0.64)	3.12 (0.66)

presents participant characteristics and baseline comparability by instructional group [52,53,54].

Both groups were from the same university, the same academic year, and the same undergraduate physics teacher education programme. This arrangement helped reduce potential variation related to institutional context, programme structure, year level, and prior curriculum exposure. Although the non-equivalent control group design could not fully eliminate selection effects, the use of pre-test measures provided a basis for examining baseline comparability and for interpreting subsequent changes in TPACK and collaborative problem solving. The conventional group received instruction through the regular teacher education approach used in the physics teacher education programme. This approach was mainly lecture based, demonstration oriented, and discipline-centred. Teaching was organised around the systematic explanation of physics content, basic pedagogical knowledge, and conventional classroom application. Following the regular teaching plan, the instructor explained the experimental content, demonstrated experimental procedures, and provided guidance during students’ task completion.

Students in the conventional group completed the assigned tasks in small groups. They were allowed to use ordinary non-AI classroom support resources according to their own learning needs, including online searching, short instructional videos, peer discussion, and consultation with the instructor. However, AI tools were not permitted in the conventional group. This arrangement ensured that the conventional group represented regular classroom learning supported by ordinary auxiliary resources, rather than AI-supported or CTD-PBL-based instruction.

In this study, the instructional condition was defined as participation in the AI-supported CTD-PBL module. AI support was embedded within a broader CTD-PBL structure that included problem-based tasks, structured collaboration, technology-supported inquiry, teacher scaffolding, peer feedback, and reflective refinement. The comparison group received conventional instruction covering the same physics teacher education content during the same intervention period, and both groups were taught by the same course instructor. These design arrangements helped reduce instructor-related, content-related, and time-related differences between groups. However, because the study used intact classes and participants were not randomly assigned to conditions, the between-group differences in pre–post gains should be interpreted as associations between the instructional condition and learning outcomes. Selection effects and other unmeasured baseline differences between classes cannot be ruled out.

Table 2. Instructional differences between the CTD-PBL group and the conventional group.

Aspect	CTD-PBL Group	Conventional Group
Instructional approach	Structured CTD-PBL instructional approach	Regular conventional teaching approach
Teaching orientation	Problem-based, collaborative, technology-supported, and inquiry-oriented	Lecture-based, demonstration-oriented, and discipline-centred
Classroom process	Students worked through problem situations, inquiry tasks, collaborative planning, peer feedback, and reflective revision	The instructor explained physics content, demonstrated experimental procedures, and guided students during task completion
Student task organisation	Small-group work with structured collaboration, role allocation, shared responsibility, and collective outputs	Small-group task completion with ordinary peer discussion
Learning resources	Physics learning resources, simulation or digital tools, online materials, collaborative platforms, and AI prompt support provided as part of the CTD-PBL tasks	Ordinary non-AI resources, such as online searching, short instructional videos, peer discussion, and teacher consultation
AI support	AI support was embedded within the CTD-PBL learning process, including the provision of AI prompts to support inquiry, planning, and refinement	AI tools were not permitted
Teacher role	The instructor provided staged scaffolding, task guidance, feedback, and support for collaborative inquiry	The instructor provided explanation, demonstration, procedural guidance, and help when students requested support
Main instructional distinction	Learning was organised through the CTD-PBL structure, where AI was one embedded support component	Learning followed the regular teaching plan, with ordinary non-AI support used according to students’ needs

summarises the key instructional differences between the CTD-PBL group and the conventional group.

2.2. Intervention

The CTD-PBL module was developed as part of the doctoral dissertation project through a systematic design, development, and validation process [48,49]. The present article focuses specifically on the intervention implementation and effectiveness evaluation phase of that larger project. For the purposes of this study, only the intervention features directly relevant to the empirical evaluation are reported. The intervention lasted eight weeks and was organised around a problem-based task chain specifically designed for pre-service physics teachers. The instructional sequence was structured to progressively engage participants in authentic physics teaching problems requiring the integration of content understanding, technological representation, collaborative task design, AI-supported resource use, peer feedback, and reflective revision.

Its instructional architecture was grounded in the pedagogical logic of problem-based learning and technology-supported collaborative inquiry. The CTD-PBL module was operationalised through six mutually reinforcing instructional techniques: active learning, collaborative learning, inquiry-based learning, technology integration, formative assessment, and scaffolded support. These techniques collectively structured student-centered problem solving, technology-enhanced inquiry, collaborative knowledge construction, iterative feedback use, and reflective refinement throughout the intervention process. Each weekly session required participants to work collaboratively on discipline-specific instructional design tasks situated within authentic physics teaching scenarios. Participants were required to analyse instructional problems, propose technology-supported pedagogical solutions, evaluate alternative approaches, revise instructional products based on peer and instructor feedback, and document reflective justification for design decisions. Generative AI tools were integrated as bounded instructional scaffolds rather than autonomous solution providers. Their use was limited to supporting idea generation, resource organisation, explanation refinement, feedback checking, and revision planning. All AI-supported outputs were critically evaluated and revised by participants to ensure pedagogical appropriateness, conceptual accuracy, and alignment with intended instructional objectives. Implementation fidelity was maintained through structured instructional procedures, including standardised task sequencing, collaborative role allocation, instructor monitoring, guided reflection protocols, and systematic documentation of participant outputs across all intervention stages.

Table 3. Instructional techniques incorporated in the CTD-PBL module.

Teaching Technique	Description
Active Learning	Emphasises student-centered learning through group discussions, case studies, and problem-solving tasks to promote engagement and deeper understanding.
Collaborative Learning	Students work in small groups, with roles assigned to ensure contribution and foster communication, teamwork, and leadership skills.
Inquiry-Based Learning	Encourages students to approach problems with inquiry, ask questions, investigate solutions, and critically evaluate different approaches.
Technological Integration	Students use digital tools like simulation software, online databases, and collaborative platforms to enhance learning and visualise physics phenomena.
Formative Assessment	Continuous assessment through quizzes, peer evaluations, reflective journals, and instructor observations to provide ongoing feedback.
Scaffolded Support	Provides gradual support through guided instructions, learning resources, and timely feedback to help students build confidence in tackling challenges.

summarises the instructional techniques incorporated into the CTD-PBL module.

The module used problem-based learning as the main pedagogical structure. Participants worked in small groups to analyse authentic physics teaching problems, design feasible instructional solutions, generate evidence through technological tools or classroom-oriented representations, and transform their solutions into teachable products. The instructor guided the process by clarifying task requirements, monitoring group progress, supporting the evaluation of physics content accuracy, and helping participants connect technical solutions with pedagogical purposes.

