Cognitive load theory (CLT), proposed by John Sweller in the 1980s, explores the limitations of the human cognitive system in learning processes and their impact on learning outcomes (Sweller, 1988). CLT classifies cognitive load into three types: intrinsic load, extraneous load, and germane load, emphasizing the interplay between the complexity of learning materials and learners’ prior knowledge. Intrinsic load is inherent to the content and is directly related to task difficulty (Klepsch and Seufert, 2020); extraneous load stems from the way information is presented and can be minimized through instructional design optimization; and germane load refers to the cognitive resources dedicated to knowledge integration and long-term memory formation (Seufert et al., 2007). As research on CLT has expanded, the theory has been increasingly applied in educational technology and online learning. Scholars have found that an optimal cognitive load enhances learning efficiency, while excessive cognitive load can hinder learning outcomes (Phan et al., 2017).

Sweller’s research suggests that while working memory capacity is limited, it can be extended through the use of schemas stored in long-term memory. Initially, CLT focused on instructional design strategies aimed at reducing unnecessary cognitive load to improve learning efficiency (Duy et al., 2025). Over time, the theory has been extended to encompass various instructional methods and learning environments, including multimedia learning, problem-solving approaches, and collaborative learning. More recently, with advancements in technology, CLT has been applied to AI-assisted learning, investigating how technological tools can optimize learning experiences, mitigate cognitive overload, and enhance learning efficiency (Duy et al., 2025; Khoa and Tran, 2024).

In the context of AI-assisted conversation tools, CLT provides a valuable theoretical lens for understanding how AI-driven interactions influence Perceived Usefulness (PU), Perceived Ease of Use (PEoU), and Perceived Learning Efficiency (Abdullah et al., 2016). Frequent AI-assisted dialogues offer real-time feedback and personalized learning experiences, reinforcing learners’ perception of the tool’s usefulness. This aligns with the concept of germane load in CLT, as an optimal cognitive load facilitates knowledge comprehension and application (Abdullah et al., 2016). Furthermore, as learners become increasingly familiar with AI-assisted conversation tools, the extraneous load associated with their usage decreases, thereby enhancing perceived ease of use (Lange et al., 2017; Sweller, 2010). Lastly, regular engagement in AI-mediated dialogues enables learners to apply acquired knowledge in real-world scenarios, promoting knowledge integration and transfer (Orru and Longo, 2019). This is consistent with CLT’s assertion that appropriately managed germane load contributes to improved learning efficiency.

Technology acceptance model (TAM), proposed by Fred D. Davis in 1986, is a theoretical framework designed to explain and predict individuals’ acceptance of new technologies (Davis, 1989; Davis et al., 1989a; Davis et al., 1989b). Rooted in the Theory of Reasoned Action (TRA), TAM posits that an individual’s acceptance of a new technology is primarily determined by their cognitive perceptions and attitudes toward it (Davis et al., 1989a; Davis et al., 1989b). The model comprises two core constructs: Perceived Usefulness (PU) and Perceived Ease of Use (PEoU). PU refers to the extent to which an individual believes that using a particular technology will enhance their performance, whereas PEoU reflects the degree to which an individual perceives the technology as easy to use (Venkatesh et al., 2003). These two factors influence users’ attitudes and behavioral intentions, ultimately determining their acceptance of the technology.

Integrating TAM with constructs like CPD and COA (from Cognitive Load Theory) is theoretically sound and supported by literature that extends TAM with cognitive and motivational variables. In this study, TAM is employed to analyze learners’ acceptance of AI-assisted conversation tools and their impact on learning efficiency (Sánchez-Prieto et al., 2020). AI-assisted conversation tools, through real-time interactions, personalized feedback, and contextual simulations, have been shown to significantly enhance both learning efficiency and user satisfaction (Otto et al., 2024). According to TAM, learners’ perceptions of PU and PEoU are critical determinants of their acceptance and continued use of AI tools. Specifically, AI-assisted tools facilitate more efficient language learning, thereby enhancing learners’ PU (Fedzechkina et al., 2012). Simultaneously, their user-friendly design and intuitive operation reduce the perceived difficulty of use, lowering the PEoU threshold and further promoting acceptance and adoption.

