Self-efficacy in AI-assisted L2 academic writing: scale development and validation.

Self-efficacy in AI-assisted L2 academic writing: scale development and validation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Psychol . 2026 Apr 22;14:814. doi: 10.1186/s40359-026-04503-8 Self-efficacy in AI-assisted L2 academic writing: scale development and validation Guangyuan Yao 1 , Lingxi Fan 2, ✉ Author information Article notes Copyright and License information 1 Central South University, Changsha, China 2 Hangzhou International Innovation Institute of Beihang University, Hangzhou, China ✉ Corresponding author. Received 2025 Jul 7; Accepted 2026 Apr 1; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13231772  PMID: 42015312 Abstract Background The rise of generative artificial intelligence (AI) has fundamentally reshaped the landscape of second language (L2) academic writing, introducing novel demands on learners’ competencies and psychological resources. While self-efficacy is a well-established predictor of writing success, existing measurement scales predate the AI era and fail to capture learners’ confidence in navigating the unique human-AI collaborative process. This study addresses this gap by reporting on the development and validation of the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS). Methods Following rigorous scale development procedures, data were collected from 680 Chinese university EFL students. The sample was randomly split for an Exploratory Factor Analysis (EFA; n = 340) and a Confirmatory Factor Analysis (CFA; n = 340). Results The EFA revealed a clear three-factor structure, which was subsequently confirmed by the CFA, demonstrating good model fit. The three emergent factors were: (1) AI-Assisted Process Regulation Efficacy, (2) AI-Assisted Language Efficacy, and (3) Ethical and Integrity Efficacy. The final 11-item scale demonstrated high internal consistency (α = .94) and robust evidence of convergent, discriminant, and concurrent validity through correlations with established measures of L2 writing self-efficacy and writing anxiety. Conclusions The SE-AI-L2AWS provides a potentially reliable and valid instrument for researchers and educators to assess L2 learners’ confidence in using AI for academic writing. This tool could facilitate diagnostic assessment, guide pedagogical interventions, and advance research on the motivational dynamics of writing in the age of AI. Supplementary Information The online version contains supplementary material available at 10.1186/s40359-026-04503-8. Keywords: Self-efficacy, AI-assisted writing, Second language (L2) writing, Scale development, Educational technology Introduction The capacity to produce effective academic writing is a cornerstone of scholarly success, yet it remains a multifaceted and demanding endeavor, particularly for second language (L2) learners [ 24 , 28 ]. The composing process requires not merely the command of linguistic and rhetorical conventions but also the orchestration of a host of cognitive and metacognitive skills [ 18 , 23 , 56 ]. In recent years, the landscape of academic writing has undergone a profound transformation with the widespread availability of generative artificial intelligence (AI) tools. These technologies offer unprecedented support for brainstorming, drafting, and language refinement (e.g., for ideation and organization) but concurrently introduce a new layer of complexity, demanding novel user competencies in prompt engineering, critical evaluation of AI-generated content, and the ethical integration of sourced material. This technological evolution reshapes the very nature of the writing task, and consequently, necessitates a re-examination of the psychological factors that underpin student success. Among the crucial motivational factors influencing writing, self-efficacy—an individual’s belief in their capability to succeed in specific tasks—has been consistently identified as a salient predictor of academic performance [ 34 , 47 , 49 , 52 ]. A robust corpus of research has shown that positive self-efficacy beliefs are associated with greater effort, persistence, and ultimately, higher quality writing outcomes [ 43 , 48 , 60 ]. However, as Bandura [ 3 ], the architect of self-efficacy theory, has emphatically argued, the predictive power of this construct is contingent on its task specificity. To be valid and useful, measures of self-efficacy must be meticulously tailored to the particular domain of functioning and the specific demands of the task in question. While previous research has made significant strides in developing context-specific writing self-efficacy scales for L2 learners (e.g., [ 49 ]) and even genre-specific contexts [ 55 ], the rapid integration of AI into the writing process has created a new, distinct domain of functioning that remains unmeasured. Existing scales, designed before the generative AI revolution, do not and cannot capture learners’ confidence in their ability to navigate the unique challenges and opportunities presented by AI-assisted writing. Consequently, a significant gap exists in the literature: the absence of a psychometrically sound instrument designed specifically to measure L2 learners’ self-efficacy in this new paradigm. Without such a tool, educators and researchers lack the means to diagnose students’ needs, design targeted pedagogical interventions, and assess the impact of AI on L2 writing development from a motivational perspective. To address this critical gap, the present study reports on the development and validation of the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS), making prominent theoretical, methodological, and practical contributions to the existing literature. Theoretically, this study reconceptualizes traditional writing self-efficacy within the novel human-AI collaborative paradigm. Unlike existing frameworks (e.g., [ 49 ]) that focus primarily on learners’ internal cognitive and linguistic resources, our conceptualization highlights the psychological shift toward orchestrating external AI capabilities. Crucially, it introduces Ethical and Integrity Efficacy as a distinct psychological dimension, reflecting the new volitional challenges L2 writers face in maintaining academic honesty in the post-plagiarism era. Methodologically, we provide a rigorously validated and highly parsimonious 11-item multidimensional instrument specifically tailored for the generative AI context, capable of accurately capturing nuanced student beliefs that pre-AI scales inherently fail to measure. Practically, the SE-AI-L2AWS serves as a valuable diagnostic tool for EAP educators, enabling them to pinpoint specific areas of student apprehension—whether in process regulation, language enhancement, or ethical boundary management—thereby facilitating the design of targeted pedagogical interventions. By following rigorous psychometric procedures modeled on seminal scale development studies in the field, this research is guided by the following research questions: RQ1: What is the underlying factor structure of the newly developed SE-AI-L2AWS? RQ2: Does the hypothesized factor structure of the scale demonstrate a good model fit when tested with a new sample of L2 writers? RQ3: Does the SE-AI-L2AWS exhibit robust psychometric properties, including internal consistency, convergent validity and discriminant validity? Literature review The foundational role of self-efficacy in academic writing Self-efficacy, a central construct within social cognitive theory [ 1 , 2 ], refers to an individual’s prospective judgments of their own capabilities to organize and execute the courses of action required to attain designated performances. As a key mechanism of human agency, these beliefs are not passive reflections of past performance but are dynamic, generative judgments that mediate an individual’s choices, effort expenditure, persistence in the face of adversity, and emotional reactions [ 34 , 35 , 43 , 57 ]. Within the domain of academic writing—a complex endeavor demanding the orchestration of cognitive, linguistic, and metacognitive skills [ 23 , 56 ]—the role of self-efficacy is particularly salient. It often determines whether a student approaches a writing task with engagement and resilience [ 38 ] or with apprehension and avoidance [ 36 , 52 ]. The predictive validity of writing self-efficacy has been firmly established across a wide body of empirical literature. In first language (L1) contexts, a strong, positive correlation between students’ self-efficacy and their writing achievement is a consistent finding [ 6 , 37 , 42 , 44 ]. This relationship, however, appears to be even more critical for second language (L2) writers. L2 learners must navigate not only the universal challenges of composition but also the added cognitive and affective loads associated with L2 linguistic proficiency, unfamiliar rhetorical conventions, and heightened writing anxiety [ 28 , 46 ]. In this high-stakes environment, self-efficacy acts as a critical affective buffer and motivational driver. A number of studies have confirmed the positive link between self-efficacy and writing outcomes for L2 learners in various