The intervention did not position AI as a separate add-on activity. Instead, AI-supported resources were embedded into the design, review, and revision stages of each task. Participants used DeepSeek-V4 to support lesson idea generation, resource organisation, explanation comparison, feedback prompting, and revision planning. However, all AI-generated outputs had to be examined by participants and reviewed in relation to physics content accuracy, classroom feasibility, pedagogical appropriateness, and ethical use. The six CTD-PBL task modules were systematically aligned with core physics topics, required student outputs, evidence sources, and dominant TPACK/CPS targets, thereby translating the intervention design into observable learning tasks that supported technological–pedagogical reasoning and collaborative problem solving development.

Table 4. The six CTD-PBL task modules, required outputs, and targeted TPACK/CPS development.

Module	Core Physics Focus	Required Output	Evidence Type	Dominant TPACK/CPS Targets
Safe DC Power Supply Box	Low-voltage power design	Group design plan + instructional explanation	Planning sheet, prototype notes	TCK, TPCK, task regulation
RC Transient and Sensing	RC circuit and sensor logging	Data-logging task design	Data sheet, design rationale	TK, TPK, learning, and knowledge building
Magnetic Field and Force Mapping	Field representation	Mapping activity + learner explanation	Measurement record, concept explanation	TCK, TPCK, perspective taking
Electromagnetic Induction and Energy Conversion	Generator optimisation	Demonstration plan + refinement	Trial records, peer feedback	PCK, TPCK, social regulation
AC to DC Transformation for Sensors	Transformation for classroom sensors	Instructional product	Product rubric, reflection	TCK, task regulation
EM Waves, Shielding, and Interference Detection	EMC survey	Mini inquiry/report	Final product, peer review	CPS participation, shared knowledge construction

presents the six CTD-PBL task modules, their required outputs, and the targeted areas of TPACK and CPS development.

The six task modules addressed physics education problems involving low-voltage DC power-supply design, RC-based low-cost sensor data logging, magnetic-field mapping using a smartphone magnetometer or Hall sensor, optimisation of a hand-cranked generator, AC-to-DC conversion for classroom sensors, and a mini electromagnetic-compatibility survey. These tasks were selected because they required participants to connect physics concepts, technological tools, classroom demonstrations, experimental or modelling evidence, and collaborative instructional design. The figure below synthesises the intervention timeline, the recurrent problem-based learning cycle, and the placement of pre-test and post-test measurements. The six modules were implemented through iterative phases of problem encounter, analysis, planning, investigation, solution construction, presentation, and reflection. The CTD-PBL module incorporated staged scaffolds and human verification mechanisms across problem launch, problem analysis, investigation, solution construction, and presentation and reflection, ensuring that students’ problem-solving processes were guided, checked, and documented through traceable revision evidence.

Table 5. Instructional scaffolds, verification boundaries, and revision evidence in the CTD-PBL module.

Stage	Scaffold/Support	Human Verification Mechanism	Revision Evidence
Problem launch	Facilitation prompts; role setup	Instructor clarification	Initial plan sheet
Problem analysis	Planning templates; misconception prompts	Peer/instructor review	Brainstorm notes; misconception list
Investigation	Tool-use guidance; evidence templates	Content/feasibility check	Datasheets; setup photos
Solution construction	Prototype/testing scaffold	Rubric-based check	Prototype; trial records
Presentation and reflection	Peer feedback form; reflection questions	Peer review + self-check	Final product; reflection

summarizses the instructional scaffolds, verification boundaries, and revision evidence embedded in the CTD-PBL module.

The intervention exposure for each group lasted eight instructional weeks. However, because the same course instructor taught both the CTD-PBL and conventional instruction groups, the implementation was arranged across a 16-week calendar period. To reduce instructor-related variation, the two intact classes were taught in an alternating sequence rather than simultaneously. Thus, the 16 weeks represent the calendar schedule of implementation, whereas each group received the same 8-week instructional sequence, including pre-test, orientation to tools and methods, six instructional task cycles, and post-test. Figure 1 provides an overview of the eight-week timeline and pedagogical flow of the AI-supported CTD-PBL intervention.

2.3. AI Support Operationalisation

DeepSeek was selected as the generative AI platform for the CTD-PBL module for contextual and pedagogical reasons. First, it was accessible to participants through a publicly available web interface during the intervention period and therefore did not require institution-specific software installation or a paid experimental system. Second, it supported Chinese-language interaction, which was important because the participants were Chinese pre-service physics teachers and the instructional design tasks involved discipline-specific pedagogical reasoning in the local teacher education context. Third, its functions were suitable for the bounded scaffolding purposes of the module, including lesson idea generation, resource organisation, explanation comparison, feedback prompting, and revision planning. In this study, DeepSeek was operationalised as an instructional scaffold rather than as an autonomous instructional agent. Its function was not to provide final answers, replace teacher judgement, or determine the quality of participants’ instructional products. Instead, it supported specific learning actions within the problem-based design process. To improve procedural consistency, participants in the CTD-PBL group accessed DeepSeek through the same public web interface, used it during the designated CTD-PBL tasks, and followed the same task-specific prompt guidelines and human verification requirements. The study therefore treated AI integration as a structured pedagogical condition rather than as unrestricted tool use.

Since generative AI platforms are frequently updated, the study does not claim to evaluate a fixed DeepSeek model version or to compare the performance of different model versions. During the implementation period, the platform may have changed in interface design, response style, accessibility, or underlying model capability. This creates a limitation for exact technical reproducibility. However, the purpose of the study was not to test model-specific performance, but to examine whether participation in an AI-supported CTD-PBL instructional condition was associated with differential changes in TPACK and CPS. For this reason, DeepSeek is reported as the generative AI platform used for pedagogical scaffolding, while the replicable element of the intervention lies primarily in the instructional design: bounded AI use, task-specific prompts, collaborative review, teacher monitoring, and human verification of AI-generated outputs.

