Integrating Sustainable Development Goals (SDGs) and Generative AI to Enhance Language Digital Literacy and Creativity in EFL Learning Environments

The global landscape of English as a Foreign Language (EFL) education is undergoing a profound transformation, driven by the dual imperatives of technological innovation and the United Nations' Sustainable Development Goals (SDGs)¹. While the demand for English proficiency continues to surge in non-Anglophone regions, fueled by globalization, academic mobility, and digital economies², traditional pedagogical approaches often fail to equip learners with the critical competencies needed for 21st-century communication³. Conventional EFL methodologies, which emphasize rote memorization and structural accuracy over creative expression and digital engagement, risk exacerbating disparities in access to quality education⁴. This misalignment not only hinders progress toward SDG 4 (inclusive and equitable quality education) but also overlooks the transformative potential of emerging technologies⁵, particularly generative artificial intelligence (AI), in reshaping language learning ecosystems⁶. Despite the rapid expansion of AI-driven tools like ChatGPT, their integration within EFL contexts remains underexplored, particularly in terms of sustainability frameworks⁷. This study bridges this critical gap by investigating how the strategic convergence of SDGs and generative AI can enhance digital literacy, creativity, and motivation in EFL learning environments, advancing pedagogical innovation and global sustainability agendas. Significant gaps persist in the existing literature regarding the intersection of AI-assisted language learning and SDG-aligned education⁸. While prior research has examined AI's role in personalized language instruction⁹, automated assessment¹⁰, and virtual exchange programs¹¹, these studies rarely engage with sustainability frameworks. Conversely, literature on EFL and sustainability has predominantly focused on content-based instruction (environmental texts) rather than leveraging technology to restructure pedagogical paradigms¹². This disconnect is particularly problematic given the escalating digital divides in Global South contexts, where linguistic marginalization intersects with limited access to cutting-edge educational tools¹³. Furthermore, while SDG 4.4 explicitly prioritizes digital literacy, most EFL curricula fail to link digital competence with sustainability outcomes, leaving learners ill-prepared for the demands of employment, entrepreneurship, and decent work¹⁴. The absence of an integrative framework that aligns AI-driven pedagogy with sustainability competencies represents a critical oversight¹⁵. Moreover, the framework is readily applicable in undergraduate EFL courses where students possess basic digital literacy for operating generative AI tools like ChatGPT. Implementation is designed for standard technology-integrated classrooms, requiring educators to provide structured, SDG-aligned task prompts to guide the AI collaboration effectively.

This study fills these gaps by proposing a novel, SDG-informed generative AI framework for EFL education, grounded in sociocultural theory¹⁶ and critical digital literacies¹⁷. Unlike prior work that treats AI as an instructional tool, researchers position it as a collaborative partner in scaffolding learners' creative and critical thinking¹⁸. This shift aligns with SDG targets on inclusive, learner-centered education. Empirically, researchers employ partial least squares structural equation modeling (PLS-SEM) to analyze data from n = 420 Chinese undergraduate EFL learners, examining how SDG integration (SDG) and ChatGPT use (UCGPT) jointly influence motivation (SM), digital literacy (DL), Students 'creativity (SC), and language learning engagement (LLE). The theoretical contributions of this study are threefold. First, researchers extend self-determination theory (SDT) by demonstrating how SDG-aligned AI tools satisfy learners' intrinsic needs for autonomy, competence, and relatedness, enhancing motivation. Second, researchers refine engagement theory by revealing how digital literacy and creativity interact dynamically in AI-mediated environments, contingent on learners' engagement levels. Third, researchers advance critical digital literacies by embedding ethical AI use within sustainability discourse, urging learners to interrogate data biases, environmental costs, and labor practices in AI systems. Empirically, the findings provide actionable insights for educators and policymakers, illustrating how ChatGPT can democratize access to SDG-related language tasks (climate change debates, gender equality simulations) while fostering higher-order cognitive skills. This article is structured as follows: First, the introduction that discusses the significance/originality of this research, the existing gap in the literature, and the contribution of this study. Next, there are reviews of the literature on SDGs in EFL, generative AI, and digital literacies. The Protocol section details the PLS-SEM methodology, while the Results section presents the findings. The Discussion section elaborates on the key findings and theoretical and practical implications. Finally, the article concludes with limitations and future directions.

Literature review and hypotheses development

Sustainable Development Goal 4 and EFL Students' Motivation
A pedagogical approach that integrates English as a Foreign Language (EFL) curricula with Sustainable Development Goal 4¹⁹, which emphasizes inclusive and equitable quality education and lifelong learning for all, has emerged as a key strategy for enhancing student motivation²⁰. SDG 4's targets, specifically 4.4 (skills for employment) and 4.7 (education for sustainable development), underscore the importance of developing linguistically competent learners with critical thinking, digital literacy, and ethical engagement skills²¹. The principles of these competencies are additionally supported by the Responsibility to Protect and the 'Rights-Based Education' SDG 4.5 targets. Several studies suggest that when climate action, gender equality, and fighting poverty are covered in language learning, it promotes students' motivation for the subject²². This method supports the Self-Determination Theory^23,24, which believes that the experience of autonomy, mastering skills, and social connection keeps people motivated²⁵. Generative AI works even better as a motivation tool when interactive content related to Sustainable Development Goals is used²⁶. AI with chatbot simulated summits and platforms in several languages that focus on sustainability make it possible for students to learn about complex subjects through language exercises²⁷. Applying scenarios that require students to discuss water scarcity solutions in English helps them become interested. It sharpens their problem-solving abilities and gives them a greater sense of being in charge and importance²⁸. AI feedback is standardized and given instantly (which was pointed out, helping meet the diverse educational needs, and it supports equity as explained in SDG 4²⁹, all of which calms students and ensures equality. Educational experts need to take SDG connections a step further since they should focus on developing skills that enable students to identify systemic inequalities³⁰. When there is an AI gap between developed and underdeveloped countries, this increases the number of reasons for students to feel discouraged and calls for inclusive design to be adopted³¹. Based on the above literature, this hypothesis is proposed: H1: Integrating SDGs into EFL lessons increases students' motivation to learn English.