A mixed-method investigation by Eymur and Çetin (2024) explored preservice teachers’ acceptance of ChatGPT for metacognitive self-regulated learning and confirmed that perceived usefulness (PU), perceived ease of use (PEOU), enjoyment, and trust significantly influenced behavioral intention, reinforcing TAM’s applicability in AI-enhanced education. Similarly, a study on college-level second-language learners demonstrated strong predictive relationships between PU and PEOU and actual usage of AI tools, indicating TAM’s validity in language acquisition contexts (Wang et al., 2024). Meta-analytic evidence further supports TAM’s explanatory power. A synthesis of educator adoption studies by Rothberg et al. (2023) identified TAM as the dominant model explaining AI integration across K–12 and higher education settings. This aligns with case studies in higher education where TAM was effectively applied to evaluate instructor acceptance of AI-assisted grading systems (Salem et al., 2023). Beyond core constructs, TAM has been successfully integrated with cognitive and motivational dimensions. The model has also been employed in more complex frameworks such as TPACK-TAM hybrids to evaluate video generative AI adoption among K–12 teachers, highlighting TAM’s flexibility in capturing new AI modalities (Li et al., 2024). The theoretical robustness of TAM is further evidenced by its extended versions (TAM2, TAM3) and by its influence on later models like UTAUT, which maintain PU and PEOU as foundational constructs (Venkatesh et al., 2003).

Furthermore, this study introduces Cognitive Processing Depth (CPD) and Cognitive Ability (COA) as moderating variables to examine how individual differences influence the effectiveness of AI tools in improving learning efficiency (Bai, 2024; Low et al., 2025). By integrating TAM, this research not only elucidates the mechanisms through which AI technology contributes to education but also offers new perspectives and practical insights for the development of personalized learning strategies (Liu, 2023).

This study will employ Partial Least Squares Structural Equation Modeling (PLS-SEM) to analyze the relationships between constructs within the research model (Ringle et al., 2020). PLS-SEM is chosen because it can handle complex models with multiple constructs and is well-suited for exploratory research. The study will assess the reliability and validity of the measurement model, including reliability indicators (Cronbach’s alpha, reported for reference, and composite reliability as a more robust index of internal consistency), convergent validity (using the Average Variance Extracted, AVE), and discriminant validity (using the Fornell-Larcker criterion and the Heterotrait-Monotrait ratio, HTMT). The structural model will be evaluated to test the hypothesized relationships between constructs. Path coefficients, t-values, and R² values will be examined to determine the strength and significance of these relationships. Bootstrapping with 5,000 subsamples will be used to assess whether the estimated structural paths are statistically significant (Streukens and Leroi-Werelds, 2016). Additionally, Necessary Condition Analysis (NCA) will be conducted to identify the critical conditions under which AI-assisted tools enhance learning efficiency. This dual-method approach strengthens the robustness of the findings by addressing both statistical association and necessary conditions.

To collect data, a structured questionnaire was designed using validated scales adapted from existing literature. The questionnaire consists of four sections: (1) demographic information, (2) frequency and context of AI-assisted tool usage, (3) perceived usefulness and ease of use of the tool, and (4) self-reported learning efficiency, cognitive processing depth, and cognitive ability. All items are measured on a five-point Likert scale, ranging from “strongly disagree” to “strongly agree.” The target population includes English learners with experience using AI-assisted conversational tools. A purposive sampling technique is employed to ensure that participants have relevant experience with the tools under investigation.

In this study, data collection was conducted via an online questionnaire hosted by Wenjuanxing, a widely recognized survey platform in China. Wenjuanxing provides access to a diverse participant pool exceeding 2.6 million respondents, offering specialized paid sampling services that enhance both sample representativeness and data reliability (Wang et al., 2021). To ensure that survey respondents accurately matched the targeted user characteristics, the platform incorporated screening procedures at the survey outset. These screening questions effectively filtered potential respondents, guaranteeing that participants included in the final sample had relevant experience with AI-assisted conversational tools in English language learning contexts.

To further strengthen data quality and validity, multiple quality-control strategies were employed (Su et al., 2024). Firstly, the platform utilized automated mechanisms to detect and remove potentially invalid responses, such as those from duplicate IP addresses, participants completing the survey in unusually short times, or respondents exhibiting suspicious answering behaviors (Lu et al., 2021). Secondly, only responses completed within a realistic and reasonable duration were retained for subsequent analysis; this step ensured that the data represented thoughtful and authentic participant engagement (Wang et al., 2021). Thirdly, attention-check questions were strategically embedded throughout the survey to identify and exclude inattentive or careless respondents.