AI support was used in five main ways. First, participants used DeepSeek for lesson idea generation. For example, they could request alternative ways to introduce a physics concept, organise a classroom demonstration, or connect a technical task with a secondary school physics topic. Second, AI was used for task refinement. Participants could ask DeepSeek to identify missing steps in an activity sequence, check whether task instructions were clear, or suggest ways to make a group task more feasible for classroom implementation. Third, AI was used to generate feedback prompts. Instead of accepting AI feedback as authoritative, participants used prompts to guide peer review, such as checking alignment among learning objectives, activities, representations, and assessment tasks. Fourth, AI was used for resource organisation. Participants could ask DeepSeek to summarise relevant teaching resources, compare possible digital tools, or organise background information for a design task. Fifth, AI was used for checking pedagogical appropriateness. This included examining whether a proposed explanation was suitable for learners’ prior knowledge, whether a representation might create misconceptions, or whether a task sequence supported inquiry rather than mechanical completion. The AI-supported CTD-PBL design was grounded in TPACK, constructivism, cognitive load theory, collaborative learning, and situated learning, with each theoretical perspective translated into specific design principles, module-level tasks, verification evidence, and outcome dimensions to ensure theoretical coherence and traceable human-machine verification across the intervention.

To further standardise AI use across the six CTD-PBL modules, participants were provided with task-specific prompt templates rather than being allowed to use DeepSeek in an unrestricted manner. These templates guided participants to use AI for five bounded purposes: generating initial lesson ideas, organising relevant resources, comparing alternative explanations or representations, producing peer-feedback prompts, and planning revisions. Before the intervention, participants received orientation on how to access DeepSeek through the same public web interface, how to formulate prompts, how to record AI-supported suggestions in group planning sheets, and how to verify generated content before it could be incorporated into instructional artefacts. AI use was allowed only during designated CTD-PBL task phases and was linked to the required outputs of each module, such as design plans, calibration sheets, visual representations, demonstration scripts, peer-feedback forms, and reflection notes.

All AI-generated suggestions were treated as provisional and required human verification. Participants were instructed to check AI outputs against four criteria: physics content accuracy, pedagogical appropriateness, classroom feasibility, and ethical or safety acceptability. When an AI-generated response contained uncertain, unsupported, inaccurate, or overly generic information, participants were required to reject, revise, or verify it using course materials, measurement evidence, group discussion, peer review, or instructor feedback. DeepSeek outputs were not used directly as research data, scoring evidence, or statistical input. The study recorded AI-supported work through planning sheets, revision notes, peer-feedback records, reflection documents, and final artefacts; however, complete raw AI interaction logs were not systematically collected. A module-level prompt framework, including sample prompts, expected AI-supported outputs, and verification procedures, is provided in Appendix B.

Table 6. Theoretical foundations, design principles, and verification evidence of the AI-supported CTD-PBL module.

Theory	Design Principle	CTD-PBL Design Choice	Module Anchor Task	Evidence	Outcome Dimension
TPACK	Tool–content–pedagogy alignment; explicit rationale	Tool bound to content rep. + teaching script	M5 ripple vs. load with rectifier + filter; M3 field heatmap + micro-lesson	Objectives–instrument mapping; DQ/MQ/EQ rubric; micro-lesson script	TPACK.xs subscales (TK, TPK, TCK, PCK, TPACK)
Constructivism	ERC scaffolds; public artefacts; iteration	Launch → investigate → ERC memo → demo	M2 calibration → model selection ERC; M4 optimisation memo	ERC memos; peer-review forms; iteration logs	Higher-order Bloom evidence (A/E/C artefacts)
CLT	Reduce extraneous; manage intrinsic; increase germane	Pre-training; segmented worksheets; templates; safety/fidelity	M1 stability test with pre-trained meters; M5 annotated waveforms	Fidelity checklist; error/redo logs; time-on-task	MQ gains (repeatability, uncertainty)
Collaborative learning	Interdependence + accountability + social skills	Role rotation; peer-review; CPS prompts	All CTD weeks: team roles and review cycles	CPS questionnaire; peer-review means; collab notes	CPS facets (coordination, monitoring, decision)
Situated learning	Authentic tasks; teacher-ready packs; transfer	Deliver teacher packs; demo plans; constraints-aware design	M6 EMC survey for real classrooms; M1 safety demo	Teacher-pack checklist; classroom-fit rating	Transfer; enacted TPACK

summarises the theoretical foundations, design principles, and verification evidence underpinning the AI-supported CTD-PBL module.

The operational boundary of AI use was made explicit to participants. AI outputs could be used only as provisional materials for discussion, comparison, and revision. They were not accepted as evidence unless participants could verify them through physics principles, course materials, experimental reasoning, or instructor feedback. DeepSeek was not used for data generation, statistical analysis, scoring, or research conclusion decision-making. This boundary was important for both ethical and methodological reasons. It ensured that AI functioned as a scaffold for professional learning while preserving human responsibility for pedagogical judgement.

2.4. Instruments

TPACK competencies were measured using a validated 28-item TPACK self-assessment questionnaire [55]. The instrument assessed seven TPACK domains: Pedagogical Knowledge (PK), Content Knowledge (CK), Technological Knowledge (TK), Pedagogical Content Knowledge (PCK), Technological Pedagogical Knowledge (TPK), Technological Content Knowledge (TCK), and Technological Pedagogical Content Knowledge (TPCK). Items were rated on a five-point Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree). Mean scores were calculated for the overall TPACK scale and for each domain. Higher mean scores indicated higher self-reported levels of the corresponding TPACK competence. Participants completed the questionnaire before and after the intervention.

CPS was measured using a questionnaire-based operationalisation aligned with the OECD PISA framework [29,56]. The CPS measure covered five dimensions: Participation, Perspective Taking, Social Regulation, Task Regulation, and Learning and Knowledge Building. These dimensions corresponded to the collaborative and problem-solving processes targeted by the CTD-PBL module, including active involvement in group work, understanding others’ perspectives, regulating team interaction, organising task progress, and constructing shared knowledge. Items were rated on a five-point Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree). Mean scores were calculated for the overall CPS scale and for each dimension. Higher mean scores indicated higher self-reported levels of the corresponding collaborative problem-solving behaviour.

The TPACK and CPS instruments were adopted from previously validated measures. The present study did not modify the item content, dimensional structure, response format, or scoring procedure of either instrument. Because the participants were Chinese pre-service teachers, the instruments were translated into Chinese for administration. The translation process focused on semantic equivalence and educational terminology consistency rather than construct adaptation. The translated items were reviewed by the research team to ensure that key terms related to technological pedagogical content knowledge, collaborative problem solving, and physics teacher education were clear and contextually appropriate for the target participants. No substantive changes were made to the constructs or measurement dimensions. The same Chinese version was used in both the pre-test and post-test, and all responses were scored according to the original scoring procedures.