Generative AI Usage and EFL Students' Motivation
Generative AI tools, which include language models and chatbots, are increasingly valued for helping to boost motivation in EFL education by encouraging interaction, addressing personal needs, and promoting creative activities³². These approaches enable learners to interact directly with AI, practice communicating in realistic situations, and receive quick and adjustable feedback, which is known to lower anxiety and sustain their interest³³. For instance, AI-generated educational resources, such as personalized texts and scenarios tailored to a student's level and interests, can meet their psychological needs for autonomy and a sense of capability. Since generative AI can provide resources in multiple languages and consider cultural contexts, it enables learners to explore global issues in English, prompting them to think deeply about ethical decisions³⁴. However, experts advise that relying too heavily on AI may limit communication among people and prevent them from developing critical thinking, both of which are essential for children's growth. In addition, disadvantages in internet access and the effects of biased algorithms can exacerbate existing inequalities for certain learning groups, further deepening current disparities. Based on the above literature, the following hypothesis is proposed: H2: Using ChatGPT in EFL classrooms increases students' motivation to engage in language learning.

EFL Students' Motivation and Digital Literacy
The relationship between digital literacy and EFL motivation is not merely correlational but synergistic. Students with higher digital literacy are better equipped to seek out and engage with authentic, multimodal content that aligns with their personal interests and sociocultural contexts^35,36. This capacity directly fuels intrinsic motivation by enhancing learners' sense of autonomy and competence, as they can navigate digital environments more effectively. However, this positive cycle is not automatic. The motivational impact is contingent on the quality of digital engagement; simply using technology does not guarantee motivation unless the tasks are meaningful and culturally responsive³⁷.

Moreover, although Generative AI is currently being marketed as a mechanism to accelerate this synergy with the help of personalized scaffolding^38, a critical approach reveals that there are several contingencies to this. The success of AI technologies in facilitating cross-cultural dialogue and demystifying media bias³⁹ is based on equal access and underlying digital competency, which is frequently lacking in under-resourced environments⁴⁰. This results in a motivational gap, with digital illiteracy causing frustration and a sense of disengagement, thereby sabotaging the very motivation AI aims to generate⁴¹. Thus, AI integration should be accompanied by rigorous digital literacy models that encourage the learner to understand ethical connotations, so that inadvertent propagation of current imbalances does not occur with the use of technologies⁴². Based on the above literature, the following hypothesis is proposed: H3: Higher motivation in EFL students leads to improved digital literacy skills.

EFL Students' Digital Literacy and EFL Students' Creativity
In English as a Foreign Language (EFL), digital literacy catalyzes creativity by empowering learners with the technical and critical ability to re-contextualize language use by innovative digital modes⁴³. Students become proficient at using devices such as generative AI, multimedia platforms, and collaborative virtual spaces, allowing them to move beyond text-based exercises and use them as tools for creating multimodal projects like digital stories, podcasts, or even AI-produced narratives⁴⁴. For example, AI-based platforms for co-creating multilingual poetry or simulating virtual reality scenarios help learners to experiment with linguistic and cultural hybridity and combine language acquisition and artistic exploration⁴⁵.

Digital literacy correlates with creativity through equal access opportunities and teaching approaches with defined purposes⁴⁶. Some marginalized groups lack equal access to generative AI and digital tools, resulting in creative stifling and passive learning rather than active creation⁴⁷. Educational professionals should select critical methods that focus on sustainable teaching practices to help their students examine both the ethical approach to AI content generation and the digital workflow impacts on the environment⁴⁸. Students can develop ethical craftsmanship through hands-on digital design activities that combine sustainable project development with algorithmic bias examination, strengthening SDG 4.7's sustainability and justice targets. Based on the above literature, the following hypotheses are proposed: H4: EFL students with stronger digital literacy skills demonstrate greater creativity in language tasks. H5: EFL students' language learning engagement demonstrates greater creativity in language tasks.

Moderating Effect of Language Learning Engagement
Specifically regarding EFL learning contexts, pedagogical interventions that integrate Sustainable Development Goals (SDGs) and generative AI are more effective when language learning engagement, as defined by the cognitive, behavioral, and emotional investment in language acquisition, is a multilaterally pivotal moderating variable⁴⁹. In implementing SDG-related content, high levels of engagement increase the motivational and cognitive gains, as participants in SDG-aligned project-based tasks or AI-simulated tasks are more persistent, ask more questions, and claim more ownership over their learning journeys⁵⁰. For example, student co-creation of artificial intelligence (AI) sustainability campaigns or discussion of global issues in English are shown to enhance linguistic capabilities and creativity, given the students' emotional connection to real-life problems⁵¹. Engagement is a moderator determining how learners handle ethical and cultural aspects of EFL learning environments driven by AI. Students who demonstrate disengagement view AI as an authoritarian system that intensifies their concerns and leads them to passively depend on automated feedback⁵². Based on the above literature, the following hypothesis is proposed: H6: Language learning engagement moderates the relationship between digital literacy and creativity.