Moreover, ethical standards were strictly upheld throughout the research process, including obtaining informed consent from participants, emphasizing voluntary participation, and ensuring anonymity and confidentiality during data collection and analysis. The rigorous sampling approach and comprehensive data validation techniques provided by the Wenjuanxing platform contributed substantially to the overall reliability and representativeness of the dataset, thereby laying a solid methodological foundation for subsequent empirical analyses.

The target population of this study consisted exclusively of undergraduate and graduate students enrolled at University. We distributed the online questionnaire via the platform and student mailing lists, ensuring voluntary participation. A total of 330 questionnaires were collected, of which 297 were valid after data cleaning. Thus, the analyses were conducted based on these 297 responses. Table 1 presents the demographic characteristics of the participants, including gender, age, and frequency of AI-assisted tool usage.

Regarding gender, 207 respondents (69.7%) were female, while 90 respondents (30.3%) were male, indicating a notable gender imbalance. In terms of age distribution, 46.8% of participants were 26 years or above, followed by 33.3% aged 18–20, 12.8% aged 21–22, and 7.1% aged 23–25, suggesting that the majority of the sample consisted of students in Tertiary Education encompassing all levels.

In terms of AI-assisted tool usage, the largest proportion of participants reported using such tools 4–8 times per week (37.4%, n = 111), followed by 1–3 times per week (28.6%, n = 85), 9–15 times per week (16.8%, n = 50), and 16 times or more (5.7%, n = 17). Notably, 11.4% (n = 34) indicated that they had never used AI-assisted tools.

Empirical analysis is typically conducted using PLS-SEM. PLS-SEM does not impose distributional requirements on data and can accommodate extended models based on existing theories (Hair et al., 2021). PLS-SEM combines factor analysis and multiple regression modeling to explore the path relationships between observed and unobserved variables. Therefore, this study employs PLS-SEM for empirical analysis and utilizes Smart-PLS 4.0 software. Smart-PLS offers high estimation accuracy and user-friendly data processing operations.

Necessary condition analysis (NCA) is an innovative research method explicitly designed to identify necessary conditions within complex statistical associations—that is, conditions that are indispensable for the occurrence of an outcome variable (Dul, 2016b). Unlike traditional analytical approaches such as regression or structural equation modeling, NCA not only identifies the existence of necessary conditions but also quantifies their strength and scope of constraint, thereby effectively uncovering the distinctive “necessary but not sufficient” relationships between independent and dependent variables. By calculating the minimal levels of necessary conditions required to achieve specific outcome thresholds—known as “ceiling lines”—NCA precisely pinpoints bottleneck factors and assesses their effect sizes (Radanliev, 2025; Ye and Kuang, 2025). Thus, it effectively complements traditional sufficiency analyses and provides a more comprehensive explanatory framework for causal mechanisms.

A typical application of NCA involves two main stages. First, latent variable scores are extracted through Partial Least Squares Structural Equation Modeling (PLS-SEM), ensuring that the data fulfill the modeling requirements for multi-level statistical associations (Richter et al., 2020). Second, the NCA analysis is conducted using dedicated analytical packages integrated within statistical software such as SMART-PLS (e.g., the NCA plug-in), following standardized procedures outlined by Dul (Dul, 2016b). Specifically, NCA employs “ceiling lines” to delineate the boundaries of necessary conditions. These ceiling lines, drawn tangentially to the upper-leftmost data points within an x-y scatter plot, visually illustrate the minimal necessary level of an independent variable required to achieve certain levels of the dependent variable (see Figure 2). The strength of this approach lies in its integration of statistical rigor and visual interpretability, allowing researchers not only to rigorously test theoretical hypotheses but also to generate actionable and practical recommendations based on identified constraints.