Reliability evidence for the two instruments was considered before the outcome analyses. For the TPACK questionnaire, previous validation work supported its construct validity, with standardised validity coefficients ranging from 0.71 to 0.93. In the present study context, the translated Chinese version was pilot-tested with pre-service teachers before formal implementation, and the reliability analysis indicated satisfactory internal consistency, with Cronbach’s alpha values ranging from 0.82 to 0.91 across the TPACK subscales. The overall internal consistency of the TPACK instrument was also satisfactory, with Cronbach’s alpha reported at 0.91. For the CPS questionnaire, prior reliability evidence indicated high internal consistency for the overall scale and acceptable to strong reliability across its subscales. The overall CPS scale showed Cronbach’s alpha values above 0.90 for both the pre-test and post-test administrations, and the reliability coefficients for the CPS factors met or exceeded 0.75. One negatively worded CPS item was reverse-scored before analysis so that higher scores consistently indicated stronger CPS-oriented behaviour. Taken together, these reliability results supported the use of the TPACK and CPS mean scores in the subsequent pre–post and between-group analyses.

Supplementary instruments and validation tools were used during module development and refinement. These included a needs analysis survey, an expert consultation protocol, a teaching challenge questionnaire, content validity review forms, and language validity checklists. These tools were not used as primary outcome measures in the present quasi-experimental analysis. Instead, they supported the development of module content, instructional sequencing, assessment alignment, language clarity, and implementation feasibility.

2.5. Validity, Reliability, and Module Refinement

Several procedures were used to strengthen the validity and implementation readiness of the CTD-PBL module and its associated instruments. During the development stage, the needs analysis survey identified priority learning areas and perceived challenges among pre-service physics teachers. Expert consultation with five physics teacher educators provided professional judgement on instructional importance, learner difficulties, and module relevance. Pilot feedback and preliminary reliability analysis were used to refine the clarity, contextual fit, and implementation feasibility of the module [57].

The module validation process included content validity and language validity checks. All 33 content-validity items received a minimum score of four on a five-point scale, and the language-validity review reached near-complete agreement. Based on expert and pilot feedback, task briefs, timings, prompt wording, rubric descriptors, and resource options were revised before the formal implementation. Formative evidence from student questions, trial logs, expert review, language checks, and peer-feedback records was used to identify implementation problems and make targeted instructional revisions, thereby strengthening the clarity, alignment, usability, and fidelity of the CTD-PBL intervention before full implementation.

Table 7. Instructional revisions based on formative evidence.

Problem Identified	Evidence	Revision Made
Ambiguity in data sheet instructions	Student confusion during setup; repeated questions observed	Reworded instruction; added example entries; updated worksheet version to 1
Misalignment between product rubric and expected student outputs	Expert noted scoring mismatch; inconsistent marking in trial run	Revised rubric descriptors; aligned with typical student products; added scoring notes
Confusion in tool selection process during sensor experiment	Trial logs showed delays; students used incorrect instruments	Added visual tool guide; relocated tool list to pre-lab checklist
Terminology inconsistency across modules	Language checklist flagged variation; 50% of experts suggested correction	Standardised key terms; applied across all modules; changes tracked in version 1.3
Low engagement in peer review activity	Student feedback showed unclear purpose; peer form underused	Simplified peer form layout; clarified role in task sheet; added oral prompt

presents the instructional revisions made to the CTD-PBL module based on formative evidence collected during implementation.

Implementation fidelity was monitored through a structured observation and documentation procedure across all CTD-PBL modules. The fidelity checks addressed five aspects of implementation: adherence to the planned CTD-PBL sequence, collaborative participation and role fulfilment, pedagogically aligned technology use, instructor scaffolding and feedback, and completion of required process documentation. For each module, the instructor verified whether the intended learning phases were enacted, including task induction, problem clarification, planning, implementation with data collection, presentation with peer review, and reflection. Process evidence included role allocation records, planning sheets, data logs, peer-feedback forms, reflection notes, and final instructional artefacts. Field notes were used to record implementation issues, deviations from the planned procedure, and corrective support provided during group work. All records were linked to coded identifiers, stored securely, and analysed in aggregate form to protect participant confidentiality. These fidelity records provided procedural evidence that the CTD-PBL condition was implemented with sufficient consistency to support interpretation of the outcome analyses.

2.6. Data Collection Procedure

Data collection was conducted in three main stages. First, before the intervention, participants in both groups completed the TPACK and CPS pre-test measures. These data were used to describe baseline levels and assess initial group comparability. Second, the CTD-PBL group received the CTD-PBL module, whereas the conventional group received conventional instruction on the same physics content. During the intervention, the CTD-PBL group completed collaborative instructional design tasks, technology supported activities, peer feedback, and reflective outputs. Third, after the intervention, both groups completed the TPACK and CPS post-test measures. The pre-test and post-test structure enabled examination of change over time within each group and comparison of developmental patterns between the CTD-PBL and conventional groups. Although each group received an eight-week instructional exposure, the implementation was spread across 16 calendar weeks because the same instructor taught the two intact classes in alternating weeks. This arrangement was used to maintain instructor consistency across conditions.

2.7. Data Analysis

All quantitative analyses were conducted using SPSS version 27.0. Before inferential analysis, the dataset was screened for completeness, matched pre-test and post-test responses, missing values, distributional normality, and suitability for parametric analysis. The final dataset included 130 matched cases, with 65 participants in the CTD-PBL condition and 65 participants in the conventional instruction condition.

The primary analysis was a 2 × 2 mixed repeated-measures ANOVA conducted separately for TPACK and CPS. In each model, group was specified as the between-subjects factor with two levels: CTD-PBL and conventional instruction. Test time was specified as the within-subjects factor with two levels: pre-test and post-test. The participant was treated as the repeated-measures units. The model examined the main effect of group, the main effect of time, and the group × time interaction. The group × time interaction was specified a priori as the primary inferential indicator because it tested whether the two instructional conditions showed different patterns of pre–post change. In line with the intact-class quasi-experimental design, this interaction was interpreted as evidence of differential change associated with instructional condition rather than as definitive causal evidence of intervention effectiveness.

Secondary repeated-measures analyses were conducted for the seven TPACK domains and the five CPS dimensions to explore whether the patterns were reflected at the dimensional level. These dimensional analyses were used to support interpretation of the outcomes and were not treated as stronger evidence than the primary TPACK and CPS models.

Supplementary analyses were conducted to clarify baseline comparability and simple pre–post patterns. Independent-samples t-tests were used to examine between-group differences at pre-test and post-test. Paired-samples t-tests were used to describe within-group changes from pre-test to post-test. These t-tests were treated as supplementary analyses rather than as the primary basis for evaluating differential change. Descriptive statistics, including means and standard deviations, were reported by group and test time. For the repeated-measures models, F values, numerator and denominator degrees of freedom, p values, partial eta squared, estimated marginal means or mean changes, and 95% confidence intervals were reported where available. The significance level was set at α = 0.05. Because the within-subjects factor had only two levels, the sphericity assumption was not applicable.