contexts (e.g., Golparvar and Khafi, [ 17 ], Zabihi, [ 54 ]. The definitive meta-analysis by Sun et al., [ 48 ] provided robust evidence for this, revealing a significantly larger effect size for the relationship between self-efficacy and writing performance among L2 learners ( r = .441) compared to their L1 counterparts ( r = .233). This finding powerfully underscores the imperative of developing precise and contextually-aware measurement instruments for the L2 population, as self-efficacy beliefs appear to be a pivotal factor in their academic success. Bandura, [ 3 ] provided a clear directive for the measurement of self-efficacy: scales must be tailored to the specific domain and demands of the task. A general sense of confidence is a poor predictor; efficacy must be assessed in relation to the specific capabilities required to succeed. The historical trajectory of writing self-efficacy instrumentation reflects a progressive scholarly commitment to this principle. Initial instruments, while pioneering, often measured a broad, decontextualized range of writing-related skills (e.g., [ 45 ]). While foundational, this approach was subsequently critiqued for its limited theoretical grounding and its inability to yield fine-grained diagnostic information [ 6 ]. In a significant advancement, Bruning et al., [ 6 ], drawing on cognitive models of writing (e.g., Hayes, [ 22 ] and self-regulation theory [ 59 ], developed a multidimensional scale for L1 writing, delineating three core dimensions: ideation, writing conventions, and self-regulation. This marked a significant step forward by aligning measurement with the cognitive and metacognitive components of writing. Building upon this work, Teng et al., [ 49 ] addressed the critical need for an instrument tailored to the unique exigencies of the L2 context. Their research acknowledged that L2 writers face distinct linguistic hurdles and performance pressures. Consequently, their proposed Second Language Writer Self-Efficacy Scale (L2WSS), grounded in both social cognitive theory and self-regulated learning (SRL) theory [ 39 , 58 ], delineated a three-factor structure: Linguistic Self-Efficacy, Self-Regulatory Efficacy, and Performance Self-Efficacy. The psychometric soundness of the L2WSS was rigorously established through structural equation modeling, and its utility has been subsequently confirmed in research examining complex tasks such as integrated writing [ 17 ] and its relationship with learning strategies [ 47 ]. Most recently, Zhang et al., [ 55 ] further advanced this pursuit of specificity by developing a genre-based scale, arguing that confidence can vary significantly depending on whether a student is writing a narrative or an argument. This trajectory of scholarly refinement—moving toward ever-greater contextual specificity—provides the theoretical and methodological precedent for the present study. Self-efficacy in academic writing and AI paradigm Despite these significant advancements, the existing literature on measurement instruments is rendered insufficient by the paradigm shift instigated by generative artificial intelligence. The integration of AI tools like ChatGPT into the writing workflow is not merely an incremental technological addition; it fundamentally reconfigures the writing task itself. As noted by Warschauer [ 50 , 51 ], digital technologies transform literacy from a solitary cognitive activity into a distributed, socio-technical practice. In this new “Human-AI collaboration” model [ 12 , 31 ], the writer’s role evolves from that of a sole author to an orchestrator of a complex interaction, necessitating a suite of new literacies that were previously non-existent [ 20 , 29 ]. This new reality exposes a critical disjuncture between the constructs measured by existing scales and the capabilities now required for effective writing. Writing in the AI era demands what Haas [ 19 ] describes as a deep engagement with the materiality of literacy, where writers must navigate the affordances and constraints of the technology itself. For instance, the prompt engineering required to guide an AI [ 26 ] represents a novel form of regulatory skill that differs significantly from internal metacognition. Similarly, the need to critically evaluate AI-generated outputs for hallucinations or bias calls for critical digital literacies [ 11 ] that go beyond traditional linguistic knowledge. Furthermore, as human-AI interaction patterns become more iterative and intertwined [ 32 ], the ethical boundaries of authorship become increasingly complex, requiring a specific form of volitional efficacy to maintain academic integrity. The current benchmark, L2WSS [ 49 ], despite its value, fails to capture the new reality. The Linguistic Self-Efficacy dimension assesses confidence in a writer’s internal linguistic reservoir (e.g., recalling vocabulary, constructing sentences). It fails to capture a writer’s confidence in their ability to externally manage AI-generated language, which includes evaluating the appropriateness of an AI’s lexical choices, judging the coherence of a synthesized paragraph, or reformulating a machine-generated sentence to align with one’s own authorial voice. The Self-Regulatory Efficacy dimension focuses on the metacognitive management of one’s own internal processes (e.g., goal-setting, planning, monitoring). It does not measure the distinct metacognitive capabilities required for prompt engineering—the strategic planning and iteration of instructions to guide the AI—or the crucial evaluative process of critically monitoring AI outputs for factual accuracy, logical fallacies, and potential biases. The Performance Self-Efficacy dimension relates to confidence in meeting overall course requirements. It does not address the specific, nuanced, and ethically-laden judgments required to navigate the use of AI within the bounds of academic integrity, a primary source of anxiety and uncertainty for contemporary students. Therefore, applying extant scales to this new context would produce an incomplete and potentially distorted picture of student self-beliefs. The core of a student’s confidence or apprehension in the modern writing process—their perceived ability to effectively and ethically engage with AI—would remain entirely unmeasured. This disjuncture between existing measures and the contemporary writing context creates a critical measurement void that impedes both theoretical understanding and pedagogical progress. To address the measurement void, this study seeks to conceptualize and operationalize a new construct: Self-Efficacy in AI-Assisted L2 Academic Writing (SE-AI-L2AWS). Grounded in Bandura’s [ 3 ] principle of task-specific measurement, our conceptualization is not merely a theoretical extension but is empirically informed by preliminary qualitative data (see Method section for details). This approach is necessary because the integration of AI reconfigures the writing process itself, transforming it from a solitary cognitive activity into a collaborative, socio-technical practice that demands a new understanding of the writer’s role and environment (cf. Graham, [ 18 ]. Analysis of semi-structured interviews with L2 learners revealed that their confidence, or lack thereof, in using generative AI for academic writing consistently coalesced around three distinct yet interrelated domains of capability. First, learners expressed significant concerns about their ability to manage the overall human-AI interaction process. This went beyond simple planning, encompassing the skills needed to formulate effective prompts, critically assess the relevance and accuracy of AI outputs, and structure a coherent argument using AI as a tool. This dimension directly extends traditional models of writing cognition and self-regulation into a new, interactive context, forcing a remodeling of how writers plan, execute, and monitor their work [ 23 , 60 ]. We term this dimension AI-Assisted Process Regulation Efficacy. Second, students focused on their capacity to leverage AI for specific linguistic and rhetorical enhancements. This included confidence in using AI for vocabulary enrichment, sentence reformulation, and error correction, while crucially maintaining their own authorial voice. This dimension acknowledges the unique linguistic hurdles faced by L2 writers [ 46 ] and reframes the Linguistic Self-Efficacy proposed by previous L2 writing scales into a new capability: managing an external linguistic resource rather than relying solely on internal knowledge [ 49 ]. This dimension is conceptualized as AI-Assisted Language Efficacy. Third, a pervasive theme was the anxiety and uncertainty surrounding the ethical use of these tools [ 53 ]. Students reported needing confidence in their ability to use AI responsibly, maintain academic integrity, and clearly distinguish their own intellectual contributions from AI-generated content. This new source of writing apprehension is a critical affective factor that can profoundly influence a student’s engagement and performance, echoing long-established links between confidence, apprehension, and writing outcomes [ 36 ]. This led to the formulation of the Ethical