Subsequently, we conducted a further analysis to statistically assess the significance of effect sizes (d) associated with the latent variable scores, using a bootstrapping procedure with a randomized sample size of 5,000 iterations (Dul, 2016a; Piff et al., 2015). Due to the suitability of the Ceiling Regression-Free Disposal Hull (CR-FDH) method for analyzing survey data gathered via five-point Likert scales, the results obtained from the Necessary Condition Analysis (NCA) were highly consistent with this analytical approach. The detailed results are presented in Table 7. Specifically, the NCA results revealed notable interrelationships among the constructs. For instance, the effect size of AIC on PU was found to be 0.195 (p < 0.001), while the effect size of PEoU on PU was 0.166 (p < 0.001), indicating that AIC is a necessary condition for PU. Similarly, the effect size of AIC on PEoU was 0.261 (p < 0.001), confirming that AIC is also a necessary condition for PEoU. Furthermore, the effect size of PEoU on LEF was 0.173 (p < 0.001), demonstrating that PEoU is a necessary condition for LEF.

However, the effect size of PU on LEF was only 0.015 (p = 0.809), failing to reach statistical significance (Table 7). Thus, PU does not satisfy the criteria for being a necessary condition for LEF. These findings collectively suggest that while AIC and PEoU are crucial for enhancing PU, and AIC is essential for improving PEoU, and PEoU is critical for enhancing LEF, PU represents an influential but not necessary factor for fostering LEF.

The bottleneck technology of LEF further clarifies the threshold levels required to achieve specific performance levels. As shown in Table 8, to reach a 40% PU level, AIC must be no less than −3.441, and PEoU must be no less than −2.906. To achieve a 40% PEoU level, AIC must be at least −3.14. To attain a 50% LEF level, AIC must be no less than −2.518, and PEoU no less than −2.347. To achieve a 100% PU level, the necessary conditions include AIC not lower than −1.683 and PEoU not lower than −0.079. Additionally, for PU to reach a 100% level, AIC must be no less than 0.71. To achieve a 100% LEF level, AIC must be no less than 0.7, PEoU no less than −0.325, and PU no less than −1.796.

This study employed a combined NCA and PLS-SEM approach to examine the mechanism of learning efficiency of EFL acquisition through AI-assisted conversational tools, providing a comprehensive understanding and analysis.

To ensure the robustness of the measurement model, we conducted a thorough assessment of reliability and validity. Reliability was evaluated using Cronbach’s alpha and composite reliability (CR). All constructs demonstrated high reliability, with Cronbach’s alpha and CR values exceeding the threshold of 0.7, indicating substantial internal consistency and stability of the questionnaire scales. These findings, as shown in Table 2 and Figure 2, confirmed the high reliability of the questionnaire data. The validity of the measurement model was assessed through convergent validity and discriminant validity. All indicators had average variance extracted (AVE) values exceeding 0.5, indicating robust convergent validity (Fornell and Larcker, 1981). Discriminant validity was confirmed using the Fornell-Larcker criterion and the HTMT (Heterotrait-Monotrait) ratio. The square root of each variable’s AVE exceeded its correlation coefficients with other dimensions, and the HTMT ratios were below the threshold of 0.85 (Sarstedt et al., 2014a), indicating significant discriminant validity of the measurement model.

To assess collinearity, we examined the variance inflation factors (VIF) for all predictive constructs in the structural model. All VIF values ranged from 1.692 to 2.648, well below the cutoff value of 3, indicating that collinearity was not a significant issue in the model. Subsequently, we performed bootstrapping with 5,000 subsamples to evaluate the significance of the hypotheses (Hair et al., 2021). The analysis showed that the paths from CPD to PU, from CPD × AIC to PEoU, from CPD × AIC to PU, and from COA × AIC to PEoU were not significant, thus not supporting the hypotheses. Despite these insignificant paths, the model explained a substantial amount of variance in the dependent variables: R² for PEoU was 0.389 with a Q² predict of 0.249; R² for PU was 0.088 with a Q² predict of 0.058; and R² for LEF was 0.402 with a Q² predict of 0.258. The Q² values further confirmed the predictive relevance of the model, as they were all above zero. Effect size analysis (f²) indicated that multiple predictor constructs had significant effects on the dependent variables. NCA provided a unique approach to understanding complex statistical associations by identifying necessary conditions that impact the outcome variable. Unlike traditional methods, NCA quantifies the size and constraints of these necessary conditions, making it particularly adept at distinguishing “necessary but not sufficient” relationships between the dependent and independent variables (Dul, 2016b). As an enhancement to traditional sufficiency analysis, NCA offers numerical measurements of the prerequisites needed to achieve a given outcome level, thereby providing deeper insights into effect sizes and potential bottlenecks.