2.8. Ethical Considerations

This study involved human participants and was conducted in accordance with institutional ethical requirements. Ethical approval was obtained from the relevant ethics review board before data collection. The approval procedures were aligned with university regulations and local educational requirements. Participant data were anonymised using coded identifiers, stored securely, and accessed only by the research team. All data were analysed and reported in aggregate form to protect participants’ confidentiality. All participants were informed of the purpose of the study, the research procedures, the voluntary nature of participation, confidentiality arrangements, and their right to withdraw from the study at any time without penalty. Written informed consent was obtained from all participants before data collection.

AI-supported resources were used as part of the CTD-PBL instructional environment to support instructional planning, feedback use, resource organisation, and explanation refinement. These AI-supported elements were embedded within the broader CTD-PBL module and were not isolated as an independent experimental variable. Therefore, the present study evaluates the association between participation in the integrated CTD-PBL instructional condition and changes in TPACK and CPS, rather than the independent effect of AI tools alone. The results should be interpreted as evidence that the AI-supported CTD-PBL condition was associated with greater gains, not as evidence that the addition of AI by itself caused those gains.

3. Results

3.1. Data Screening and Baseline Comparability

A total of 130 valid matched responses were retained for the final analysis, with 65 participants in the CTD-PBL group and 65 participants in the conventional group. The dataset was screened before inferential analysis. Normality checks for TPACK, the seven TPACK domains, CPS, and the five CPS dimensions indicated no substantial deviation from normality. Therefore, the data were considered suitable for the primary mixed repeated-measures ANOVA and the supplementary independent-samples and paired-samples t-tests.

Pre-test comparisons indicated that the two groups were broadly comparable before the intervention. For TPACK, no statistically significant pre-test difference was found between the CTD-PBL group and the conventional group on the overall score, t(128) = −1.102, p = 0.273. No significant pre-test differences were found for PK, CK, TK, PCK, TPK, TCK, or TPCK, with all p values greater than 0.05. For CPS, no statistically significant pre-test difference was found for the overall score, t(128) = −0.928, p = 0.320. Similarly, no significant pre-test differences were observed for Participation, Perspective Taking, Social Regulation, Task Regulation, or Learning and Knowledge Building. These results indicate that the two groups were sufficiently comparable at baseline for subsequent pre-test and post-test comparisons. Implementation fidelity records indicated that the CTD-PBL module was delivered according to the planned instructional sequence. The six task modules were implemented through the required phases of problem launch, problem analysis, investigation, solution construction, presentation, and reflection. Instructor monitoring notes and process documents showed that group role allocation, peer-feedback activity, reflection records, and artefact production were used across the module activities. The required evidence archive included planning sheets, data or trial records, peer-review forms, reflection notes, and final instructional products. No major implementation deviation was recorded that required exclusion of any module session or participant group from the outcome analysis. These records support the consistency of the CTD-PBL implementation while recognising that fidelity was monitored descriptively rather than treated as a separate quantitative outcome. TPACK scores represent self-reported TPACK measured by a self-assessment questionnaire; CPS scores represent questionnaire-based perceived CPS processes and should not be interpreted as objective performance scores.

3.2. Primary Mixed Repeated-Measures Model Results for TPACK and CPS

The primary analysis examined whether the CTD-PBL and conventional instruction conditions showed different patterns of pre–post change in TPACK and CPS.

Table 8. Primary mixed repeated-measures ANOVA results for TPACK and CPS.

Outcome	Group	N	Pre-Test M (SD)	Post-Test M (SD)	Gain [95% CI]	Difference in Gain [95% CI]	Group × Time F (df)	p	Partial η2
TPACK	Conventional	65	3.16 (0.59)	3.40 (0.61)	0.24 [0.08, 0.40]	0.79 [0.55, 1.03]	F(1, 128) = 41.787	<0.001	0.139
TPACK	CTD-PBL	65	3.01 (0.64)	4.04 (0.75)	1.03 [0.82, 1.24]
CPS	Conventional	65	2.99 (0.65)	3.05 (0.66)	0.06 [0.00, 0.12]	0.44 [0.31, 0.57]	F(1,128) = 46.882	<0.001	0.152
CPS	CTD-PBL	65	3.12 (0.66)	3.62 (0.65)	0.50 [0.34, 0.66]

presents the primary mixed repeated-measures ANOVA results and the corresponding pre–post mean changes for both outcomes.

The primary mixed repeated-measures ANOVA showed a significant group × time interaction for TPACK, F(1,128) = 41.787, p < 0.001, partial η² = 0.139. The conventional group increased from M = 3.16 at pre-test to M = 3.40 at post-test, whereas the CTD-PBL group increased from M = 3.01 to M = 4.04. The estimated mean change was 0.24 [95% CI: 0.08, 0.40] in the conventional group and 1.03 [95% CI: 0.82, 1.24] in the CTD-PBL group, yielding a difference in change of 0.79 [95% CI: 0.55, 1.03]. This indicates that the CTD-PBL condition was associated with stronger pre–post gains in TPACK.

For CPS, the primary mixed repeated-measures ANOVA also showed a significant group × time interaction, F(1,128) = 46.882, p < 0.001, partial η² = 0.152. The conventional group increased from M = 2.99 at pre-test to M = 3.05 at post-test, whereas the CTD-PBL group increased from M = 3.12 to M = 3.62. The estimated mean change was 0.06 [95% CI: 0.00, 0.12] in the conventional group and 0.50 [95% CI: 0.34, 0.66] in the CTD-PBL group, yielding a difference in change of 0.44 [95% CI: 0.31, 0.57]. This pattern indicates stronger CPS gains in the CTD-PBL condition than in the conventional instruction condition.

3.3. Group Differences and Pre–Post Changes in TPACK Competencies

After the primary TPACK model, supplementary t-test analyses and secondary dimensional repeated-measures analyses were conducted to clarify the pattern of TPACK results. The supplementary independent-samples t-tests showed no significant baseline differences in TPACK or its seven dimensions before the intervention period. At post-test, the CTD-PBL group scored significantly higher than the conventional instruction group on TPACK and seven TPACK domains. Specifically, significant post-test differences were found for PK, t(128) = −4.612, p < 0.001; CK, t(128) = −5.233, p < 0.001; TK, t(128) = −4.981, p < 0.001; PCK, t(128) = −6.021, p < 0.001; TPK, t(128) = −5.442, p < 0.001; TCK, t(128) = −5.118, p < 0.001; and TPCK, t(128) = −6.884, p < 0.001. These post-test comparisons should be interpreted as supplementary evidence of between-group differences, not as the primary test of differential change.

