Sage Journals: Discover world-class research

Abstract

Generative artificial intelligence (AI) tools like DeepSeek hold transformative potential for data-driven learning (DDL) in translator education, yet their efficacy, equity implications, and impact on learner engagement remain underexplored, particularly outside Western-dominated research. Grounded in the Translatorial Competence Model and Cognitive Load Theory, this mixed-methods study investigated the integration of generative AI into DDL activities for 80 translation majors from a comprehensive university in eastern China over a 12-week intervention. Results indicated that AI-enhanced DDL was associated with significantly improved terminology mastery and cultural adaptation compared to traditional methods. The experimental group showed greater gains in accuracy (23% vs. 14%), terminology consistency (28% vs. 17%), and cultural appropriateness (19% vs. 9%). However, efficacy was not uniform. A significant rural-urban divide mediated outcomes, with urban students with prior AI exposure outperforming rural peers (β = .32, p = .008). This disparity was significantly reduced through structured pedagogical scaffolding, with the rural-urban proficiency gap narrowing from 14% to 5%. Learners also reported heightened motivation (M = 4.2 vs. 3.5 control, p < .001) but raised concerns about AI’s potential for errors and Western-centric biases. The study underscores that the integration of generative AI is highly conditional, requiring targeted strategies to address socioeconomic equity gaps and ensure cultural relevance. These findings offer a cautious framework for integrating AI into translator education, highlighting the prerequisites for equitable, discovery-based learning in diverse contexts.

Plain Language Summary

Generative AI in Translation Education: Enhancing Learning, Equity, and Engagement

Generative AI tools, such as DeepSeek, have the potential to transform data-driven learning (DDL) in translation education. However, their effectiveness, fairness, and impact on student engagement are not well understood, especially in non-Western settings like China. This study explores how generative AI can be integrated into DDL activities in Chinese undergraduate translation programs, where teaching is often teacher-focused and access to real bilingual materials is limited. Eighty translation students were divided into two groups: one used DeepSeek to create customized bilingual materials and practice real-world translation tasks, while the other used traditional static materials. The study measured improvements in translation accuracy, terminology use, and cultural understanding through tests before and after the experiment. Qualitative data from journals, group discussions, and observations provided insights into student engagement and fairness. The results showed that AI-enhanced DDL improved students’ mastery of terminology and cultural adaptation more than traditional methods. However, students from urban areas, who had more prior experience with AI, performed better than their rural peers, though structured training helped reduce this gap. Students reported feeling more motivated but expressed concerns about AI’s Western biases. The study highlights how generative AI can make specialized translation resources more accessible while emphasizing the need for teaching strategies to address fairness and cultural relevance. These findings offer a roadmap for using AI in translation education, supporting China’s “AI + Education” goals and promoting equitable, discovery-based learning in non-Western contexts.

Keywords

generative AI data-driven learning translator education educational equity learner engagement

Introduction

The globalization of communication has intensified the demand for skilled translators capable of producing culturally nuanced work (Dung, 2024). In response, translator education has increasingly embraced data-driven learning (DDL), an inductive approach where learners derive linguistic patterns from analyzing authentic language corpora. Despite its pedagogical value, the implementation of DDL faces significant constraints in many contexts, including China. These constraints include teacher-centered pedagogies, a scarcity of high-quality bilingual corpora, and an assessment culture favoring rote memorization over exploratory learning (Chung et al., 2024; Lusta et al., 2023). Consequently, students often rely on outdated textbook examples and static online dictionaries rather than engaging with the dynamic, real-world language data that defines modern translation work (Abou-Khalil et al., 2021).