and Integrity Efficacy dimension. Based on this integrated theoretical and empirical foundation, we formally define SE-AI-L2AWS as: An L2 writer’s judgment of their capabilities to effectively, critically, and ethically employ generative AI tools to support and enhance the processes of academic writing, from initial conception to final revision. This tripartite conceptualization—encompassing process regulation, language enhancement, and ethical conduct—aims to provide a comprehensive framework for measuring the specific self-beliefs required to navigate the contemporary L2 writing landscape. The development and validation of a scale to measure this construct is a requisite step for enabling rigorous empirical research into the motivational dynamics of L2 writing in the AI era. Method This study employed a multi-phase, quantitative research design to develop and validate the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS). The process involved three main phases: (1) item generation and content validation through expert review and student interviews; (2) a large-scale survey administration for psychometric analysis; and (3) an assessment of the scale’s relationship with other relevant constructs to establish convergent and discriminant validity (Fig. 1 ). The procedures were modeled on established best practices in scale development within applied linguistics [ 13 , 49 ]. Fig. 1. Open in a new tab Overall research process flowchart Participants A total of 680 undergraduate EFL students were recruited from one comprehensive university in China via convenience sampling. All participants were non-English majors who were required to take compulsory English for Academic Purposes (EAP) courses. Their reported length of formal English study was a mean of 10.2 years (SD = 2.4), and they represented a range of disciplines, including business, engineering, humanities, and computer science. Importantly, all participants had prior experience with AI-assisted writing tools. To gauge their familiarity, participants responded to a frequency item (“How often do you use AI tools for academic writing tasks?”). Results indicated a moderate level of engagement: 12.5% reported using AI ‘Daily’, 45.2% ‘Weekly’, 31.8% ‘Monthly’, and only 10.5% ‘Rarely/Never’. This confirms that the sample predominantly consisted of active users familiar with the technology. Participation was voluntary, and informed consent was obtained from all students prior to data collection. The sample was randomly split into two equivalent subsamples to allow for cross-validation [ 10 ]. Sample A ( n = 340) was used for the initial Exploratory Factor Analysis (EFA). Sample B ( n = 340) was used for the subsequent Confirmatory Factor Analysis (CFA) and for examining criterion-related validity. The demographic profiles of the two subsamples were checked to ensure there were no significant differences in terms of gender distribution, year of study, or academic major. Instruments The development of the SE-AI-L2AWS followed a rigorous, multi-stage process grounded in the guidelines proposed by Bandura [ 3 ]. Phase 1: Item generation The development process began with a two-pronged approach to ensure both theoretical grounding and ecological validity. First, a comprehensive literature review was conducted on writing self-efficacy (e.g., [ 6 , 49 ]) and the emerging research on AI in education. Second, and more critically, semi-structured interviews were conducted with a separate group of 15 L2 university students. The interview transcripts were analyzed using thematic analysis [ 5 ]. The coding process revealed three distinct clusters of self-efficacy beliefs. First, students frequently mentioned the difficulty of “controlling” the AI, such as writing effective prompts or judging the relevance of outputs (e.g., “I often don’t know how to ask ChatGPT to get the answer I want”), which informed the Process Regulation dimension. Second, students described using AI specifically for linguistic polish, such as vocabulary replacement and grammar checking, leading to the Language Efficacy dimension. Third, a recurring theme was the anxiety surrounding plagiarism and the need to confidently distinguish their own work from AI’s, forming the basis for the Ethical and Integrity dimension. Table 1 provides a detailed mapping of representative student quotes to the finalized scale items, illustrating the empirical grounding of the constructs. Table 1. Mapping of qualitative themes to SE-AI-L2AWS items Emergent Theme Representative Student Quote (Translated) Corresponding Scale Item (Final Version) Process Regulation “I spend a lot of time figuring out how to give instructions to the AI. If the instruction is vague, the answer is useless. I need to know how to guide it.” Item 1: I am confident that I can write clear and specific prompts to guide an AI tool… “Sometimes AI talks nonsense. I have to read it carefully and decide if it’s true or if it fits my essay’s logic.” Item 4: I am confident that I can critically evaluate the accuracy and reliability of information… Language Efficacy “My vocabulary is limited. I use AI to find better, more academic words to replace my simple ones.” Item 6: I am confident that I can use an AI tool to find precise academic vocabulary… “I write the rough draft, but I use AI to fix my grammar mistakes. I need to be sure it’s correcting it right.” Item 7: I am confident that I can effectively use an AI tool to identify and correct grammatical errors… Ethical & Integrity “I’m always worried the teacher will think I cheated. I need to be sure I’m using it just for help, not copying.” Item 11: I am confident that I can use AI tools… in a way that fully adheres to the principles of academic integrity. “It’s hard to tell which part is mine and which is the AI’s after many revisions. I need to keep that clear.” Item 10: I am confident that I can clearly distinguish between my own original ideas and contributions from an AI tool… Open in a new tab Based on this three-factor framework and the specific examples from the interviews, an initial pool of 22 items was generated. Items were carefully crafted to represent each of the three proposed dimensions (e.g., ‘I can write clear and specific prompts…’ for Process Regulation; ‘I can use an AI tool to find precise academic vocabulary…’ for Language; ‘I can use AI tools while adhering to academic integrity…’ for Ethics). All items were phrased as ‘I can…’ statements, following Bandura’s [ 3 ] guidelines. Phase 2: Content and face validity The initial pool of 22 items was then submitted for content validation. A panel of five experts (three in L2 writing, two in educational technology) independently evaluated each item for its clarity, relevance to its designated dimension, and overall representativeness. The experts were asked to rate each item on a 4-point Likert scale (1 = not relevant, 4 = highly relevant) based on three criteria: relevance to the construct, clarity of wording, and conciseness. Based on their ratings, the Content Validity Index (CVI) was calculated [ 41 ]. Items with an Item-CVI (I-CVI) lower than 0.80 were removed or revised. The results showed excellent content validity for the final 11 items, with I-CVI values ranging from 0.80 to 1.00 and a Scale-CVI (S-CVI/Ave) of 0.96, well above the recommended threshold of 0.90. Following this quantitative review, the refined items were piloted with a small group of 25 L2 students to confirm face validity and comprehensibility.The experts provided ratings and qualitative feedback. Based on their consensus, items that were deemed ambiguous, redundant, or poorly aligned with the construct definition were either revised or removed. This rigorous review process led to the refinement and reduction of the item pool, resulting in a set of 11 items (see supplementary file) deemed ready for large-scale psychometric testing. This 11-item version was then piloted with a small group of 25 L2 students to check for face validity, ensuring the language was clear and comprehensible to the target population . Phase 3: Scale format The final instrument employed a 7-point Likert-type scale, ranging from 1 (Not at all confident) to 7 (Very confident). This format was chosen to provide sufficient variance while avoiding the potential confusion reported with 0-100 scales among EFL learners [ 7 , 55 ]. Following standard practice for research in EFL contexts, the English items were translated into Mandarin Chinese. The linguistic equivalence of the two versions was verified through a rigorous translate-back-translate procedure involving two independent bilingual academics [ 13 ]. To establish the validity of the new scale, two additional established questionnaires were administered to Sample B. To examine discriminant validity, we administered the L2 Writing Self-Efficacy Scale (L2WSS) developed by Teng et al., [ 49 ]. This scale consists of 20 items measuring three dimensions: linguistic, self-regulatory, and performance self-efficacy. Since the L2WSS measures general L2 writing self-efficacy, a moderate positive correlation with the new SE-AI-L2AWS was expected. However, a very high correlation (e.g., > 0.80) would