Table 9. Supplementary post-test independent-samples t-test results for TPACK and its dimensions.

Variable	Conventional M (SD)	CTD-PBL M (SD)	Mean Difference [95% CI]	t (df)	p	Cohen’s d
TPACK	3.40 (0.61)	4.04 (0.75)	0.64 [0.42, 0.86]	−5.842 (128)	<0.001	1.02
PK	3.45 (0.60)	4.00 (0.81)	0.55 [0.31, 0.79]	−4.612 (128)	<0.001	0.81
CK	3.47 (0.69)	4.05 (0.76)	0.58 [0.36, 0.80]	−5.233 (128)	<0.001	0.92
TK	3.35 (0.65)	4.13 (0.70)	0.78 [0.47, 1.09]	−4.981 (128)	<0.001	0.87
PCK	3.42 (0.61)	4.44 (0.58)	1.02 [0.68, 1.36]	−6.021 (128)	<0.001	1.06
TPK	3.40 (0.59)	4.49 (0.54)	1.09 [0.69, 1.49]	−5.442 (128)	<0.001	0.95
TCK	3.20 (0.65)	4.47 (0.54)	1.27 [0.78, 1.76]	−5.118 (128)	<0.001	0.90
TPCK	3.37 (0.54)	4.48 (0.50)	1.11 [0.79, 1.43]	−6.884 (128)	<0.001	1.21

reports the supplementary post-test independent-samples t-test results for overall TPACK and its dimensions.

Supplementary paired-samples t-tests showed that both groups improved from pre-test to post-test, but the CTD-PBL group showed larger within-group changes. In the conventional group, TPACK increased significantly, t(64) = −2.98, p = 0.004. In the CTD-PBL group, the increase in TPACK was stronger, t(64) = −9.84, p < 0.001. Significant within-group gains were also observed across seven TPACK domains in the CTD-PBL group, including PK, CK, TK, PCK, TPK, TCK, and TPCK.

Secondary dimensional repeated-measures analyses showed significant group × time interaction effects for PK, CK, TK, PCK, TCK, and TPCK. As shown in Figure 2, the CTD-PBL group showed larger mean gains than the conventional instruction group across most TPACK dimensions. The TPK result should be reported cautiously: it improved over time and differed significantly at post-test, but its group × time interaction did not support a clearly differentiated developmental trajectory between the two instructional conditions. Therefore, the strongest dimensional evidence for differential TPACK change should be interpreted through domains with significant interaction effects, especially TCK, PCK, and TPCK. Figure 2 illustrates the mean changes from pre-test to post-test across the seven TPACK dimensions in the conventional and CTD-PBL groups.

3.4. Group Differences and Pre–Post Changes in Collaborative Problem Solving

After the primary CPS model, supplementary t-test analyses and secondary dimensional repeated-measures analyses were conducted to clarify the CPS result pattern. The supplementary independent-samples t-tests showed no significant baseline differences in CPS or its five dimensions before the intervention period. At post-test, the CTD-PBL group scored significantly higher than the conventional instruction group on CPS, t(128) = −5.024, p < 0.001. Significant post-test differences were also found for five CPS dimensions: Participation, t(128) = −4.115, p < 0.001; Perspective Taking, t(128) = −4.332, p < 0.001; Social Regulation, t(128) = −4.554, p < 0.001; Task Regulation, t(128) = −4.772, p < 0.001; and Learning and Knowledge Building, t(128) = −4.889, p < 0.001. These post-test comparisons provide supplementary evidence that the CTD-PBL group reported higher perceived CPS process scores after the intervention period.

Table 10. Supplementary post-test independent-samples t-test results for CPS and its dimensions.

Variable	Conventional M (SD)	CTD-PBL M (SD)	Mean Difference [95% CI]	t (df)	p	Cohen’s d
CPS	3.05 (0.66)	3.62 (0.65)	0.57 [0.35, 0.79]	−5.024 (128)	<0.001	0.88
Participation	3.58 (0.62)	4.10 (0.80)	0.52 [0.27, 0.77]	−4.115 (128)	<0.001	0.72
Perspective Taking	3.05 (1.07)	3.62 (1.41)	0.57 [0.31, 0.83]	−4.332 (128)	<0.001	0.76
Social Regulation	3.69 (0.61)	4.66 (0.49)	0.97 [0.55, 1.39]	−4.554 (128)	<0.001	0.80
Task Regulation	3.57 (0.57)	4.49 (0.59)	0.92 [0.54, 1.30]	−4.772 (128)	<0.001	0.84
Learning and Knowledge Building	3.63 (0.62)	4.73 (0.43)	1.10 [0.65, 1.55]	−4.889 (128)	<0.001	0.86

reports the supplementary post-test independent-samples t-test results for CPS and its dimensions.

Supplementary paired-samples t-tests showed statistically significant improvement in both groups. In the conventional instruction group, CPS increased significantly from pre-test to post-test, t(64) = −2.08, p = 0.041. In the CTD-PBL group, the increase in CPS was stronger, t(64) = −6.34, p < 0.001. Significant gains were also found across five CPS dimensions in the CTD-PBL group, including Participation, Perspective Taking, Social Regulation, Task Regulation, and Learning and Knowledge Building.

At the dimensional level, secondary repeated-measures analyses showed significant group × time interaction effects for all five CPS dimensions. The interaction effect was significant for Participation, F(1,128) = 24.736, p < 0.001; Perspective Taking, F(1,128) = 4.818, p = 0.029; Social Regulation, F(1,128) = 118.447, p < 0.001; Task Regulation, F(1,128) = 55.706, p < 0.001; and Learning and Knowledge Building, F(1,128) = 157.494, p < 0.001. The CTD-PBL group showed larger mean gains than the conventional instruction group across all five dimensions. The clearest dimensional differences were observed in Learning and Knowledge Building and Social Regulation, suggesting that the CTD-PBL condition was particularly associated with stronger gains in collaborative knowledge construction and regulation of group processes. Figure 3 illustrates the mean changes from pre-test to post-test across the five CPS dimensions in the conventional and CTD-PBL groups.