To overcome these logistical barriers, Generative artificial intelligence (AI) presents a paradigm-shifting opportunity. Tools like DeepSeek can dynamically generate tailored, domain-specific parallel texts, offering a potential solution to the corpus scarcity that has long hampered DDL (Crosthwaite & Baisa, 2023). Theoretically, this aligns with two critical frameworks. The PACTE Group’s (2003) Translatorial Competence Model highlights the need to develop instrumental sub-competence and strategic sub-competence. Concurrently, Cognitive Load Theory (CLT; Sweller, 2010) provides a lens for understanding how AI could optimize learning. By automating the manual processes of traditional DDL, AI may reduce extraneous cognitive load, freeing mental resources for the germane load required to build strategic translation competence.

However, the integration of generative AI into pedagogy is not inherently beneficial; its efficacy is highly conditional and raises critical concerns. First, AI models are prone to generating errors and are often trained on Western-centric data, risking the perpetuation of cultural biases and the production of inaccurate terminology (Hwang & Chang, 2023; Yang & Wang, 2023). Beyond these questions of pure efficacy, a second and central concern is that the benefits of AI may be inequitably distributed. Pre-existing socioeconomic disparities, particularly the pronounced rural-urban digital divide, threaten to mediate AI’s effectiveness, potentially exacerbating educational inequalities rather than alleviating them (Zhang & Li, 2021). From a CLT perspective, for disadvantaged students, the tool intended to reduce cognitive load could instead become a new source of it. Third, an over-reliance on automated outputs could potentially undermine the development of the critical, strategic decision-making that defines expert translation (PACTE Group, 2003).

These challenges are acutely visible within China’s unique educational landscape, which is characterized by top-down “AI + Education” initiatives (Knox, 2020) yet also shaped by Confucian-heritage pedagogical traditions (Tan, 2015) that may conflict with the learner-centered ethos of AI-enhanced DDL. Consequently, while the potential is significant, empirical research on AI’s role in translator education remains dominated by Western contexts (Bo et al., 2025), leaving a critical gap in understanding its efficacy, equity implications, and impact on learner engagement in non-Western settings.

As such, this study argues that the efficacy of generative AI in DDL is not automatic but is fundamentally mediated by socioeconomic factors and pedagogical design. Therefore, our primary research question is not merely if AI-enhanced DDL works, but for whom and under what conditions it works. Grounded in the PACTE and CLT frameworks, this mixed-methods study investigates the integration of generative AI into a Chinese undergraduate translation program to address the following questions:

1. What is the comparative impact of AI-enhanced DDL versus traditional methods on developing translation competence?

2. How do socioeconomic factors, specifically the rural-urban divide, mediate the efficacy of AI-enhanced DDL?

3. How do learners perceive and engage with AI-enhanced DDL in terms of motivation and critical autonomy?

Literature Review

The integration of generative AI into DDL for translator education represents a convergence of technological innovation, pedagogical theory, and sociocultural dynamics. To contextualize this study’s focus on efficacy, equity, and engagement, this literature review examines three interrelated domains: (1) the theoretical foundations of DDL and translatorial competence, (2) the emerging role of AI in translator education, and (3) equity challenges in technology-enhanced language learning, with a focus on China’s educational landscape. By synthesizing these strands, this section establishes a framework for evaluating how generative AI can reshape DDL practices while addressing the unique demands of translator training in non-Western contexts.

DDL and Translatorial Competence

DDL, rooted in corpus linguistics, emphasizes inductive learning through direct engagement with authentic language data (cf. Lusta et al., 2023). In translator education, DDL enables learners to analyze parallel corpora (McGuire, 2019)—collections of source texts and their translated counterparts—to identify patterns in terminology, collocations, and discourse conventions. This approach aligns with the Translatorial Competence Model proposed by the PACTE Group (2003), which defines translation competence as a multicomponential construct comprising bilingual, extralinguistic, instrumental, strategic, and cultural sub-competences. Of particular relevance to DDL is the instrumental sub-competence, which involves the ability to use tools and resources—such as corpora, terminology databases, and translation memory systems—to solve translation problems. By engaging with corpora, students develop equivalents, and adhere to genre-specific conventions (Farroni, 2024).