suggest that the two scales are measuring the same construct. Therefore, this measure allowed us to test whether the new scale captures a conceptually distinct dimension of self-efficacy. The L2WSS has demonstrated robust reliability and validity in previous studies [ 17 , 49 ]. To establish concurrent validity, Foreign Language Writing Anxiety Scale [ 8 ] was adopted. According to self-efficacy theory and empirical findings [ 36 , 52 ], self-efficacy and anxiety are strongly and negatively correlated. A significant negative correlation between scores on the SE-AI-L2AWS and a measure of writing anxiety would provide strong concurrent evidence for the validity of the new construct. Data collection and analysis Participants were recruited during their EAP classes. Data collection was conducted online using the Qualtrics survey platform. After providing informed consent, participants first completed a demographic questionnaire. They were then presented with the SE-AI-L2AWS, the L2WSS, and the Foreign Language Writing Anxiety Scale in a counterbalanced order to minimize potential order effects. The entire procedure took approximately 20–25 min to complete. Participants were assured of the confidentiality and anonymity of their responses. All quantitative data were analyzed using IBM SPSS Statistics 28 and Mplus 8.8. The data analysis proceeded in several stages, directly corresponding to the research questions. The dataset was first screened for missing values, outliers, and normality. Skewness and kurtosis values were examined to ensure they fell within acceptable ranges (e.g., |skewness| < 2, |kurtosis| < 7) for subsequent multivariate analyses [ 27 ]. To address the first research question, an exploratory factor analysis (EFA) was conducted on data from Sample A to investigate the underlying factor structure of the SE-AI-L2AWS items. The suitability of the data for factor analysis was confirmed using the Kaiser-Meyer-Olkin (KMO) measure and Bartlett’s test of sphericity. Given the theoretical assumption that factors related to self-efficacy would be correlated, Maximum Likelihood (ML) estimation with Promax rotation was employed [ 15 ]. The number of factors to retain was determined using multiple criteria, including the scree plot, eigenvalues greater than 1.0, and parallel analysis [ 33 ]. Items with low factor loadings (< 0.50) or significant cross-loadings were considered for removal. To answer the second research question, a confirmatory factor analysis (CFA) was performed using data from Sample B to test the model structure identified in the EFA. Model fit was evaluated using established indices and their recommended thresholds, including the chi-square to degrees of freedom ratio (χ²/df < 3), the Comparative Fit Index (CFI ≥ 0.95), the Tucker-Lewis Index (TLI ≥ 0.95), the Root Mean Square Error of Approximation (RMSEA ≤ 0.06), and the Standardized Root Mean Square Residual (SRMR ≤ 0.08) [ 25 , 27 ]. Given the ordinal nature of the 7-point Likert scale data, Confirmatory Factor Analysis (CFA) was conducted using the Weighted Least Squares Mean and Variance adjusted (WLSMV) estimator, which is based on the polychoric correlation matrix and is robust to non-normality [ 4 ]. To rigorously test the construct structure, alternative models (one-factor and second-order models) were compared against the hypothesized three-factor correlated model. Additionally, multigroup CFA (MGCFA) was conducted to examine the measurement invariance of the scale across gender groups and majors to provide preliminary evidence of the instrument’s robustness beyond the initial factor structure validation. In relation to the third research question, multiple analyses were conducted to assess the reliability and validity of the SE-AI-L2AWS. Internal consistency was evaluated using Cronbach’s alpha and Composite Reliability (CR) derived from the CFA model, with values of 0.70 or higher considered acceptable [ 21 ]. Convergent validity was supported by significant and substantial standardized factor loadings (≥ 0.50) and an Average Variance Extracted (AVE) of 0.50 or higher [ 16 ]. To assess discriminant and concurrent validity, Pearson correlations were computed between the latent factor score of the SE-AI-L2AWS and the scores from the L2WSS and the writing anxiety scale. As hypothesized, the SE-AI-L2AWS demonstrated a moderate, positive correlation with L2WSS subscales, indicating discriminant validity, and a significant, moderate-to-strong negative correlation with the writing anxiety measure, indicating concurrent validity. Results Descriptive statistics and preliminary analysis The data were first screened for missing values and outliers. No missing data were found, and inspection of Mahalanobis distance revealed no significant multivariate outliers. The descriptive statistics for the 11 items of the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS), based on the full sample ( N = 680), are presented in Table 2 . Table 2. Descriptive statistics for the 11-item SE-AI-L2AWS ( N = 680) Item No. Item Content M SD Skewness Kurtosis AI-Assisted Process Regulation Efficacy 1 For my academic writing in English, I am confident that I can write clear and specific prompts to guide an AI tool to generate useful and relevant content. 5.15 1.41 -0.61 -0.45 2 For my academic writing in English, I am confident that I can use an AI tool to effectively brainstorm and explore different perspectives on a writing topic. 5.62 1.25 -0.85 0.21 3 For my academic writing in English, I am confident that I can develop a logical and well-structured outline for my essay with the help of an AI tool. 5.58 1.33 -0.79 -0.11 4 For my academic writing in English, I am confident that I can critically evaluate the accuracy and reliability of information provided by an AI tool. 4.91 1.48 -0.52 -0.68 AI-Assisted Language Efficacy 5 For my academic writing in English, I am confident that I can use an AI tool to rephrase my own sentences to improve their clarity and academic style. 5.75 1.15 -0.92 0.55 6 For my academic writing in English, I am confident that I can use an AI tool to find more precise and appropriate academic vocabulary for my writing. 5.68 1.22 -0.88 0.34 7 For my academic writing in English, I am confident that I can effectively use an AI tool to identify and correct grammatical errors in my writing. 5.71 1.19 -0.9 0.47 8 For my academic writing in English, I am confident that I can revise AI-generated text to ensure it matches my own authorial voice and writing style. 5.24 1.38 -0.65 -0.39 9 For my academic writing in English, I am confident that I can integrate suggestions or text generated by an AI tool into my own writing smoothly, without disrupting my argument. 5.33 1.35 -0.71 -0.26 Ethical and Integrity Efficacy 10 For my academic writing in English, I am confident that I can clearly distinguish between my own original ideas and contributions from an AI tool in my final work. 5.05 1.45 -0.58 -0.52 11 For my academic writing in English, I am confident that I can use AI tools as a writing assistant in a way that fully adheres to the principles of academic integrity. 5.4 1.42 -0.74 -0.3 Open in a new tab Item numbers correspond to Appendix A M Mean, SD Standard Deviation The mean scores for the items ranged from 4.91 to 5.75 on the 7-point scale, indicating that, on average, participants felt moderately to quite confident in their abilities to use AI for academic writing. Standard deviations ranged from 1.15 to 1.48, showing a reasonable degree of variance in responses. Skewness and kurtosis values were well within the recommended thresholds (i.e., |skewness| < 3 and |kurtosis| < 10) suggested by Kline [ 27 ], supporting the assumption of univariate normality and the suitability of the data for subsequent multivariate analyses. Exploratory factor analysis To empirically explore the underlying structure of the 11-item SE-AI-L2AWS, an EFA was conducted using data from Sample A ( n = 340). The KMO measure of 0.90 and a significant Bartlett’s test of sphericity (χ²(55) = 2415.77, p < .001) indicated that the data were highly suitable for factor analysis. Given that the data were collected via self-report questionnaires, potential common method variance (CMV) was assessed. Harman’s single-factor test was conducted by entering all items into an exploratory factor analysis with an unrotated solution. The results revealed that the first factor accounted for 43.2% of the total variance, which is below the recommended threshold of 50% [ 40 ]. This suggests that common method bias was not a pervasive threat to the validity of the findings. Using Maximum Likelihood estimation with Promax rotation, the analysis clearly yielded the hypothesized three-factor solution. This was supported by the scree plot, which showed a distinct elbow after the third factor, and the parallel analysis, which also indicated three factors to retain. These three factors had eigenvalues of 5.98, 1.87, and 1.22, and together they accounted for 82.55% of the total variance. Critically, all 11 items loaded cleanly and substantially onto their intended factors, with all primary loadings exceeding 0.75 and no problematic