Mediating Effect of EFL Students' Motivation in EFL Settings
The motivational levels of EFL students play a vital role in how pedagogical strategies linking to SDGs and using generative AI affect their digital skills and creative abilities as well as their proficiency in the English language⁵³. Instructional design meets learner achievement through motivation, which uses Self-Determination Theory⁵⁴, to strengthen psychological conditions for deep engagement. Student motivation related to autonomy (decision-making for advocacy topics), competence (learning sustainability vocabulary), and relatedness (participating in global activities) serves as the mechanism through which SDG-themed content determines the effectiveness of students' critical thinking and digital skill development. Student engagement can rise during the first use of generative AI tools because of their newness, but ongoing improvements in literacy and creativity need students to recognize AI interactions as both identity-relevant and autonomy-promoting⁵⁵.

Researchers establish motivation as the primary factor that connects various settings through empirical evidence. SDG-focused projects exposed to EFL learners led to associated creativity because the learners experienced heightened motivation levels simultaneously, as reported by self-assessments⁵⁶. Writing outcomes from AI-generated personalized feedback showed maximum improvements in students who demonstrated strong existing motivation^57,58. The study proved that motivation regulates the effectiveness of technological interventions. Another factor that hinders this process is cultural, along with socioeconomic elements. Resource-challenged settings tend to prioritize external incentives, which relate to career opportunities mentioned in SDG 4.4, resulting in altered relationships between motivation and digital tool delivery^59,60. Based on the above literature, the following hypotheses are proposed: H7: Students' motivation mediates the relationship between ChatGPT use and digital literacy skills. H8: Students' motivation mediates the relationship between SDGs and digital literacy skills. Figure 1 shows the research model of the present study, which includes all variables and their path relationships with each other. Moreover, based on the path relationships between variables, all hypotheses are also indicated in the research model.

This study was conducted in compliance with ethical guidelines for human subject's research. Informed consent was obtained from all participants prior to their involvement, ensuring they were fully informed of the study's purpose, survey-based procedures, voluntary nature, and their right to withdraw at any time without consequence. All methods adhered to ethical standards for human research, ensuring participant confidentiality, data anonymization, and protection throughout the study. The study involved non-invasive, survey-based data collection with minimal risk to participants.

1. Research design

This study adopted a quantitative, cross-sectional research design to examine the relationships between Sustainable Development Goal (SDG) integration, generative AI usage, digital literacy, creativity, motivation, and language learning engagement in EFL contexts. The design prioritized hypothesis testing through structured surveys, grounded in positivist epistemology, enabling statistical generalization of findings across similar academic settings. A partial least squares structural equation modeling (PLS-SEM) approach was selected to analyze complex mediation and moderation effects. PLS-SEM accommodates smaller sample sizes and exploratory model structures while minimizing assumptions about data normality⁶¹. The design aligned with SDG 4's call for data-driven educational strategies, focusing on measurable outcomes to inform scalable pedagogical innovations in AI-enhanced language learning.

2. Data and sampling

First, to construct the sampling frame, researchers identified and selected all intact, technology-integrated EFL classes from the university's program roster for the Spring 2024 semester, resulting in a convenience sample of 420 undergraduate students. Second, participant eligibility was confirmed against the criteria of being aged 18-24 and enrolled as first- or second-year undergraduates with an intermediate level of English proficiency, as per university placement records; the final cohort comprised 62% female and 38% male participants, with 95% originating from Eastern China. Third, during a scheduled class session, the researcher administered the survey instrument in person: after distributing a detailed information sheet, the researcher read aloud the informed consent script, allowed time for questions, and collected signed consent forms before distributing the questionnaire, which participants had 20 min to complete anonymously with no incentives provided. Fourth, to control for the extraneous variable of prior AI experience, the survey instrument itself incorporated a dedicated filter section at the outset, capturing data on participants' familiarity with AI tools for subsequent statistical control. Finally, the collected data were screened for completeness, and the sample size of n = 420 was verified for statistical adequacy by applying the "10 times rule" for PLS-SEM, confirming it exceeded the minimum requirement of 10 cases per predictor variable in the model's most complex regression equation.

1. Practical integration of SDGs and ChatGPT
   To ground the survey measures in a concrete pedagogical context, the study was conducted within a specific EFL module that explicitly integrated SDG themes with ChatGPT assistance. Students participated in a series of task-based learning activities over a four-week period. For instance, one task required students to use ChatGPT to research a specific SDG (SDG 4 education, SDG 13: Climate Action, or SDG 5: Gender Equality), generating a summary of key issues and then critically evaluating the AI-provided information for potential bias. In another assignment, students used ChatGPT as a collaborative tool to co-write a persuasive essay or a social media campaign script advocating for a local sustainability initiative, thereby connecting global goals to their immediate context. These tasks were designed to be multimodal, often culminating in video blogs or presentations, requiring students to synthesize information, exercise critical thinking, and use English for a tangible purpose. This practical implementation ensured that students' survey responses were informed by direct experience with the integrated SDG-AI learning approach.