However, traditional DDL faces limitations in translation classrooms. Static corpora often lack domain-specific or culturally nuanced texts (Zhang et al., 2024), particularly for language pairs involving less-resourced languages. Moreover, the manual compilation and annotation of corpora demand significant time and expertise, rendering them inaccessible to many educators. Generative AI tools like DeepSeek offer a potential solution by automating the creation of tailored bilingual texts (Woo et al., 2024). For instance, an instructor could prompt DeepSeek to generate parallel texts on specialized topics or simulate client briefs with specific stylistic requirements. This capability aligns with the strategic sub-competence outlined in the PACTE model, which emphasizes problem-solving and decision-making in real-world scenarios. Yet, the pedagogical efficacy of AI-generated corpora remains under-researched (Pack & Maloney, 2024), particularly in relation to their impact on translation accuracy and cultural appropriateness.

Generative AI in Translator Education

Recent advances in generative AI have spurred interest in its applications for translator training. Unlike rule-based machine translation systems, generative AI models like DeepSeek produce fluid, context-aware outputs that mimic human language use, making them valuable for simulating authentic translation tasks (Hu et al., 2024). Studies in Western contexts suggest that AI can enhance DDL by providing instant feedback, generating practice materials, and fostering learner autonomy (e.g., Mizumoto, 2023). For instance, students might use DeepSeek to extract collocations from AI-generated texts or compare multiple translation variants to identify optimal solutions.

Such activities must be examined through the lens of CLT (Sweller, 2010), which provides a crucial framework for understanding the mental architecture involved in learning complex skills like translation. CLT distinguishes between three types of cognitive load: intrinsic (the inherent difficulty of the material, e.g., translating a complex legal text), extraneous (load imposed by the instructional design or tool interface), and germane (load devoted to schema construction and automation). Effective learning occurs when instructional design minimizes extraneous load and optimizes germane load (Sweller, 2010) to manage intrinsic load.

The integration of AI into DDL presents a dual potential from a CLT perspective. On one hand, by automating the laborious process of corpus compilation and providing immediate, contextualized examples, AI can significantly reduce extraneous cognitive load (Li, 2024). Students are freed from the technical hurdles of building and querying complex databases, allowing them to dedicate more cognitive resources to the intrinsic challenge of comparative text analysis and strategic decision-making (Al-Obaydi & Pikhart, 2020). This aligns with studies suggesting that well-designed tools can offload routine tasks (Moreno & Park, 2010), facilitating deeper engagement with core concepts.

On the other hand, poorly integrated AI can increase extraneous load (Lampou, 2023). For novice learners or those with low digital literacy, the process of formulating effective prompts, interpreting potentially flawed or biased AI outputs, and troubleshooting the technology itself can create a new source of cognitive overhead (Walter, 2024). This risk is particularly acute in contexts with a pronounced digital divide, where unfamiliarity with the tool could overwhelm the cognitive benefits it offers. Therefore, the central pedagogical challenge is not merely to introduce AI, but to design scaffolding that ensures it functions as a load-reducing facilitator of germane processing rather than a load-inducing obstacle.

However, the existing literature, while identifying potential, suffers from a lack of critical empirical investigation into these concerns. Many studies champion AI’s potential but fail to adequately test its limitations under real-world pedagogical conditions (cf. Díaz & Nussbaum, 2024). First, the probabilistic nature of generative AI means its outputs can contain errors—including factual inaccuracies, terminological inconsistencies, and stylistic infelicities—that require critical evaluation by the user. This is compounded by the models’ training on vast datasets dominated by English and other high-resource languages, which embeds Western-centric biases and can increase error frequency for certain language pairs or domains (Singh, 2024). For Chinese translator students, this raises the risk of internalizing inaccurate or culturally inappropriate equivalents, particularly when translating culturally specific concepts (e.g., traditional Chinese medicine terms). Second, over-reliance on AI-generated content could undermine the development of strategic sub-competence, as students may prioritize algorithmic suggestions over critical analysis (Zábojník & Hromada, 2024; Zhai et al., 2024). Third, the “black box” nature of AI systems obscures the decision-making processes behind their outputs, complicating efforts to teach students how to evaluate and justify their translation choices (Franzoni, 2023). These challenges underscore the need for pedagogical frameworks that leverage AI’s strengths while mitigating its limitations.