3.5. Summary of Empirical Findings

The primary mixed repeated-measures ANOVA results showed significant group × time interactions for both TPACK and CPS. These results indicate that the CTD-PBL group showed stronger pre–post gains than the conventional instruction group across the two main outcomes. Supplementary post-test comparisons further showed that the CTD-PBL group achieved significantly higher post-test performance than the conventional instruction group in both self-reported TPACK and perceived CPS processes. Because the study used an intact-class quasi-experimental design, these findings should be interpreted as differential changes associated with the instructional condition, rather than as definitive evidence of individual-level causal effects.

For TPACK, the evidence indicates stronger pre–post gains in the CTD-PBL group, with especially consistent support for domains involving technology and content integration, such as TCK, and more integrated professional knowledge, such as TPCK. For CPS, the evidence was more consistent across dimensions, with significant post-test differences and significant interaction effects for all five CPS dimensions. The clearest CPS differences were found in Learning and Knowledge Building, Social Regulation, and Participation.

4. Discussion

The present study examined whether participation in an AI-supported Collaborative TPACK Competency Development module based on problem-based learning (CTD-PBL) was associated with differential pre–post changes in pre-service physics teachers’ self-reported TPACK and perceived CPS processes. The discussion is organised around five points: the main findings, the possible instructional mechanisms, the relationship between the findings and previous literature, the limitations of the study, and the practical implications for physics teacher education.

4.1. Main Findings

The main finding was that participants in the CTD-PBL condition showed stronger pre–post gains and higher post-test questionnaire scores than participants in the conventional instruction condition. The primary mixed repeated-measures analyses indicated significant group × time interactions for both TPACK and CPS, suggesting that the two instructional conditions were associated with different developmental patterns across the intervention period. These results should be interpreted as questionnaire-based evidence that the CTD-PBL condition was associated with stronger self-reported TPACK and stronger perceived CPS processes, rather than as direct evidence of objectively demonstrated teaching competence or independently observed collaborative performance.

The dimensional analyses further clarified the pattern of change. For TPACK, the clearest differential gains were found in domains that required coordination among technology, pedagogy, and physics content, especially TCK, PCK, and TPCK. This suggests that the CTD-PBL condition was particularly associated with participants’ reported growth in integrated instructional knowledge rather than only with general familiarity with digital tools. For CPS, the CTD-PBL group showed stronger gains across the five CPS dimensions, with especially clear differences in Learning and Knowledge Building, Social Regulation, and Participation. This pattern is consistent with the collaborative design of the module, in which participants were required to work through shared tasks, exchange feedback, regulate group work, and revise collective artefacts. Taken together, the findings indicate that the CTD-PBL condition was associated with more favourable self-reported and perceived developmental outcomes in both technology-integrated instructional knowledge and collaborative problem-solving processes.

4.2. Mechanism Interpretation

The most plausible explanation for these findings lies in the integrated instructional architecture of the CTD-PBL module. The module did not treat physics content, technology use, collaboration, and AI support as separate learning activities. Instead, participants repeatedly worked through problem-based cycles involving problem analysis, planning, technological representation, collaborative design, feedback checking, revision, and reflection. This structure may have supported self-reported TPACK because participants had to connect physics concepts with teachable representations, digital or experimental tools, learner difficulties, instructional sequencing, and classroom-oriented explanation. In this process, technology was not introduced as a stand-alone device or platform, but as part of pedagogical decision-making for specific physics teaching problems.

The same instructional structure may also explain the stronger gains in perceived CPS processes. The module did not treat collaboration as informal group work. Participants were required to negotiate shared goals, distribute responsibilities, compare alternative solutions, justify instructional choices, respond to peer feedback, and revise group products. These requirements correspond to key CPS processes, including participation, perspective taking, social regulation, task regulation, and shared knowledge construction. Therefore, the CPS gains were not incidental to the intervention. They were aligned with the collaborative demands of the CTD-PBL task cycle.

AI support should be interpreted within this broader instructional configuration. DeepSeek was used as a bounded scaffold for idea generation, resource organisation, explanation comparison, feedback prompting, and revision planning. It was not used as an autonomous source of final answers, nor was it used for data generation, scoring, statistical analysis, or research conclusion-making. Its educational role depended on the task structure around it. Participants were expected to treat AI-generated suggestions as provisional, check them against physics content accuracy, pedagogical appropriateness, classroom feasibility, safety requirements, and available evidence, and revise them through peer or instructor feedback. The observed gains should therefore not be attributed to AI alone. They are more defensibly interpreted as outcomes associated with an integrated pedagogical condition in which PBL provided the professional problem context, collaboration provided the social structure, technology provided the representational challenge, and AI provided flexible support for checking and refinement.

4.3. Relationship to Previous Literature

The findings are consistent with prior TPACK research suggesting that technology-related teacher knowledge develops more strongly when technology use is embedded in subject-specific pedagogical design rather than taught as a generic digital skill [21,22,23,24,25,26,27]. In physics teacher education, the central challenge is not simply whether pre-service teachers can operate digital tools, simulations, sensors, or AI platforms. The more important issue is whether they can transform these tools into pedagogically meaningful representations of abstract physics ideas. The stronger gains in integrative TPACK domains support this interpretation. Tasks such as magnetic field mapping, RC sensing, electromagnetic induction demonstration, and AC-to-DC transformation required participants to align physics content, technological representation, learner interpretation, and instructional explanation.

The CPS findings also align with literature on problem-based and collaborative learning, which emphasises shared problem representation, role negotiation, task regulation, feedback use, and knowledge building as important features of collaborative problem solving [28,29,30,36,38,56]. The stronger perceived CPS gains in Learning and Knowledge Building, Social Regulation, and Participation are coherent with the module design because participants had to produce shared artefacts, evaluate alternative explanations, respond to peer feedback, and revise their work toward a classroom-oriented product. This supports the view that CPS development in teacher education is more likely when collaboration is structured around meaningful professional tasks rather than left as unstructured group discussion.

The study also contributes to emerging research on generative AI in teacher education. Much existing work on AI in education focuses on perceptions, acceptance, general affordances, or policy-level claims, while fewer studies show how AI can be pedagogically governed within an implemented teacher education intervention [37,39,40,41,42,43,44,45,46]. The present study addresses this gap by positioning AI as a regulated curriculum scaffold rather than as an independent instructional solution. The contribution is therefore not that DeepSeek itself improved teacher competence, nor that one AI platform is superior to another. A more bounded contribution is that the study provides a clearly specified curriculum design case in which generative AI was integrated with explicit purposes, verification requirements, and task constraints within authentic disciplinary teaching activities. This case illustrates one possible way of organising AI-supported learning in teacher education without extending claims beyond the observed questionnaire-based outcomes.