3. Ethical considerations

The present study was conducted in accordance with rigorous ethical considerations. All participants were informed about the purpose and procedures of the study, as well as their rights, in detail through both written and oral means, and then provided informed consent before participating in the study. This agreement ensured anonymity, allowing them to withdraw at any time without reprisals and guaranteeing the confidentiality of their data. The information was gathered through encrypted digital systems and stored securely in password-protected servers. All analysis was conducted on aggregated data to prevent the identification of any individual. Consistent with the SDG 4.7 principles, the study included a separate debriefing section to address ethical issues unique to generative AI, clearly presenting potential concerns such as the possibility of biases in ChatGPT responses and its environmental impact. Lastly, the study's limitations, such as self-reporting biases and the geographical location of the sample, were openly acknowledged to promote academic integrity.

4. Measures

Constructs were operationalized using validated scales adapted to the EFL context. SDG integration was measured using a 36-item scale with three subscales: economy (13 items), society (9 items), and environment (14 items). This scale was derived from Atmaca et al.⁶², assessing exposure to sustainability themes in language tasks. Digital literacy was evaluated using a 29-item scale, expanded to six dimensions adapted from Rodríguez-de-Dios et al.⁶³, including Technological Literacy (7 items), Personal Security Literacy (5 items), Critical Literacy (5 items), Device Security Literacy (4 items), Information Literacy (5 items), and Communication Literacy (3 items). Moreover, EFL Students' Creativity was captured through Govindasamy et al.⁶⁴, a 12-item instrument, covering four subscales, each with three items: originality, flexibility, fluency, and elaboration. ChatGPT usage was quantified via an 8-item scale adapted from Abbas et al.⁶⁵, focusing on frequency, purpose, and perceived efficacy in language tasks. Motivation was assessed using Obiosa⁶⁶, with a 05-item scale. Finally, language learning engagement was evaluated using a 9-item scale from Eerdemutu et al.⁶⁷. All scales employed 5-point Likert responses, with pilot testing (n=30) confirming reliability (Cronbach's α > 0.82) and confirmatory factor analysis validating construct distinctiveness. The analysis for the present study was conducted based on first-order reflective-reflective constructs, as second-order reflective-reflective constructs are not directly tested.

5. Data analysis

The data analysis was executed through a sequential two-phase procedure utilizing SPSS version 28 and SmartPLS version 4. The process commenced in SPSS 28, where researchers navigated to Analyze > Descriptive Statistics > Descriptives and Analyze > Correlate > Bivariate to compute preliminary descriptive statistics (means, standard deviations) and Pearson correlation coefficients, respectively, for the sample of n = 420 cases. Subsequently, the primary analysis shifted to SmartPLS 4 to conduct Partial Least Squares Structural Equation Modeling (PLS-SEM). The first step in SmartPLS involved validating the measurement model by running the built-in Calculate algorithm to assess internal consistency, requiring composite reliability values to exceed the threshold of 0.70, and convergent validity, confirmed by an Average Variance Extracted (AVE) greater than 0.50 for all constructs. Discriminant validity was then verified using the software's report functions to apply the Fornell-Larcker criterion and ensure all heterotrait-monotrait (HTMT) ratios were below the conservative benchmark of 0.85. Following this validation, the structural model was evaluated by using the Bootstrapping routine configured with 5,000 subsamples and a two-tailed test at the 0.05 significance level to generate bias-corrected confidence intervals for all path coefficients (β) and indirect effects, thereby testing the hypothesized direct, mediating, and moderating relationships. Finally, the model's predictive relevance was ascertained by examining Stone-Geisser's Q² value, obtained through the Blindfolding procedure, with values above zero indicating adequate predictive power, and the substantive impact of predictors was determined by calculating effect sizes (f²) from the PLS algorithm results.

Descriptive statistics
A descriptive analysis provides important insights into patterns of variables, central locations, and their dispersion, enabling researchers to verify the health, normality, and theoretical congruence of the initial data. In this study, the mean values are used to measure the level of kurtosis, while the parameters of standard deviation are employed to assess the level of skewness. Additionally, the effects of skewness are also considered. This suggests that the 420 participants are suitable for parametric tests and exhibit the proper interactions in SDG-based AI-enhanced EFL learning settings.

Table 1 presents descriptive statistics analysis that provides essential breakdowns of how contributors engage with EFL learning, including the integration of AI functions and Sustainable Development Goals principles. Participants demonstrated tremendous confidence in their digital skills and ChatGPT applications, scoring their abilities at 3.98 (SD = 0.76) and 3.78 (SD = 0.85), respectively. These high scores support the institution's focus on technology-based educational approaches. The participants reported diverse levels of creative self-assessment, with creativity exhibiting the lowest mean rating of 3.21, alongside the highest measurement variability (SD = 0.94). All variables exhibited standard distribution patterns (Field, 2018) because their skewness and kurtosis values remained below ±2, and participants utilized the complete response range from 1 to 5. Students perceive SDG Integration (M = 3.52) and Motivation (M = 3.65) at a moderate level, highlighting a potential to strengthen links between sustainability themes and the curriculum, which can boost student drive.

Figure 2 illustrates the mean performance values (solid line) for six variables, SDG, DL, SC, UCGPT, SM, and LLE, over a specified sequence or condition. The shaded band around the mean trend line represents the ±1 standard deviation (SD) range, indicating variability in the data. Trends in the mean values highlight relative performance differences among the variables, while the SD bands provide insight into the consistency and spread of each measure.