Equity in AI-Enhanced Translator Education in China

The digital divide is a global challenge, but its manifestations are locally specific (Afzal et al., 2023; Clark & Gorski, 2001). China presents a compelling case study due to the scale and policy context of its rural-urban disparities. The integration of AI into education is often framed as a democratizing force (Prinsloo & Khalil, 2024), yet its benefits are unevenly distributed. In China, where socioeconomic disparities between urban and rural regions persist, access to technology and digital literacy skills vary significantly (Song, 2023). Urban universities in cities like Beijing or Shanghai often boast advanced infrastructure and partnerships with tech firms, enabling students to engage with cutting-edge AI tools. In contrast, rural institutions may lack reliable internet access or funding for AI training programs, exacerbating existing inequities in translator education. This divide aligns with Afzal et al. (2023) concept of the digital divide, which encompasses not only physical access to technology but also the skills and cultural capital required to use it effectively—a challenge common to many education systems.

Within translation classrooms, equity issues manifest in multiple ways. Students from rural backgrounds may struggle with the technical demands of AI tools, such as formulating effective prompts for DeepSeek or interpreting its outputs (Vandenberg et al., 2023; Walter, 2024). Additionally, generative AI’s reliance on English-centric training data may disadvantage students with weaker English proficiency (Kshetri, 2024), as they might misinterpret AI-generated suggestions or fail to recognize cultural biases. These challenges are compounded by China’s Confucian-heritage educational culture (You, 2018), which traditionally prioritizes teacher authority and standardized assessments over exploratory, learner-centered approaches. While AI-enhanced DDL has the potential to shift this dynamic by fostering autonomy, it also risks marginalizing students who lack the confidence or resources to navigate self-directed learning environments.

To address these inequities, emerging studies highlight strategies (e.g., Crosthwaite & Steeples, 2022; Woo et al., 2024). Structured training programs that scaffold AI literacy—such as workshops on prompt engineering or critical evaluation of AI outputs—have been shown to reduce performance gaps between students with varying levels of prior experience (Walter, 2024). Furthermore, institutional policies that mandate equitable access to technology, such as subsidized software licenses or device-sharing initiatives, can mitigate hardware-related barriers (Gonzales et al., 2018). However, few studies have examined how these strategies apply to translator education, particularly in contexts where linguistic and cultural specificity are paramount.

Gap Synthesis and Present Study

This review synthesizes three critical and interconnected lacunae that this study is designed to address, moving the discourse from mere application to critical expansion.

First, a pronounced contextual gap exists. Empirical research on AI-enhanced DDL remains predominantly situated in Western, high-resource contexts, which assume norms of digital access and learner autonomy that are not universal. The efficacy of these approaches in non-Western settings, particularly within China’s socio-educational landscape, remains severely underexplored.

Second, and most critically, there is a significant equity-mediated efficacy gap. While the digital divide is often acknowledged abstractly, there is a scarcity of empirical research investigating how specific socioeconomic factors (e.g., the rural-urban divide) directly mediate learning outcomes. The literature frequently champions AI’s potential without critically examining its power to exacerbate existing inequities.

Third, a theoretical-pedagogical gap persists. While CLT is occasionally referenced, its specific application to the cognitive architecture of AI-augmented translation tasks is underdeveloped. Furthermore, discussions of engagement often overlook the critical competence required to navigate AI’s inherent cultural biases—a skill paramount to translation.

Consequently, this study does not merely transpose a Western model but seeks to critically expand the discourse. It investigates the integration of generative AI into DDL in China to provide a nuanced framework for evaluating how AI can be harnessed to bridge—rather than widen—existing gaps.