4.4. Limitations

Several limitations define the interpretation of the findings. First, the study used an intact-class quasi-experimental design. Although pre-test comparisons indicated baseline comparability, and both groups were taught by the same instructor, individual random assignment was not possible. Therefore, the findings should be interpreted as differential changes associated with the instructional condition rather than as definitive individual-level causal effects. Selection effects and unmeasured class-level differences cannot be fully ruled out.

Second, the study examined the AI-supported CTD-PBL module as an integrated instructional condition. It did not isolate the independent contribution of AI from the contributions of problem-based learning, group collaboration, technological tasks, instructor scaffolding, peer feedback, or reflective revision. This is a methodological boundary. The findings therefore speak to the association between participation in the integrated CTD-PBL condition and stronger questionnaire-based gains, not to the isolated causal effect of DeepSeek or any single module component.

Third, the outcome measures were questionnaire-based. TPACK was measured through a self-assessment questionnaire, and CPS was measured through participants’ perceived collaborative problem-solving processes. Therefore, the findings should not be interpreted as direct evidence of objectively demonstrated teaching competence, actual classroom performance, or independently observed collaborative problem-solving behaviour. Future studies should combine self-report measures with classroom observations, group interaction records, design artefact analysis, performance-based TPACK tasks, stimulated recall interviews, and externally rated collaborative problem-solving tasks. Such evidence would help determine whether self-reported and perceived gains correspond to observable changes in professional reasoning, technology-integrated instructional design, and collaborative behaviour.

Fourth, the AI process evidence was incomplete. Although AI use was bounded by task-specific prompt templates, human verification requirements, instructor monitoring, and artefact-based documentation, complete raw DeepSeek interaction logs were not systematically archived for every group. As a result, the analysis cannot reconstruct the full sequence of prompts, model responses, rejected outputs, or revision decisions at the level of individual AI conversations. The available evidence mainly consisted of planning sheets, revision records, peer-feedback forms, reflection notes, instructor monitoring notes, and final instructional artefacts. Future studies should collect structured prompt histories, AI output samples, verification annotations, and revision traces to examine more precisely how generative AI contributes to instructional design reasoning and collaborative regulation.

Fifth, the study was conducted with third-year pre-service physics teachers at one public university in western China. The transferability of the findings to other institutions, regions, science subjects, year levels, and teacher education systems requires further examination. Implementation may also depend on instructor expertise, institutional support, access to digital tools, and participants’ prior experience with inquiry-based or AI-supported learning.

4.5. Practical Implications

The findings suggest several practical implications for physics teacher education. First, AI-supported teacher education should be organised around authentic instructional problems rather than around isolated tool training. Pre-service teachers need opportunities to use technology and AI while designing explanations, representations, demonstrations, assessment tasks, and revision plans for concrete physics teaching situations. Such tasks can help them connect physics content, pedagogical reasoning, technological representation, and learner-oriented explanation.

Second, generative AI should be introduced with explicit pedagogical boundaries. Participants should be trained to use AI for bounded purposes such as idea generation, resource organisation, explanation comparison, feedback prompting, and revision planning. They should also be trained to verify AI-generated outputs before incorporating them into instructional artefacts. Verification should include checks for physics content accuracy, pedagogical suitability, classroom feasibility, safety, and ethical acceptability. This is particularly important in physics teacher education because AI-generated explanations may be fluent but still conceptually inaccurate, oversimplified, or pedagogically inappropriate.

Third, collaborative learning should be deliberately structured. The CTD-PBL module suggests that role allocation, peer review, shared artefact production, and reflection can help connect technology-integrated instructional design with perceived CPS processes. Teacher education programmes should therefore treat collaboration as a designed learning condition, not as a simple arrangement in which students work in groups.

Fourth, implementation requires institutional support. Teacher educators need access to module materials, prompt templates, verification checklists, assessment rubrics, and examples of acceptable instructional artefacts. Institutions should also provide guidance on ethical AI use, data privacy, hallucination handling, and documentation of AI-supported learning processes. Such support is necessary if AI-supported PBL is to be implemented in a structured and reproducible manner.

Finally, for course design, programmes can organise modules around iterative problem-based cycles that explicitly require the integration of physics content, pedagogy, and technological representation, with clearly defined deliverables at each stage. For AI use boundaries, instructors should provide prompt templates, verification checklists, and explicit rules specifying when AI outputs can be used, revised, or rejected. For collaborative scaffolding, structured roles, peer-review protocols, and revision logs can be used to make group processes visible and accountable. For teacher educator preparation, training should include how to monitor AI-supported tasks, guide verification practices, and evaluate artefacts that combine human and AI contributions. These suggestions are intended as operational design considerations derived from the present case rather than as general prescriptions beyond the study context.

5. Conclusions

This study examined how generative AI can be responsibly integrated into discipline-specific teacher education through an AI-supported CTD-PBL module. Using an intact-class quasi-experimental pre-test/post-test design, it investigated whether participation in the module was associated with changes in pre-service physics teachers’ self-reported TPACK and perceived collaborative problem-solving processes. The CTD-PBL group showed stronger pre–post gains and higher post-test scores than the conventional instruction group, indicating more favourable questionnaire-based developmental outcomes. Dimensional results further suggested reported gains in integrative TPACK domains and perceived collaborative processes involving knowledge building, social regulation, and participation. These findings should be interpreted cautiously. The study did not isolate the independent effect of AI, compare different DeepSeek model versions, archive complete raw AI interaction logs, or use objective performance tasks and classroom observations as primary outcomes. Therefore, the results do not demonstrate that AI alone improved teaching competence, collaborative skill, or classroom performance. Rather, they suggest that the integrated CTD-PBL instructional condition was associated with stronger self-reported developmental patterns. The study contributes a bounded curriculum design case for AI-supported physics teacher education. In the module, AI was embedded in problem-based collaborative tasks and constrained by prompt templates, verification requirements, artefact production, peer feedback, and reflective revision. DeepSeek functioned as a scaffold for planning, checking, revising, and reflecting, while professional judgement remained with teacher educators and pre-service teachers. For practice, teacher education programmes should treat AI-supported learning as a curriculum design issue rather than simple tool adoption. Modules should include authentic disciplinary tasks, explicit AI-use boundaries, collaborative scaffolds, verification checklists, and teacher educator guidance. Future research should incorporate objective performance assessments, classroom observations, archived AI interaction data, and comparative AI-supported instructional designs.