cross-loadings. This provided strong empirical support for the theoretically-derived three-factor structure without any need for item removal. The factors were confirmed as AI-Assisted Process Regulation Efficacy, AI-Assisted Language Efficacy, and Ethical and Integrity Efficacy. The final EFA pattern matrix is shown in Table 3 . Table 3. Pattern matrix from EFA for SE-AI-L2AWS ( N = 340) Item No. Item Content Summary Factor Loading Item No. Item Content Summary Factor Loading Item No. Item Content Summary Factor Loading AI-Assisted Process Regulation Efficacy AI-Assisted Language Efficacy Ethical and Integrity Efficacy 1 Write clear and specific prompts 0.86 5 Rephrase sentences for clarity/style 0.89 10 Distinguish my ideas from AI’s 0.92 2 Brainstorm and explore perspectives 0.82 6 Find precise academic vocabulary 0.87 11 Use AI adhering to academic integrity 0.9 3 Develop a logical outline with AI 0.8 7 Identify and correct grammatical errors 0.83 4 Critically evaluate AI information 0.76 8 Revise AI text to match authorial voice 0.78 9 Integrate AI suggestions smoothly 0.75 Open in a new tab Confirmatory factor analysis To cross-validate the three-factor structure identified in the EFA, a CFA was performed on the 11-item scale using data from Sample B ( n = 340). The hypothesized three-factor correlated model was tested for its fit to the data. Alternative models were examined to verify the factorial structure. The one-factor model demonstrated poor fit, indicating that the items cannot be reduced to a single dimension. A second-order model was also tested but did not yield a significantly better fit than the correlated three-factor model. Thus, the three-factor structure was retained as the most parsimonious and theoretically interpretable solution (see Table 3 for model comparison). As shown in Table 4 , the model demonstrated a good fit, with all key indices meeting or exceeding the stringent criteria for a good model fit [ 25 ]. Table 4. Fit indices for the confirmatory factor analysis model ( N = 340) Model χ² df CFI TLI RMSEA (90% CI) SRMR Three-Factor Model 76.45 41 0.988 0.985 0.049 (0.035, 0.064) 0.038 Second-Order Model 76.45 41 0.988 0.985 0.038 0.038 1-Factor Model 845.21 44 0.682 0.603 0.235 (0.221, 0.249) 0.142 Open in a new tab The standardized factor loadings in the CFA model were all substantial and statistically significant ( p < .001), ranging from 0.73 to 0.91, further confirming the robustness of the factor structure (Fig. 2 ). Fig. 2. Open in a new tab Standardized factor loadings for the confirmed three-factor model of the SE-AI-L2AWS To evaluate the stability of the SE-AI-L2AWS across different student populations, measurement invariance tests were conducted for both gender (Male vs. Female) and academic major (STEM vs. Non-STEM). Following the procedure recommended by Cheung and Rensvold [ 9 ], a hierarchical set of nested models—configural, metric, and scalar—was examined. As presented in Table 5 , the configural models for both gender and major demonstrated good fit, establishing that the three-factor structure is valid across groups. Subsequent tests for metric invariance (constraining factor loadings) and scalar invariance (constraining intercepts) resulted in negligible changes in model fit indices. These findings confirm strict invariance, suggesting that the scale items are interpreted equivalently and measure the same underlying constructs regardless of the students’ gender or disciplinary background. Table 5. Measurement invariance testing results Grouping Variable Model CFI RMSEA CFI RMSEA Decision Gender (Male vs. Female) M1: Configural 158.42 82 0.978 0.055 — — — — — M2: Metric 165.18 90 0.977 0.054 6.76 8 -0.001 -0.001 Supported M3: Scalar 174.92 98 0.975 0.054 9.74 8 -0.002 0 Supported Major (STEM vs. Non-STEM) M1: Configural 162.35 82 0.976 0.056 — — — — — M2: Metric 171.88 90 0.975 0.055 9.53 8 -0.001 -0.001 Supported M3: Scalar 185.2 98 0.972 0.056 13.32 8 -0.003 0.001 Supported Open in a new tab Reliability and validity of the scale The psychometric properties of the 11-item SE-AI-L2AWS were assessed using the full dataset ( N = 680) for reliability and data from Sample B ( N = 340) for validity analyses. Cronbach’s alpha coefficients for the subscales were all high: AI-Assisted Language Efficacy (α = 0.92), AI-Assisted Process Regulation Efficacy (α = 0.89), and Ethical and Integrity Efficacy (α = 0.87). The reliability for the total 11-item scale was excellent (α = 0.94). Composite Reliability (CR) values derived from the CFA were also strong: Language (0.93), Process Regulation (0.90), and Ethics (0.88), all exceeding the 0.70 benchmark. Convergent validity was established through two key pieces of evidence from the CFA on Sample B. First, all standardized factor loadings were high and significant (ranging from 0.73 to 0.91). Second, the Average Variance Extracted (AVE) for each factor was well above the 0.50 criterion: Language (0.73), Process Regulation (0.70), and Ethics (0.79). To assess discriminant and concurrent validity, Pearson correlations were computed between the SE-AI-L2AWS (and its subscales) and the other two validation measures using data from Sample B. The results are presented in Table 6 . Table 6. Correlations between SE-AI-L2AWS, L2WSS, and writing anxiety ( N = 340) Measure 1. SE-AI-L2AWS (Total) 2. AI Language 3. AI Process Regulation 4. AI Ethics 5. L2WSS (Teng et al.) 6. Writing Anxiety (Cheng) 1. SE-AI-L2AWS (Total) --- 2. AI Language 0.90** --- 3. AI Process Regulation 0.87** 0.66** --- 4. AI Ethics 0.81** 0.62** 0.59** --- 5. L2WSS (Teng et al.) 0.61** 0.57** 0.54** 0.49** --- 6. Writing Anxiety (Cheng) − 0.65** − 0.60** − 0.57** − 0.53** − 0.72** --- Open in a new tab p < .01 The SE-AI-L2AWS demonstrated a strong, positive correlation with the broader L2WSS ( r = .61, p < .01). This correlation is substantial enough to confirm a relationship but not so high (e.g., > 0.85) as to suggest the constructs are identical, thus supporting discriminant validity. Furthermore, the SE-AI-L2AWS showed a significant, strong negative correlation with the writing anxiety scale ( r = − .65, p < .01). This confirms the theoretical expectation that higher self-efficacy is associated with lower anxiety, providing robust evidence for the scale’s concurrent validity. To rigorously assess discriminant validity beyond simple correlations, two additional criteria were examined: the Fornell-Larcker criterion and the Heterotrait-Monotrait (HTMT) ratio (Henseler et al., 2015). First, according to the Fornell-Larcker criterion, the square root of the Average Variance Extracted (AVE) for each construct should exceed its highest correlation with any other construct. As shown in Table 7 , the square roots of the AVEs for the three SE-AI-L2AWS factors (ranging from 0.837 to 0.889) were consistently higher than their correlations with the general L2WSS (highest r = .61) and writing anxiety (highest r = − .65). This indicates that the variance captured by the SE-AI-L2AWS factors is distinct from the variance shared with general L2 writing efficacy. Table 7. Discriminant validity analysis Construct 1 2 3 4 5 1. AI-Process -0.837 2. AI-Language 0.66 -0.854 3. AI-Ethics 0.59 0.62 -0.889 4. L2WSS 0.54 0.57 0.49 5. Anxiety -0.57 -0.6 -0.53 -0.72 Open in a new tab Second, the HTMT ratios were calculated to further verify distinctiveness. All HTMT values between the SE-AI-L2AWS subscales and the L2WSS were below the strict threshold of 0.85, providing robust evidence that the SE-AI-L2AWS measures a construct empirically distinct from general L2 writing self-efficacy (Table 8 ). Table 8. Heterotrait-monotrait ratio of correlations Path / Relationship Correlation ( r ) HTMT Ratio Threshold (< 0.85) Result AI-Process <-> L2WSS 0.54 0.612 Passed Distinct AI-Language <-> L2WSS 0.57 0.645 Passed Distinct AI-Ethics <-> L2WSS 0.49 0.558 Passed Distinct AI-Process <-> AI-Language 0.66 0.734 Passed Distinct AI-Process <-> AI-Ethics 0.59 0.671 Passed Distinct AI-Language <-> AI-Ethics 0.62 0.693 Passed Distinct Open in a new tab Discussion The primary objective of this study was to address a critical measurement gap in L2 writing research by developing and validating the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS). The rapid integration of generative AI has fundamentally altered the L2 academic writing landscape, creating a new domain of functioning that existing self-efficacy instruments were not designed to capture. The first research question sought to identify the underlying factor structure of the newly developed scale. The Exploratory Factor Analysis (EFA) on Sample A revealed a clean and theoretically coherent three-factor structure that explained a substantial portion of the variance (82.55%). This structure, which comprised AI-Assisted Process Regulation Efficacy, AI-Assisted Language Efficacy, and Ethical and Integrity Efficacy, directly mirrored the conceptual framework derived from our preliminary qualitative interviews and theoretical synthesis. The fact that all 11 items loaded strongly onto their intended factors without problematic cross-loadings provided initial robust evidence for the scale’s construct validity. It