Initially, the PLS-SEM method was used to obtain scores for the latent variables (Richter, 2020; Richter et al., 2020). Subsequently, NCA analysis was performed using the NCA package within the Smart-PLS software, following the guidelines set by Vis and Dul (2018). The fundamental step of NCA involved plotting a ceiling line that intersected the upper-left data points on an x-y scatter plot. Figure 2 illustrates the scatter plots for all relevant relationships.

Subsequent analyses evaluated the statistical significance of the effect sizes (d) associated with the latent variable scores, using 10,000 random samples (Dul, 2016b; Piff et al., 2015). Given that the CR-FDH (Cumulative Relative Frequency - Full Disjunctive Hypothesis) line is well-suited for survey data derived from a five-point Likert scale, the interpretation of the NCA results was consistent with this method. As shown in Table 6, the results indicated that AIC (d = 0.195, p < 0.000) and PEOU (d = 0.166, p < 0.001) were necessary conditions for PU adaptation; AIC (d = 0.261, p < 0.001) was a necessary condition for PEOU; similarly, PEOU (d = 0.173, p < 0.000) was a necessary condition for LEF. However, PU (d = 0.015, p = 0.809) was not significant and thus did not constitute a necessary condition for LEF.

This study focuses on investigating the impact of AI-assisted conversational tools on English learning efficiency (LEF). Employing quantitative analysis centered around key variables such as perceived usefulness (PU), perceived ease of use (PEoU), interaction frequency of AI-assisted conversation (AIC), and learning efficiency (LEF), the research develops an integrated theoretical framework that synthesizes the Technology Acceptance Model (TAM) with Cognitive Load Theory (CLT). This integrative approach has yielded several important theoretical contributions. This research enriches the theoretical scope of TAM by incorporating insights derived from CLT. Traditional TAM primarily emphasizes users’ subjective acceptance intentions toward technology use. By contrast, this study further elucidates the mediating effect of cognitive processing depth (CPD)—triggered by AIC usage—on learning efficiency. This finding aligns with Gkintoni et al. (2025) cognitive load model, which posits that learning outcomes depend not merely on technological attributes but also on individuals’ cognitive processing strategies. Hence, the present research provides theoretical support for understanding the cognitive mechanisms underlying technology acceptance behaviors.

By adopting Necessary Condition Analysis (NCA), this research explicitly identifies the threshold values of key variables required to achieve specific levels of learning efficiency. For instance, to achieve a 100% level of PU, the analysis reveals that AIC must not fall below 0.71, and PEoU must not be lower than −0.079 (see Table 7). Serving as a valuable complement to Partial Least Squares Structural Equation Modeling (PLS-SEM), the NCA method effectively identifies the “bottleneck” roles of certain variables—that is, critical thresholds below which the desired outcomes become unattainable (Dul, 2016b). This methodological approach addresses the inherent limitation of traditional path analysis in failing to identify critical cut-off points, thereby enriching the boundary conditions research of TAM and CLT in practical applications. This study introduces cognitive ability (COA) as a moderating variable, revealing the role of individual differences in moderating AIC learning outcomes. Empirical results indicate that learners with higher cognitive abilities exhibit greater CPD when using AIC, thereby enhancing their learning efficiency. This finding is consistent with the theory of Self-Regulated Learning (Zimmerman, 2002), which suggests that individuals with higher cognitive capabilities are better at autonomously regulating their learning paths and strategies. The moderating role of COA underscores the necessity of considering individual differences in future technological designs and educational interventions, advocating for the creation of more cognitively adaptive AI systems. This study fills an existing research gap in AI education literature concerning non-native English learners (specifically, the EFL population). Current research on AI-assisted learning has predominantly focused on STEM disciplines and native English-speaking student populations (Luckin et al., 2016). By concentrating specifically on Chinese university students’ English learning contexts, this research expands the applicability of the TAM and CLT frameworks to cross-cultural and second-language educational environments.