Common Method Variance (CMV) bias
To overcome possible Common Method Variance (CMV) bias in self-reported data collected from the same source, researchers employed procedural and statistical remedies. Procedurally, researchers made questions random to reduce consistency bias and ensured that respondents remained anonymous. The single-factor test statistically performed by Harman indicated that the first factor explained 32.7% of the variance (under the 50% mark), indicating that there was no dominant CMV effect.

Table 2 shows that a Common Method Variance (CMV) assessment was conducted using three robust statistical approaches. Harman's single-factor test revealed that only 32.7% of the variance was explained by the first factor, well below the 50% threshold, indicating no dominant CMV effect. The marker variable technique confirmed that adjusted correlations remained statistically significant (p < 0.05), suggesting CMV did not substantially bias the results. Furthermore, complete collinearity assessment showed all Variance Inflation Factors (VIFs) were below 1.8, significantly under the conservative threshold of 3.3, demonstrating the absence of CMV-induced multicollinearity. These comprehensive tests collectively confirm that common method bias does not pose a significant threat to the validity of findings.

Measurement model results
The measurement model was rigorously evaluated to ensure construct reliability, convergent validity, and discriminant validity, adhering to the standards of PLS-SEM. All reflective constructs demonstrated robust psychometric properties, with outer loadings exceeding the recommended threshold of 0.70, confirming that items reliably captured their respective latent variables. Composite reliability (CR) values ranged from 0.85 to 0.93, surpassing the 0.70 benchmark, while average variance extracted (AVE) scores (0.57-0.68) exceeded the 0.50 criterion, affirming convergent validity. Outer variance inflation factor (VIF) values remained below 3.0, indicating no multicollinearity concerns. These results validate the internal consistency and unidimensionality of the scales, establishing a solid foundation for testing structural relationships in the hypothesized model.

Table 3 presents the measurement model results for the reflective constructs in the study, demonstrating that the constructs generally possess strong internal consistency, as indicated by their high Cronbach's Alpha and Composite Reliability (rho_c) scores, which are predominantly above the 0.7 threshold. However, the low Average Variance Extracted (AVE) values for the second-order constructs Sustainable Development Goals (SDG) at 0.370 and Digital Literacy (DL) at 0.256 require specific justification. These AVE values fall below the conventional benchmark of 0.50 because SDG and DL are broader, multidimensional constructs comprised of distinct first-order factors (Economic, Social, and Environmental sustainability for SDG; and Technological, Critical, Information Literacy, etc., for DL). The variance within each of these specific first-order dimensions is high (as shown by their acceptable individual AVE values), but the variance between these different dimensions is lower. Since AVE is a measure of how much a construct explains the variance of its indicators relative to measurement error, a second-order construct with highly diverse sub-dimensions will naturally have a lower AVE because its indicators (the first-order constructs) are not perfectly correlated and are capturing different facets of the broader concept. Therefore, for such complex, formative-like second-order models, the high Composite Reliability (which is acceptable for both SDG and DL) is a more appropriate indicator of internal consistency than AVE, suggesting the constructs are reliable despite their heterogeneous components. Overall, the measurement model demonstrates strong psychometric properties, with most constructs meeting or exceeding reliability and validity standards.

Figure 3 is a measurement model using PLS SEM. In the measurement model, the factor loadings of all items and variables can be seen. The values in the measurement model indicate that the model of the present study is conceptually sound.

Table 4 presents discriminant validity evidence derived from HTMT testing, which passes robust standards because all calculated ratios (0.58 to 0.80) remain below a conservative cut-off point of 0.85, thereby demonstrating construct distinction and redundancy avoidance. The theoretical connection between Motivation and Engagement (HTMT = 0.80) and Digital Literacy and ChatGPT Usage (HTMT = 0.78) stands out through their highest HTMT values in AI-enhanced EFL contexts due to the natural tendency of motivated students to engage deeply while their digital skills enhance AI tool interaction. All constructs maintain roles in the SDG-AI-EFL framework because the HTMT values stay below 0.85. The learner-focused outcome Creativity remains separate from Digital Literacy and technical competencies because its HTMT values consistently fall between 0.63 and 0.74.

Figure 4 shows that HTMT ratios between construct pairs are well below the 0.85 threshold, confirming strong discriminant validity among the constructs. The spider chart below visually demonstrates these relationships.

Hypotheses testing
The structural model was tested to evaluate the direct and indirect correlations between SDGs integration, the use of generative AI, digital literacy, creativity, motivation, and engagement. Direct tests were evaluated using the path coefficient (β) and significance (p), whereas the indirect impacts were evaluated with bootstrapped mediation analysis (5000 subsamples). Findings proved that SDG integration and the use of ChatGPT were strongly predictors of motivation and digital literacy, respectively, and motivation was the mediating variable in main pathways between sustainability themes and creativity and engagement. All tests were made at p =.05 and bias-corrected confidence intervals (CI) were given of all indirect effects.