Methodology

Participants

The study involved 80 undergraduate translation majors (Female = 62; Male = 18; Mean age = 20.4, SD = 1.2) enrolled in a 4-year English-Chinese translation program at a comprehensive university in eastern China. This institution was selected for its strategic relevance, as it actively participates in national “AI + Education” initiatives, and its student body reflects a balanced mix of urban and rural backgrounds (approximately 60% urban, 40% rural in this sample), providing a microcosm of the national socioeconomic diversity within Chinese higher education.

Participants were recruited from second—(n = 45) and third-year (n = 35) cohorts to ensure foundational proficiency in both languages and familiarity with translation principles. All participants had passed the TEM-4 (Test for English Majors-Band 4), with scores ranging from 65 to 82 (M = 72.5, SD = 4.3), indicating an intermediate-to-upper-intermediate level of English proficiency.

Stratified random sampling was employed to divide participants into experimental and control groups (40 students each). Stratification criteria included urban/rural residency (Urban = 47; Rural = 33), prior exposure to AI tools (assessed via a pre-intervention survey: Novice = 32, Intermediate = 35, Advanced = 13), and academic performance (GPA from the previous semester, divided into three tiers: top 30% [GPA ≥ 3.62], middle 40% [3.20 ≤ GPA < 3.62], and bottom 30% [GPA < 3.20]; M = 3.41, SD = .38). Urban residency was defined as having completed secondary education in Tier-1 or Tier-2 cities (e.g., Beijing, Shanghai, Hangzhou), while rural residency encompassed students from counties, townships, or villages. This stratification ensured balanced group composition for a robust analysis of equity-related outcomes. Independent samples t-tests confirmed no pre-intervention differences in translation proficiency between the experimental and control groups, t(78) = .92, p = .36. The demographic and proficiency profile of the participants is summarized as follows (Table 1).

Table 1.

Participant Demographics and Baseline Characteristics.

Characteristics	Total (N = 80)	Experimental (n = 40)	Control (n = 40)
Gender
Female	62	30	32
Male	18	10	8
Year of study
Second year	45	23	22
Third year	35	17	18
Residency
Urban	47	24	23
Rural	33	16	17
Prior AI experience
Novice	32	16	16
Intermediate	35	18	17
Advanced	13	6	7
TEM-4 score (mean ± SD)	72.5 ± 4.3	72.8 ± 4.1	72.2 ± 4.5
GPA (mean ± SD)	3.41 ± .38	3.44 ± .36	3.38 ± .40

Intervention

The 12-week intervention integrated the generative AI tool DeepSeek (DeepSeek-Coder 1.0 via API) into DDL activities for the experimental group, while the control group utilized traditional static corpora. To ensure standardization, a fixed prompt template was used for all AI-generated materials: “Generate a parallel English-Chinese text of approximately (N) words on the topic of (topic). Ensure the text includes (N) domain-specific key terms and reflects a (register) register appropriate for (context).” For example: “generate a parallel English-Chinese text of approximately 300 words on the topic of ‘intellectual property law in software licensing agreements. Ensure the text includes 8 to 10 domain—specific key terms and reflects a formal register appropriate for legal contexts’”.

Both groups attended weekly 90-min translation workshops focused on specialized domains (legal translation), aligned with curriculum requirements. All instructors (N = 4) participated in an 8-hr standardized training workshop to ensure consistent delivery of tasks, adherence to lesson plans, and awareness of potential AI biases to mitigate in classroom discussions.

For the experimental group, AI-enhanced DDL tasks were structured around three pillars, with extensive pedagogical scaffolding: (1) Guided Corpus Generation and Critical Evaluation: Students were trained from the outset that AI outputs are generative suggestions, not authoritative answers. They learned to generate tailored corpora using structured prompts and were explicitly trained to critically evaluate outputs for factual errors, terminological imprecision, and cultural biases. For example, a core task required students to generate a parallel text on commercial liability using the prompt: “Act as a legal translator. Generate a parallel English-Chinese text of approximately 250 words on the topic of ‘liability in commercial partnerships’. Ensure the text includes the term ‘joint and several liability’ and 7 to 9 other relevant legal terms.” (2) Pattern Identification and Analysis: Learners extracted terminology, collocations, and discourse patterns from these AI-generated texts using concordance lines. (3) Translation and Iterative Refinement: Students translated AI-generated client briefs, using the AI as an analytical aid for brainstorming and verification rather than for direct translation, followed by peer feedback loops.