The introduction of CPD and COA variables offers novel explanatory perspectives for AI educational technology research. While previous studies have frequently emphasized learning attitudes and behavioral intentions, they have largely overlooked how technological integration stimulates deeper cognitive processing pathways. By constructing an integrated model, the present study provides a new cognitive psychological interpretation of how AI technology “reshapes learning itself.” In sum, this research contributes theoretically in four significant ways: (1) deepening the understanding of AIC acceptance mechanisms through the integration of TAM and CLT; (2) employing NCA to identify critical thresholds at which variables influence learning efficiency; (3) elucidating the mediating mechanism of CPD and the moderating mechanism of COA; and (4) extending the theoretical scope of AI educational research into language learning within humanities and cross-cultural contexts. These theoretical insights not only offer a robust framework for future model optimization but also provide cognitive-level guidance for technological design and educational interventions.

With the rapid advancement of artificial intelligence technologies in educational domains, AI-assisted conversational tools have increasingly emerged as crucial supplementary resources in university-level English instruction. However, the effective implementation of such technologies is influenced not merely by their inherent sophistication, but more importantly, by their deep integration with learners’ cognitive characteristics, usage habits, and the instructional environment. This study provides several significant practical implications at multiple levels, as detailed below.

Beyond the statistical significance reported in the structural model, this study highlights a particularly noteworthy path: the positive effect of Perceived Ease of Use (PEoU) on Learning Efficiency (LEF). This result bears important pedagogical value in the context of AI-assisted conversational learning tools, especially for university-level English learners who frequently face cognitive overload and self-regulation challenges. Specifically, the finding suggests that when students perceive AIC interfaces as easy to navigate and interact with, their mental resources can be redirected from operational effort to language acquisition tasks, such as grammar refinement, argumentation, or vocabulary contextualization. This aligns with Cognitive Load Theory (CLT), which posits that reducing extraneous load—such as tool design or unintuitive input formats—can enhance germane processing and learning transfer.

Moreover, the findings indicate that cognitive processing depth (CPD) serves as a crucial mediator in the relationship between AIC use and learning efficiency. Consequently, AI tool developers should emphasize aligning system architecture with learners’ cognitive mechanisms. For instance, dialogue design could incorporate adjustable feedback modes (e.g., basic, reflective, inferential feedback), enabling the system to dynamically generate cognitively stimulating feedback based on learners’ current progress and performance, thus enhancing cognitive engagement.

Additionally, adaptive adjustment mechanisms should be incorporated into AIC systems to accommodate learners with varying cognitive abilities, as the empirical evidence confirms that cognitive ability (COA) significantly moderates learning outcomes. AI systems could automatically assess users’ cognitive capacities through historical interaction data, linguistic input quality, and task completion rates, subsequently matching appropriate interactive tasks. For example, higher cognitive-ability learners could be provided with tasks emphasizing critical thinking and language generation, whereas lower cognitive-ability students could receive more structured, scaffolded tasks to minimize cognitive load.

Furthermore, at the quantitative level, this study introduces Necessary Condition Analysis (NCA) to identify the critical threshold values for variables such as AIC, PEoU, and PU required to achieve specific levels of learning efficiency. For example, to reach a 50% LEF level, the minimum required threshold for AIC is −2.518, and for PEoU is −2.347. This analytical approach can serve as a dynamic diagnostic model for assessing instructional quality in higher education, shifting the evaluation focus from purely outcome-based indicators to process-oriented measures, thereby providing real-time insights into instructional bottlenecks. Educational managers could integrate these critical thresholds into backend monitoring systems of AI-assisted learning platforms, facilitating real-time tracking of students’ key variable performance. Timely interventions, including tailored learning recommendations and technical assistance, could thus be implemented when students persistently exhibit sub-threshold levels of AIC or PEoU, ensuring personalized instruction and differentiated service provision.

Regarding enhancing user experience and learner engagement, it is advisable to empower learners with greater control over their learning process, such as customizing their learning pace, choosing feedback frequency, and providing suggestions, thereby reinforcing the learner-centered paradigm and preventing learning apathy or technological inertia arising from excessive technological dependence. Comprehensive data ethics and privacy protection mechanisms should be established. AI systems must comply with relevant regulatory standards (e.g., the Personal Information Protection Law) when collecting and processing student data, ensuring data security, transparency of use, and controllability of outcomes. Educational institutions should also facilitate training programs and institutional arrangements to cultivate students’ healthy AI usage habits and robust data security awareness.