Table 5 presents the results of direct hypothesis testing, showing strong statistical support for all five hypotheses (H1-H5). The beta (β) values, which indicate the strength and direction of relationships, range from moderate (H2: β = 0.28) to extreme (H5: β = 0.688), with all coefficients being statistically significant (p < .001 or p < .002). The high T-values (all above 3.75) and narrow confidence intervals (none crossing zero) further confirm the robustness of these findings. Notably, H5 stands out with a considerable effect (f² = 0.473), suggesting it explains a substantial portion of the variance in the dependent variable. Overall, the results demonstrate that all hypothesized relationships are significant, with H3 (β = 0.45) and H5 (β = 0.688) having the most substantial impacts. The consistency in significance (all p-values < .01) and effect sizes reinforces the model's reliability.

Table 6 shows that Mediation effects (H7-H8) were evaluated using a bootstrap approach with 5,000 samples. Researchers computed Variance Accounted For (VAF) to determine mediation strength. Both hypotheses show partial mediation (VAF: 28%-35%), as direct effects remained significant after introducing the mediator (β = 0.31 and 0.29, p < 0.001). Moderation (H6) was tested via interaction terms, with simple slope analysis confirming the adverse buffering effect (β = −0.18, p = 0.012).

Predictive validity of the inner model
Stone-Geisser's Q2Q2 criterion evaluated the structural model's predictive capacity for endogenous constructs outside the estimation process. Assessments of this type provide essential evidence to demonstrate the practical usefulness of proposed relationships between SDG integration and the use of generative AI, alongside digital literacy, creativity, motivation, and EFL engagement. The Q2Q2 value exceeding zero confirms the predictive relevance of the model, while higher Q2Q2 values represent a stronger predictive ability. All endogenous constructs in the analysis yielded positive Q2Q2 values, validating the model's forecasting power for language learning outcomes resulting from SDG-aligned pedagogies and AI-driven interventions.

Table 7 shows that the PLS-SEM model demonstrates excellent goodness-of-fit based on three key metrics. The Standardized Root Mean Square Residual (SRMR) value of 0.056, which falls well below both the 0.08 threshold and the HI95 bootstrap threshold of 0.073, indicates strong model fit. Similarly, the unweighted least squares discrepancy (d_ULS) value of 0.412 remains below its 95% confidence interval (HI95) threshold of 0.489, confirming an acceptable model fit. Additionally, the Normed Fit Index (NFI) of 0.917 exceeds the recommended benchmark of 0.90, providing evidence of a well-specified model. Collectively, these fit indices suggest that the theoretical model adequately represents the observed data patterns, supporting the validity of the structural relationships examined in the study.

Table 8 shows that the model's predictive validity proved reliable in predicting SM (Q2 = 0.28) and DL (Q2 = 0.24), thus demonstrating its effectiveness in forecasting the effects of SDG integration and ChatGPT usage. The model's predictive power targets Creativity (Q2 = 0.20) and LLE (Q2 = 0.17) at a moderate level; however, additional factors, such as instructional design and cultural factors, can enhance prediction accuracy. The model exhibits strong predictive accuracy, characterized by minimal prediction errors, as evidenced by low MAE and RMSE values across constructs (MAE = 0.16-0.21), which supports its application in designing AI-enhanced sustainable EFL teaching methods.

Figure 5 shows that visualization effectively demonstrates the model's predictive capabilities across all constructs, clearly indicating both the Q²-predict values and associated error metrics (MAE and RMSE). The declining pattern from Motivation to Engagement shows the relative predictive strength for each construct while maintaining acceptable levels above the moderate threshold (0.15).

DATA AVAILABILITY:
Data supporting the findings of this study are provided in Supplementary File 1.

Figure 1: Research model of the study. Please click here to view a larger version of this figure.

Figure 2: Visualization of descriptive statistics of key variables. Please click here to view a larger version of this figure.

Figure 3: Measurement model. Please click here to view a larger version of this figure.

Figure 4: HTMT ratios for construct pairs. Please click here to view a larger version of this figure.

Figure 5: Predictive validity assessment of model constructs using PLS-predict. Please click here to view a larger version of this figure.

Table 1: Descriptive statistics of key variables. Please click here to download this Table.

Table 2: Common Method Variance (CMV) assessment. Please click here to download this Table.

Table 3: Results of measurement model. Please click here to download this Table.

Table 4: Discriminant validity assessment using the HTMT ratio. Please click here to download this Table.

Table 5: Direct impact (hypothesis testing). Please click here to download this Table.

Table 6: Mediation and moderation analysis. Please click here to download this Table.

Table 7: PLS-SEM model fit assessment. Please click here to download this Table.

Table 8: Predictive validity of the inner model using PLS-predict. Please click here to download this Table.

Summary of key findings
This study examined the interplay between Sustainable Development Goals (SDGs), generative AI (ChatGPT), and key educational outcomes: motivation, digital literacy, and creativity in EFL learning environments. The results confirmed all direct-effect hypotheses (H1-H5), demonstrating that SDG integration (β = 0.37) and ChatGPT use (β = 0.28) significantly enhance student motivation, which in turn fosters digital literacy (β = 0.45) and creativity (β = 0.32). Language learning engagement strongly affected creativity (β = 0.688), suggesting its pivotal role in amplifying students' innovative capacities. The moderating and mediating analyses also revealed that engagement moderates the digital literacy-creativity link (H6, β = -0.18). At the same time, motivation mediates the effects of ChatGPT (H7, β = 0.13) and SDGs (H8, β = 0.17) on digital literacy. These findings collectively highlight the synergistic potential of combining SDG-aligned pedagogy and AI tools to cultivate 21st-century skills in EFL education.'