The control group employed static corpora from established resources such as the Bilingual Corpus of Chinese Classics (BCCC). Tasks mirrored those of the experimental group in domain and difficulty but relied on pre-existing, pre-compiled corpora.

Both groups participated in identical pre-/post-tests. All assessments were scored by two independent raters using a validated rubric. Inter-rater reliability was high (Cohen’s κ = .89); discrepancies were resolved through discussion to reach a consensus.

Data Collection

Data were collected through a triangulated approach, integrating quantitative proficiency metrics, self-reported engagement surveys, and qualitative reflections on equity and learning experiences.

Quantitative data included pre-/post-tests and engagement surveys. Participants completed translation tasks in legal, technical, and literary domains, which were scored using a rubric evaluating accuracy, terminology consistency, and cultural appropriateness. Rubric validity was established through pilot testing and inter-rater reliability (Cohen’s κ = .85). Engagement surveys consisted of a 20-item Likert-scale survey adapted from the Motivated Strategies for Learning Questionnaire (Rotgans & Schmidt, 2010), measuring motivation, self-efficacy, and autonomy. Items included statements such as, “I felt confident using AI tools to solve translation problems,” rated from 1 (strongly disagree) to 5 (strongly agree).

Qualitative data were gathered through reflective journals, focus groups, and classroom observations. Students maintained weekly journals documenting challenges and successes. Four semi-structured focus groups (two per cohort) explored equity themes, with prompts such as, “Did your background affect your ability to use AI tools?” Sessions were audio-recorded and transcribed. Researchers observed eight classroom sessions (four per group), noting patterns in tool usage, peer collaboration, and instructor-student interactions.

Data Analysis

Quantitative and qualitative data were analyzed separately before integration to address the research questions.

Quantitative analysis involved ANCOVA to compare post-test scores between groups, controlling for pre-test performance, urban/rural residency, and prior AI experience. Multiple regression analysis identified variables influencing outcomes, such as access to high-speed internet or participation in AI training workshops. Descriptive statistics and t-tests compared survey responses between groups.

Qualitative data underwent thematic analysis following Braun and Clarke’s (2006) six-phase approach using NVivo 12 software. This process involved a thorough familiarization with the data from reflective journals and focus group transcripts. Initial codes (e.g., “AI dependency,”“cultural bias”) were generated and subsequently clustered into broader themes (e.g., “tool reliability,”“equity barriers”). To ensure consistency, two researchers independently coded 20% of the transcripts (initial Cohen’s κ = .78); discrepancies were resolved through discussion until a consensus was reached (final κ = 1.0), and the refined codebook was applied to the remaining data. Finally, observational notes were analyzed to identify behavioral patterns (e.g., “frequent tool troubleshooting,”“peer mentoring”); these patterns were used to triangulate and contextualize the findings from the surveys and journals.

Ethical Considerations

The study adhered to ethical guidelines for educational research. Participants provided informed consent, with explicit clarification that AI interactions would be anonymized and data stored securely. To address algorithmic bias, students participated in a workshop critiquing DeepSeek’s cultural limitations, such as its tendency to prioritize Western legal terminology over Chinese equivalents. Compliance with China’s Personal Information Protection Law (PIPL) ensured data privacy, particularly for rural students with limited digital literacy.

Results

Efficacy: AI-Enhanced DDL Fosters Critical Triangulation and Reveals Domain-Specific Limits

Quantitative analysis revealed significant differences in translation proficiency between the experimental (AI-enhanced DDL) and control (traditional DDL) groups (Table 2). ANCOVA results, controlling for pre-test scores, demonstrated that the experimental group outperformed the control group in post-test assessments across all three domains. Table 2 presents the pre-/post-test differences in translation competence by group, showing moderate to large effect sizes, particularly in the legal and technical domains.