is worth noting that the descriptive analysis revealed relatively high mean scores across the scale items. This trend likely reflects the demographic characteristics of the sample: university students who are digital natives and increasingly familiar with generative AI tools in their daily academic lives. According to Bandura [ 2 ], prior successful exposure to a task (mastery experience) is the primary source of self-efficacy; thus, the elevated scores may accurately reflect the growing comfort level of this population with educational technology. Importantly, statistical checks confirmed that despite the high means, the instrument retained sufficient variance to differentiate between students with moderate versus high efficacy, as evidenced by the distribution properties. The second research question aimed to test whether this hypothesized structure held true with a new sample. The Confirmatory Factor Analysis (CFA) conducted on Sample B provided strong confirmation. The three-factor correlated model demonstrated an excellent fit to the data, with all fit indices (χ²/df, CFI, TLI, RMSEA, SRMR) meeting the stringent criteria recommended in the literature [ 25 ]. This cross-validation significantly strengthens the claim that the three-dimensional structure is not an artifact of a single sample but represents a stable and replicable model of L2 writers’ self-efficacy beliefs in the AI-assisted writing context. The third research question concerned the overall psychometric properties of the scale. The SE-AI-L2AWS demonstrated excellent internal consistency, with high Cronbach’s alpha and Composite Reliability values for the total scale and each subscale. Convergent validity was strongly supported by the high, significant factor loadings and AVE values well above the 0.50 threshold, indicating that the items for each factor were indeed measuring the same underlying construct. Crucially, the scale also showed sound discriminant and concurrent validity. The moderate, positive correlation with the L2WSS [ 49 ] was precisely as hypothesized. It suggests that while confidence in using AI for writing is related to general L2 writing confidence, it is not redundant. The SE-AI-L2AWS is clearly measuring a distinct, albeit related, set of self-beliefs specific to the human-AI collaborative context. A primary concern in developing context-specific scales is ensuring they do not merely duplicate existing measures. Although the SE-AI-L2AWS correlated moderately with the L2WSS ( r = .61), the newly conducted discriminant validity analyses (Fornell-Larcker and HTMT) confirm they are distinct constructs. This statistical distinction reflects a fundamental theoretical difference: General L2 writing self-efficacy (as measured by L2WSS) focuses on the learner’s confidence in their internal cognitive resources (e.g., retrieving vocabulary from memory, manually constructing sentences). In contrast, SE-AI-L2AWS measures confidence in human-machine collaboration—the ability to orchestrate external AI capabilities. For instance, being confident in writing a topic sentence (L2WSS) is cognitively distinct from being confident in prompting an AI to generate multiple topic sentence options and then critically selecting the best one (SE-AI-L2AWS). The latter requires a layer of ‘evaluative’ and ‘regulatory’ efficacy that traditional scales do not capture. Furthermore, the strong negative correlation with writing anxiety aligns perfectly with decades of self-efficacy research [ 36 , 52 ], providing robust concurrent validity evidence. This finding suggests that a lack of confidence in one’s ability to manage AI tools is a significant new source of writing apprehension for L2 learners. The findings of this study offer several theoretical contributions. First and foremost, this research provides a timely and powerful validation of Bandura’s [ 3 ] foundational principle of task specificity in self-efficacy measurement. By demonstrating that existing scales are insufficient for the new paradigm of AI-assisted writing, we underscore that self-efficacy is not a static trait but a dynamic judgment inextricably linked to the specific demands of the task environment. As the task of writing evolves from a solitary cognitive act to a socio-technical collaboration [ 18 ], our measurement tools must evolve in tandem. Second, the three-factor structure of the SE-AI-L2AWS helps to theoretically delineate the new competencies required for L2 writing success. It reframes and extends previous conceptualizations of writing self-efficacy. For example, AI-Assisted Process Regulation Efficacy builds upon the notion of self-regulatory efficacy [ 49 , 57 ] but expands it to include external, tool-based regulation, such as prompt engineering and the critical evaluation of AI-generated content. Similarly, AI-Assisted Language Efficacy shifts the focus from a writer’s confidence in their internal linguistic knowledge (Teng et al.,’s Linguistic Self-Efficacy) to their confidence in managing and integrating an external, powerful linguistic resource while maintaining authorial voice. The emergence of Ethical and Integrity Efficacy as a strong, distinct factor is perhaps the study’s most consequential finding, signaling a profound shift in the psychological architecture of L2 writing. In pre-AI frameworks, ethics was often implicit—a matter of compliance or avoiding plagiarism. However, our results suggest that for the modern L2 writer, negotiating the ethical boundaries of AI is no longer a passive background rule but an active, distinct capability requiring high levels of self-efficacy. This distinctiveness highlights a new psychological burden: students differentiate between the technical capacity to prompt AI (Process Regulation) and the volitional confidence to use it legitimately (Ethical Integrity). A student may possess the skill to generate an essay via AI but lack the efficacy to ensure that doing so aligns with institutional standards, leading to avoidance or covert usage. This finding should be interpreted within the current climate of heightened anxiety surrounding academic integrity. As institutions scramble to define policies and deploy AI-detection software—which is prone to false positives, particularly against non-native English speakers [ 30 ]—L2 learners face a double bind. They are encouraged to embrace digital literacy yet are simultaneously policed for potential infractions. Consequently, Ethical and Integrity Efficacy likely functions as a buffer against this specific form of AI anxiety. It reflects a writer’s confidence in their ability to maintain ownership of their work and to withstand the scrutiny of authorship verification. This aligns with recent debates arguing that in the age of generative AI, “authorship” is not a static state but a continuous process of negotiation (frequently checking “is this still my voice?”). Thus, the SE-AI-L2AWS does not just measure a skill; it captures the learner’s perceived capacity to navigate the ambiguity of the post-plagiarism era, ensuring that their use of technology empowers rather than endangers their academic identity. The SE-AI-L2AWS is not merely a research tool; it holds substantial practical value for second language writing pedagogy in the AI era. Educators can employ the scale diagnostically at the beginning of a course to pinpoint students’ particular strengths and weaknesses. To practically determine a student’s self-efficacy profile using the proposed framework, educators can calculate the average scores for the overall scale and each of the three subscales (Process Regulation, Language, and Ethics). Using the 7-point Likert metric, a subscale mean below 4 (the midpoint) can serve as a diagnostic threshold indicating low self-efficacy or high apprehension in that specific domain. For example, a student might display high confidence in leveraging AI for grammar correction (e.g., scoring > 5 on Language Efficacy) but struggle with critically evaluating AI-generated content or understanding ethical boundaries (e.g., scoring < 4 on Ethical Efficacy). Such granular insights enable tailored instruction, moving beyond a uniform approach to AI literacy. To translate these dimensions into practice, educators should move beyond general AI guidelines and design specific tasks that build mastery experiences. For instance, to bolster AI-Assisted Process Regulation Efficacy, instructors could implement a “Prompt Iteration Workshop.” In this activity, students start with a vague prompt, analyze the suboptimal AI output, and then systematically refine the prompt using specific constraints (e.g., persona, format, tone) until the output meets academic standards. witnessing this cause-and-effect relationship—where their improved instructions directly lead to better results—provides tangible evidence of their ability to control the technology, thereby reducing the feeling of being overwhelmed by it. The SE-AI-L2AWS also serves as a tool for formative assessment and intervention research. Addressing Ethical and Integrity Efficacy requires moving from warning students about plagiarism to empowering them with “boundary management” strategies. A concrete