The nuanced finding of a negative moderation effect (H6) suggests that at very high levels of engagement, the direct relationship between digital literacy and creativity is slightly attenuated. This may indicate that highly engaged students are operating at a ceiling of their creative capacity within a given task, where further digital skills do not linearly translate to greater innovation. Furthermore, the stronger effect of SDGs compared to ChatGPT on creativity (H7, H8) can be attributed to the inherently meaningful and problem-based nature of SDG topics, which likely stimulate deeper cognitive and empathetic processes essential for original thought, whereas ChatGPT may function more as a facilitative tool. This aligns with prior literature by⁶⁸, who found that contextually rich, authentic content is a more powerful driver of intrinsic motivation and subsequent creative output than technological access alone. The significant mediation of motivation for both independent variables confirms its central role, as established by Self-Determination Theory, in channeling external resources into tangible skills development.

Critical protocol steps for successful implementation
The efficacy of the proposed SDG-AI framework is contingent upon several critical protocol steps, as illuminated by the empirical findings. First, implementation requires the deliberate design of structured, problem-based learning tasks anchored in specific SDGs (such as simulating UN climate negotiations or designing inclusive community projects), as our results confirm the stronger path coefficient from SDG integration (β = 0.37) directly fosters the intrinsic motivation necessary for creative output. Second, the role of AI must be strategically scaffolded; educators should provide explicit prompt engineering guidelines that position ChatGPT as a collaborative partner for idea generation and critical evaluation, rather than a mere information source, to activate the significant mediating role of motivation (H7, β = 0.13). Finally, the negative moderating effect of engagement (H6, β = -0.18) underscores a crucial facilitative step: instructors must continuously monitor learner engagement to dynamically adjust task complexity, ensuring that highly engaged students are provided with sufficiently challenging, open-ended objectives to prevent the observed attenuation of the digital literacy-creativity link, thereby maximizing the framework's synergistic potential.

The use of the Partial Least Squares Structural Equation Modeling in this study, which is referred to as PLS-SEM, has very specific features of benefits and disadvantages compared to the more traditional covariance-based version of Structural Equation Modeling (CB-SEM), commonly applied in educational studies. The main benefit of the PLS-SEM is that it is designed for use in predictive and exploratory research, and can effectively estimate more complex models using smaller sample sizes. This is particularly appropriate for our study, which aims to create a new theoretical framework rather than validate an existing one. Moreover, in contrast to CB-SEM, which assumes numerous distributional conditions that may be frequently breached when dealing with behavioral data, PLS-SEM does not impose any distributional requirements, which are far less restrictive in the context of real-life data in this research. This, however, has a known trade-off: PLS-SEM lacks global model fit indices (e.g., CFI, RMSEA), a characteristic of CB-SEM that is used to determine the overall goodness-of-fit between the model and the data. As a result, although our design is more focused on predictive strength and theoretical investigation, subsequent studies may attempt to test the developed framework on a bigger sample by using CB-SEM to evaluate its confirmatory validity.

Theoretical and empirical alignment
The positive relationship between SDG integration and motivation (H1) aligns with self-determination theory, which posits that contextual relevance and real-world applicability enhance intrinsic motivation. Prior studies^69,70, similarly found that thematic curricula, such as SDGs, increase engagement by connecting learning to global challenges. The significant impact of ChatGPT on motivation (H2) corroborates recent research⁷¹, emphasizing AI's role in personalizing learning experiences and fostering autonomy. However, the effect size for ChatGPT (β = 0.28) was smaller than that of SDGs (β = 0.37), possibly due to varying student familiarity with AI tools or the need for structured scaffolding to maximize their motivational benefits. The intense mediation of motivation in digital literacy development (H3, H7, H8) resonates with⁷² the assertion that motivated learners proactively acquire technological competencies. This finding extends self-determination theory by illustrating how extrinsic stimuli (AI tools, SDGs) translate into skill development through intrinsic drive. Meanwhile, the digital literacy-creativity link (H4) supports technical proficiency, which enables innovative problem-solving^73,74. However, the moderation by engagement (H6) suggests that digital skills alone are insufficient; students must be actively immersed in learning to harness their full creative potential. The negative moderation (β = -0.18) implies that high engagement may reduce the dependency on digital literacy for creativity, possibly because engaged learners leverage alternative cognitive or collaborative strategies.

From a methodological standpoint, the application of Partial Least Squares Structural Equation Modeling (PLS-SEM) was critical for this exploratory research. Key steps ensuring the validity of our results included a thorough examination of the measurement model, confirming indicator loadings > 0.7, composite reliability, and discriminant validity (HTMT < 0.9) prior to assessing the structural paths. Common issues, such as low outer loadings or cross-loadings, were mitigated by removing problematic indicators, and a sufficient sample size was secured to ensure statistical power for bootstrapping. While PLS-SEM is advantageous for predictive applications and models with formative constructs, unlike covariance-based SEM (CB-SEM), it does not provide global model fit indices, which is a key limitation. However, its robustness to non-normal data and ability to handle complex models with smaller samples justified its selection for this study, allowing us to effectively test the proposed mediating and moderating relationships within our theoretical framework.