Table 2.

Pre-/Post-Test Differences in Translation Competence by Group.

Domain	Group	Pre-test M (SD)	Post-test M (SD)	Adjusted post-test M (SE)	F	p	η²
Legal	Experimental	65.2 (7.1)	83.1 (6.5)	82.9 (.8)	12.34	<.001	.14
Legal	Control	64.8 (6.9)	74.5 (7.2)	74.7 (.8)	12.34	<.001	.14
Technical	Experimental	63.5 (8.2)	80.2 (7.1)	80.0 (.9)	9.87	.002	.11
Technical	Control	62.9 (7.8)	72.8 (7.9)	80.0 (.9)	9.87	.002	.11
Literary	Experimental	68.1 (6.3)	78.9 (6.8)	78.7 (.7)	7.02	.01	.08
Literary	Control	67.7 (6.5)	74.3 (7.1)	74.5 (.7)	7.02	.01	.08

Note. ANCOVA controlled for pre-test scores. M = mean; SD = standard deviation; SE = standard error; η² = eta-squared.

Breakdowns of rubric criteria further clarified these gains (Figure 1). In accuracy, the experimental group showed a 23% improvement, compared to 14% in the control group. For terminology consistency, AI-supported students achieved a 28% increase, versus 17% for the control group. The most notable divergence emerged in cultural appropriateness: experimental group translations exhibited a 19% enhancement, while the control group improved by only 9%. (Anomaly Handling: Prior to analysis, two outliers were identified in the control group’s post-test literary translation scores via Z-score analysis ([Z| > 3.29]. Winsorization was applied, and the analysis was re-run. This adjustment had no material impact on the significance or interpretation of the overall results).

Figure 1.

Rubric breakdown of pre- and post-test.

Qualitative data revealed that the efficacy of AI-enhanced DDL was not merely about faster access to information, but the fostering of a more sophisticated research methodology. Students developed a practice we term “critical triangulation”—using AI as a powerful starting point for a rigorous validation process against authoritative human sources. For example, Student #22 (Urban, High-Proficiency) documented a detailed process for translating the term “fiduciary duty” that exemplifies this new skill:

DeepSeek’s first suggestion was “信托责任”… I then cross-referenced this with our university’s legal database and a native-speaker forum discussion. The AI didn’t give me the answer; it gave me a curated shortlist to investigate, saving hours of searching and helping me learn the nuance.

This quote demonstrates a paradigm shift in instrumental sub-competence. Students were no longer passive recipients of information from static tools but active managers of an AI-human verification workflow, leading to deeper conceptual understanding rather than just terminological acquisition.

Interestingly, a counterintuitive yet significant finding was that AI’s limitations in literary translation became powerful pedagogical moments. Students developed a keen metacognitive awareness of the AI’s weaknesses, which in turn sharpened their own strategic decision-making. This is not common knowledge; it’s a unique insight into using AI failures pedagogically. One student (Student #57) provided a detailed critique that highlights this:

The AI translated ‘Hope is the thing with feathers’ as “希望是带有羽毛的东西”… It’s grammatically correct but dead… The AI’s failure was actually a great lesson—it forced us to define what the AI couldn’t do and why the human translator is irreplaceable in poetry.

This case illustrates the development of strategic sub-competence—the decision to override the tool—which was sparked by the AI’s specific failures. The learning occurred not despite these failures, but because of them; they forced students to articulate the nuanced, cultural, and aesthetic criteria that define quality in literary translation.

Equity: The Rural-Urban Divide is a Critical Mediator of Efficacy

Crucially, this overall efficacy was not uniform. Regression models identified urban residency (β = .32, p = .008) and prior AI experience (β = .27, p = .02) as significant predictors of post-test scores in the experimental group. Urban students with “advanced” AI familiarity achieved a 31% improvement in terminology consistency, compared to 19% for rural peers with “novice” status.