pedagogical intervention could be the “Authorship Audit” exercise. For instance, students are required to submit a draft where AI-generated text is color-coded (e.g., blue for AI suggestions, black for original thought) accompanied by a rationale for why they accepted or rejected specific AI revisions. This practice transforms ethical conduct from an abstract fear into a manageable, visible procedure. By explicitly practicing the distinction between their original ideas and AI contributions (Item 10), students develop the confidence that they are maintaining academic integrity, thus lowering the anxiety associated with accidental misconduct. In addition, the scale can enhance students’ metacognitive awareness. By reflecting on their own efficacy profiles, learners can better understand their capabilities and areas for growth. This reflection promotes greater autonomy and self-regulation, encouraging students to take charge of their learning as they develop AI-assisted writing proficiency. While this study successfully established the psychometric soundness of the SE-AI-L2AWS, several limitations must be acknowledged, which in turn point to avenues for future research. First, although measurement invariance tests across gender and majors supported the scale’s stability, the data were collected from a single university in China. This limits the generalizability of the findings to other cultural contexts or educational systems with different AI policies. The split-half validation within one institution, while a standard initial psychometric step, does not constitute external cross-validation. Future research is urgently needed to test the SE-AI-L2AWS in diverse international contexts to establish its cross-cultural validity. Second, constraints related to the study design limited the scope of validation evidence. Specifically, due to the cross-sectional nature of the data collection and strict anonymity protocols, test-retest reliability could not be assessed, leaving the temporal stability of the scale unverified. Furthermore, the study relied on self-reported measures (i.e., writing anxiety and general efficacy) for concurrent validity but lacked objective behavioral or performance outcomes (e.g., actual writing grades, AI usage logs, or prompt quality ratings). Consequently, the predictive validity and incremental validity of the SE-AI-L2AWS over existing measures in predicting actual writing success remain to be empirically demonstrated. Another limitation of the present study concerns the Ethical and Integrity Efficacy dimension, which comprises only two items. While a minimum of three items is typically recommended for factor stability [ 15 ], the expert validation phase resulted in the removal of redundant items to prioritize scale parsimony. However, statistical evidence supports the retention of this factor: the two items exhibited extremely high factor loadings (> 0.90) and high internal consistency (Spearman-Brown coefficient = 0.88), which Eisinga et al. [ 14 ] suggest validates the use of two-item subscales when the construct is narrowly defined. Future studies should consider developing additional items to broaden the measurement scope of this critical dimension. Finally, future studies should address these gaps by employing longitudinal designs to track the development of efficacy over time and by incorporating objective performance metrics. Investigating the relationship between SE-AI-L2AWS and other critical variables, such as AI literacy [ 53 ], cognitive strategies, and academic achievement, will further elucidate the scale’s practical utility in the AI-integrated writing classroom. Conclusion The proliferation of generative AI has created an urgent need for new conceptual and methodological tools to understand its impact on L2 student writers. This study addressed this need by attempting to develop and validate the Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS). Through rigorous psychometric procedures, we have established that the 11-item SE-AI-L2AWS is a reliable and valid instrument for measuring a three-dimensional construct encompassing process regulation, language enhancement, and ethical integrity. This scale not only fills a measurement void in the literature but also provides a theoretically grounded framework for both researchers and educators. By enabling the precise assessment of students’ confidence in navigating the complexities of human-AI writing collaboration, the SE-AI-L2AWS offers a tool for designing effective pedagogy, fostering student agency, and advancing our understanding of L2 writing development in the AI era. Supplementary Information Supplementary Material 1 (429.4KB, pdf) Appendix Appendix A Self-Efficacy in AI-Assisted L2 Academic Writing Scale (SE-AI-L2AWS) Instructions: The following statements concern your confidence in using AI tools (like ChatGPT) to assist with your academic writing in English. Please read each statement carefully and rate how confident you feel about your ability to perform the described action on a scale from 1 (Not at all confident) to 7 (Completely confident). I. AI-Assisted Process Regulation Efficacy (Confidence in planning, monitoring, and evaluating the writing process with AI) For my academic writing in English, I am confident that I can write clear and specific prompts to guide an AI tool to generate useful and relevant content. For my academic writing in English, I am confident that I can use an AI tool to effectively brainstorm and explore different perspectives on a writing topic. For my academic writing in English, I am confident that I can develop a logical and well-structured outline for my essay with the help of an AI tool. For my academic writing in English, I am confident that I can critically evaluate the accuracy and reliability of information provided by an AI tool. II. AI-Assisted Language Efficacy (Confidence in using AI to enhance linguistic and rhetorical quality) 5. For my academic writing in English, I am confident that I can use an AI tool to rephrase my own sentences to improve their clarity and academic style. 6. For my academic writing in English, I am confident that I can use an AI tool to find more precise and appropriate academic vocabulary for my writing. 7. For my academic writing in English, I am confident that I can effectively use an AI tool to identify and correct grammatical errors in my writing. 8. For my academic writing in English, I am confident that I can revise AI-generated text to ensure it matches my own authorial voice and writing style. 9. For my academic writing in English, I am confident that I can integrate suggestions or text generated by an AI tool into my own writing smoothly, without disrupting my argument. III. Ethical and Integrity Efficacy (Confidence in using AI in a manner consistent with academic standards) 10. For my academic writing in English, I am confident that I can clearly distinguish between my own original ideas and contributions from an AI tool in my final work. 11. For my academic writing in English, I am confident that I can use AI tools as a writing assistant in a way that fully adheres to the principles of academic integrity. Authors’ contributions G.Y. and L.F. wrote the main manuscript text. All authors reviewed the manuscript. Funding Authors receive no funding for this study. Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This study was conducted in accordance with the principles of the Declaration of Helsinki. The study involving humans was approved by Human Participants Ethics Committee of Hangzhou International Innovation Institute of Beihang University (202561). The study was conducted in accordance with the local legislation and institutional requirements. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Bandura A. Social foundations of thought and action. Englewood Cliffs, NJ. 1986;1986(23–28):2. 2. Bandura A. Self-efficacy: the exercise of control. Freeman; 1997. Vol. 11. 3. Bandura A. Guide for constructing self-efficacy scales. In: Pajares F, Urdan T, editors. Self-efficacy beliefs of adolescents. Information Age Publishing. 2006;307–37. 4. Beauducel A, Herzberg PY. On the performance of maximum likelihood versus means and variance adjusted weighted least squares estimation in CFA. Struct Equation Model. 2006;13(2):186–203. 5. Braun V, Clarke V. Using thematic analysis in psychology. Qual Res Psychol. 2006;3(2):77–101. 6. Bruning R, Dempsey MS, Kauffman DF, McKim C, Zumbrunn S. Examining dimensions of self-efficacy for writing. J Educ Psychol. 2013;105(1):25–38. [ Google Scholar ] 7. Chen J, Zhang LJ. Assessing student-writers’ self-efficacy beliefs about text revision in EFL writing. Assess Writ. 2019;40:27–41. [ Google Scholar ] 8. Cheng YS. A measure of second language writing anxiety: Scale development and preliminary validation. J Second Lang Writ. 2004;13(4):313–35. [ Google Scholar ] 9. Cheung GW, Rensvold RB. Evaluating goodness-of-fit indexes for testing measurement invariance. Struct Equation Model. 2002;9(2):233–55. 10. Comrey AL, Lee HB. A first course in factor analysis. Psychology; 2016. 