Consistencies and discrepancies
While most findings align with existing literature, the unexpectedly high effect of engagement on creativity (H5, β = 0.688) diverges from prior studies that emphasize incremental skill-based pathways⁷⁵. This discrepancy may stem from the unique EFL context, where language anxiety often hinders engagement⁷⁶, which acts as a critical gateway for creative expression when mitigated. Additionally, the small effect sizes for mediation (H7, H8) and moderation (H6) suggest that while these mechanisms are statistically significant, their practical impact may be secondary to direct motivational and pedagogical interventions. Cultural factors, such as the Chinese educational emphasis on collective learning, could also explain why individual engagement moderates rather than amplifies the digital literacy-creativity relationship⁷⁷.

Theoretical and contextual explanations
These findings should be interpreted with their methodological constraints in mind. While the results align with Self-Determination Theory, where SDGs provide relevance and ChatGPT fosters autonomy, the cross-sectional, self-report design from a single national context necessitates caution in claiming causality. The smaller effect size for ChatGPT, for instance, could reflect implementation challenges but may also be influenced by the study's inability to control for varying levels of AI familiarity among participants. Furthermore, the unique pattern of results, such as the negative moderation effect and the strong engagement-creativity link, may be shaped by the specific educational culture of the sample. To solidify these causal inferences and enhance generalizability, future research should employ longitudinal designs or controlled experiments that track the evolution of motivation and skills over time. Such rigorous designs, as called for in recent literature on educational AI, would help isolate the specific impacts of SDG and AI integrations while controlling for extraneous variables across diverse cultural settings.

This research provides vital theoretical insights into the joint operation of motivation, digital literacy skills, and creativity among English Language Foreign Students. Applying the SDGs and ChatGPT with students yields empirical evidence to validate and improve two core psychological theories, namely self-determination theory and the technology acceptance model. Self-determination theory demonstrates the motivating relationship between external curricular features and technological elements, as these aspects increase students' internal drive, thereby enhancing skill mastery and creative performance. The research investigation yielded vital discoveries concerning sustainable development goals (SDGs), alongside the implementation of the ChatGPT system and the mechanisms that motivate factors to enhance technological competencies and creative abilities in the English as a foreign language education context. Using the SDGs in conjunction with the EFL curriculum and the ChatGPT system boosts student motivation and enhances digital competencies, incorporating creative abilities. The fundamental elements of language acquisition, combined with the motivational drive of students, contribute to the development of complex digital competency methods. Previous academic investigations have gained additional scholarly understanding of digital language learning, but provide critical advice about the educational use of modern instructional sources and curricula for teaching professionals

Conclusion
This study advances the field by providing a validated model that synergistically bridges macro-level global citizenship education (SDG 4.7) with micro-level technological personalization through the use of AI. It advances SDG pedagogy beyond a mere thematic focus, empirically demonstrating its potent role as a catalyst for student motivation and 21st-century skill development. Concurrently, it refines the concept of AI-assisted learning by illustrating that tools like ChatGPT are most effective not in isolation, but when integrated within a meaningful, real-world context that gives their use purpose and direction. The critical finding that motivation mediates the effect of both independent variables confirms that the "why" of learning is as important as the "how," offering a crucial lens for future educational design.

For practitioners, these findings offer a clear mandate: to effectively harness the potential of both SDGs and AI, curriculum design must be intentionally student-centered. This involves selecting SDG topics that resonate with learners' identities and employing generative AI not just as a language tool, but as a partner for critical inquiry, such as analyzing sustainability issues or simulating cross-cultural dialogues. Success, therefore, depends on moving beyond technical adoption to fostering a supportive learning ecology where motivation is nurtured, digital access is equitable, and pedagogical strategies are explicitly designed to translate engagement into creativity and critical digital literacy. This integrated approach provides a replicable framework for aligning technological transformation with the foundational goals of quality, inclusivity, and relevance in education for all.

Limitations and future research
This study has several limitations that should be considered. First, the use of a convenience sample from a single university in Eastern China significantly restricts the generalizability of the findings. The results are inherently shaped by this specific socio-educational context, characterized by its distinct cultural dynamics, including specific attitudes towards technology in education and prevailing pedagogical traditions. Consequently, the observed relationships between SDGs, AI, motivation, and learning outcomes may not be directly transferable to other cultural or educational settings with different norms and resources. Second, the reliance on self-reported data for measuring constructs such as motivation and digital literacy introduces the potential for bias, including social desirability or inaccurate self-assessment. Finally, the cross-sectional nature of the design precludes the establishment of causal relationships or insights into the long-term effects of the variables studied.

Future research should actively address these limitations. To enhance external validity and better understand cultural nuances, studies should employ longitudinal or mixed-methods approaches across diverse geographical regions. A longitudinal design would trace the developmental trajectory of motivation and skill acquisition, establishing causality. In contrast, a mixed-methods approach would be invaluable for validation, triangulating self-reported data with behavioral observations, performance metrics, and qualitative interviews to provide a more robust and contextually rich understanding. Further investigation is also needed to explore how the interplay of SDGs and AI functions in different cultural contexts and with other emerging technologies. Beyond academic research, the methodological framework presented here offers direct practical applications for educational practice and policy. Instructional designers can deploy this model to conduct needs assessments and evaluate the efficacy of specific SDG-AI integrated modules within their own institutional contexts, using the validated survey instrument to diagnose strengths and gaps in student motivation and digital literacy. Furthermore, the approach provides an actionable blueprint for language program administrators seeking to make data-informed decisions on scaling generative AI tools, enabling them to benchmark engagement and creativity outcomes against our findings to justify resource allocation and tailor professional development for educators on effective, SDG-aligned prompt engineering.