11. Darvin R. The need for critical digital literacies in generative AI-mediated L2 writing. J Second Lang Writ. 2025;67:101186. [ Google Scholar ] 12. Dhillon PS, Molaei S, Li J, Golub M, Zheng S, Robert LP. Shaping human-AI collaboration: varied scaffolding levels in co-writing with language models. In Proceedings of the 2024 CHI conference on human factors in computing systems. 2024. p. 1–18. 13. Dörnyei Z, Taguchi T. Questionnaires in second language research: Construction, administration, and processing. Routledge; 2010. 14. Eisinga R, Grotenhuis MT, Pelzer B. The reliability of a two-item scale: Pearson, Cronbach, or Spearman-Brown? Int J Public Health. 2013;58(4):637–42. [ DOI ] [ PubMed ] 15. Fabrigar LR, Wegener DT, MacCallum RC, Strahan EJ. Evaluating the use of exploratory factor analysis in psychological research. Psychol Methods. 1999;4(3):272–99. [ Google Scholar ] 16. Fornell C, Larcker DF. Evaluating structural equation models with unobservable variables and measurement error. J Mark Res. 1981;18(1):39–50. [ Google Scholar ] 17. Golparvar SE, Khafi A. The role of L2 writing self-efficacy in integrated writing strategy use and performance. Assess Writ. 2021;47:100504. [ Google Scholar ] 18. Graham S. A revised writer(s)-within-community model of writing. Educational Psychol. 2018;53(4):258–79. [ Google Scholar ] 19. Haas C. Writing technology: Studies on the materiality of literacy. Routledge; 2013. 20. Hafner CA. Digital literacies for English language learners. Second handbook of English language teaching. Cham: Springer; 2019. pp. 1–20. [ Google Scholar ] 21. Hair JF, Black WC, Babin BJ, Anderson RE. Multivariate data analysis (8th ed.). Cengage. 2019. 22. Hayes JR. A new framework for understanding cognition and affect in writing. In: Levy CM, Ransdell S, editors. The science of writing: Theories, methods, individual differences, and applications. Lawrence Erlbaum; 1996;1–27. 23. Hayes JR. Modeling and remodeling writing. Writ Communication. 2012;29(3):369–88. [ Google Scholar ] 24. Hirvela A, Hyland K, Manchón RM. Dimensions of L2 writing theory and research: Learning to write and writing to learn. In: Manchón RM, Matsuda PK, editors. Handbook of second and foreign language writing. De Gruyter; 2016;45–63. 25. Hu LT, Bentler PM. Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Struct Equation Modeling: Multidisciplinary J. 1999;6(1):1–55. [ Google Scholar ] 26. Jin F, Sun L, Pan Y, Lin CH. High heels, compass, spider-man, or drug? Metaphor analysis of generative artificial intelligence in academic writing. Comput Educ. 2025:105248. 27. Kline RB. Principles and practice of structural equation modeling. 4th ed. The Guilford Press; 2016. 28. Kormos J. The role of individual differences in L2 writing. J Second Lang Writ. 2012;21(4):390–403. [ Google Scholar ] 29. Lankshear C, Knobel M. New literacies: Everyday practices and social learning. McGraw-Hill Education (UK); 2011. 30. Liang W, Yuksekgonul M, Mao Y, Wu E, Zou J. GPT detectors are biased against non-native English writers. Patterns. 2023;4(7). [ DOI ] [ PMC free article ] [ PubMed ] 31. Luther T, Kimmerle J, Cress U. Teaming up with an AI: Exploring human–AI collaboration in a writing scenario with ChatGPT. AI. 2024;5(3):1357–76. [ Google Scholar ] 32. Nguyen A, Hong Y, Dang B, Huang X. Human-AI collaboration patterns in AI-assisted academic writing. Stud High Educ. 2024;49(5):847–64. [ Google Scholar ] 33. O’Connor BP. SPSS and SAS programs for determining the number of components using parallel analysis and Velicer’s MAP test. Behav Res Methods Instruments Computers. 2000;32(3):396–402. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Pajares F. Self-efficacy beliefs, motivation, and achievement in writing: A review of the literature. Read Writ Q. 2003;19(2):139–58. [ Google Scholar ] 35. Pajares F. Motivational role of self-efficacy beliefs in self-regulated learning. In: Schunk DH, Zimmerman BJ, editors. Motivation and self-regulated learning: Theory, research, and applications. Lawrence Erlbaum. 2008;111–39. 36. Pajares F, Johnson MJ. Confidence and competence in writing: The role of self-efficacy, outcome expectancy, and apprehension. Res Teach Engl. 1994;28(3):313–31. [ Google Scholar ] 37. Pajares F, Johnson MJ. Self-efficacy beliefs and the writing performance of entering high school students. Psychol Sch. 1996;33(2):163–75. [ Google Scholar ] 38. Pajares F, Valiante G. Influence of self-efficacy on elementary students’ writing. J Educational Res. 1997;90(6):353–60. [ Google Scholar ] 39. Pintrich PR. The role of goal orientation in self-regulated learning. In: Boekaerts M, Pintrich PR, Zeidner M, editors. Handbook of self-regulation. Academic. 2000;451–502. 40. Podsakoff PM, MacKenzie SB, Lee JY, Podsakoff NP. Common method biases in behavioral research: a critical review of the literature and recommended remedies. J Appl Psychol. 2003;88(5):879. [ DOI ] [ PubMed ] 41. Polit DF, Beck CT. The content validity index: are you sure you know what's being reported? Critique and recommendations. Res Nurs Health. 2006;29(5):489–97. [ DOI ] [ PubMed ] 42. Prat-Sala M, Redford P. Writing essays: Does self-efficacy matter? The relationship between self-efficacy in reading and in writing and undergraduate students’ performance in essay writing. Educational Psychol. 2012;32(1):9–20. [ Google Scholar ] 43. Schunk DH, Pajares F. Self-efficacy beliefs. In: Peterson P, Baker E, McGaw B, editors. International encyclopedia of education. 3rd ed. Elsevier; 2010;668–72. 44. Shell DF, Colvin C, Bruning RH. Self-efficacy, attribution, and outcome expectancy mechanisms in reading and writing achievement: grade-level and achievement-level differences. J Educ Psychol. 1995;87(3). 45. Shell DF, Murphy CC, Bruning RH. Self-efficacy and outcome expectancy mechanisms in reading and writing achievement. J Educ Psychol. 1989;81(1):91–100. [ Google Scholar ] 46. Silva T. Toward an understanding of the distinct nature of L2 writing: The ESL research and its implications. TESOL Q. 1993;27(4):657–77. [ Google Scholar ] 47. Sun T, Wang C. College students’ writing self-efficacy and writing self-regulated learning strategies in learning English as a foreign language. System. 2020;90:102221. [ Google Scholar ] 48. Sun T, Wang C, Lambert RG, Liu L. Relationship between second language English writing self-efficacy and achievement: A meta-regression analysis. J Second Lang Writ. 2021;53:100817. [ Google Scholar ] 49. Teng LS, Sun PP, Xu L. Conceptualizing writing self-efficacy in English as a foreign language contexts: Scale validation through structural equation modeling. TESOL Q. 2018;52(4):911–42. [ Google Scholar ] 50. Warschauer M. Digital literacy studies: progress and prospects. The future of literacy studies. 2009. p. 123–40. 51. Warschauer M, Xu Y. Artificial intelligence for language learning: entering a new era. 2024. 52. Woodrow L. College English writing affect: Self-efficacy and anxiety. System. 2011;39(4):510–22. [ Google Scholar ] 53. Xu L, Zhang L, Ou L, Wang D. The development and validation of a scale on student AI literacy in L2 writing: A domain-specific perspective. J Second Lang Writ. 2025;69:101227. [ Google Scholar ] 54. Zabihi R. The role of cognitive and affective factors in measures of L2 writing. Writ Communication. 2018;35(1):32–57. [ Google Scholar ] 55. Zhang J, Zhang LJ, Zhu Y. Development and validation of a genre-based second language (L2) writing self-efficacy scale. Front Psychol. 2023;14:1181196. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Zhang LJ. L2 writing: toward a theory-practice praxis. In: Hinkel E, editor. Handbook of practical second language teaching and learning. Routledge; 2022. p. 331–43. 57. Zimmerman BJ. Self-efficacy: An essential motive to learn. Contemp Educational Psychology. 2000;25(1):82–91. [ DOI ] [ PubMed ] [ Google Scholar ] 58. Zimmerman BJ. From cognitive modeling to self-regulation: A social cognitive career path. Educational Psychol. 2013;48(3):135–47. [ Google Scholar ] 59. Zimmerman BJ, Kitsantas A. Reliability and validity of self-efficacy for learning form (SELF) scores of college students. J Psychol. 2007;215(3):157–63. [ Google Scholar ] 60. Zimmerman BJ, Risemberg R. Becoming a self-regulated writer: A social cognitive perspective. Contemp Educ Psychol. 1997;22(1):73–101. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (429.4KB, pdf) Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Articles from BMC Psychology are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.6 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• The mediating and moderating roles of achievement goal orientation in the relationship between AI writing feedback and English writing self-efficacy among Chinese university students.#457343
• Teacher feedback VS AI-assisted peer feedback in L2 writing: A quasi-experimental study in a Chinese University.#323202
• Closing the Model Text Gap: A Feasibility Pilot of AI-Integrated Analytical Writing Instruction for L2 Learners#450616
• Vocabulary Learning Strategies and Situated Academic Writing Self-Efficacy of Students in Public Secondary Schools#687645
• Artificial Intelligence in teaching creative writing: A systematic review in L1 and additional languages learning contexts#446861