The qualitative data uncovered that the equity gap was, initially, a cognitive load gap. For rural students, the AI was not a tool but a significant source of extraneous cognitive load and anxiety. This was quantitatively supported by time-on-task analysis, which showed rural students spent 37% more time troubleshooting technical issues (mean weekly troubleshooting time: rural = 12.5 min; urban = 9.1 min; t[78] = 2.91, p = .005). A subsequent regression analysis confirmed that this effect remained significant even after controlling for confounding variables such as personal internet access quality and device type.

The mental effort required to understand the tool’s interface and basic logic competed for the finite cognitive resources needed for the translation task itself. Student #08’s (Rural, Novice) description of her first week is a powerful testament to this:

I felt a real sense of panic. My urban classmates were instantly asking DeepSeek complex questions. I didn’t even know what to ask. My first prompt was just “Translate this sentence”. I got an answer, but it was no better than Google Translate. I felt behind and stupid. The tool felt like a barrier, not a bridge.

This narrative vividly exemplifies the CLT principle in action: the very tool designed to reduce extraneous load became its primary source for novices, creating an initial chasm in experience and self-efficacy that directly impacted learning capacity.

Yet, the study’s most significant finding was the demonstrable power of structured pedagogical intervention to bridge this gap. The journals documented a transformation that was not merely about skill acquisition but about identity and empowerment. The same Student #08 described this turning point after scaffolded training:

The workshop on “prompt engineering for translation” was a turning point. We learned formulas like: [Role] + [Task] + [Context] + [Format]. For my legal text, I wrote: “Act as a legal translator. Analyze the following English clause for ambiguous terms. Suggest 3 Chinese translation options for the term ‘joint and several liability’ and recommend the best one for a PRC contract.” The difference was night and day. I wasn’t just using AI; I was commanding it. I finally felt like my classmates’ equal.

This evolution from anxiety to mastery is critical. It shows that AI literacy—specifically, the ability to formulate strategic prompts—functions as a form of cultural capital that is not innate but can be explicitly taught. This finding moves beyond identifying the problem to providing a scalable pedagogical solution, explaining the quantitative result where the rural-urban proficiency gap narrowed from 14% to 5%.

Despite the overall progress, a more nuanced and pernicious inequity persisted, revealing a challenge for AI in translator education. Students with lower English proficiency faced a double disadvantage: they lacked the prior experience with the tool and the linguistic ability to detect its subtle errors. Student #15 (Rural, Lower English Proficiency) illustrated this with a critical incident:

The AI translated “the court upheld the verdict” as “法院维持了原判” (fǎyuàn wéichíle yuán pàn). I didn’t know “upheld” and trusted the AI. Later, the instructor said while technically correct, “维持” (wéichí—to maintain) is neutral, but a stronger, more precise term in Chinese legal discourse would be “驳回了上诉” (bóhuíle shàngsù—rejected the appeal) which better conveys the finality. Because my English was weaker, I couldn’t sense the nuance in the source text to question the AI’s good-enough output.

This case highlights a profound equity issue specific to translator learning with AI: the tool’s utility and reliability are contingent on the user’s pre-existing language skills. Without the ability to critically evaluate the source text, students are vulnerable to the AI’s “good-enough” outputs, potentially cementing the disadvantages of the least proficient students and creating a new, hidden layer of inequity that is harder to remediate.

Engagement: Learners Report High Motivation Amidst Critical Concerns

Quantitative surveys revealed heightened engagement in the experimental group (Figure 2). On a 5-point Likert scale, AI-supported students reported significantly higher motivation (M = 4.2 vs. 3.5 control, p < .001), self-efficacy (M = 4.0 vs. 3.3 control, p = .004), and autonomy (M = 4.1 vs. 3.4 control, p = .002).

Figure 2.

Students’ engagement with AI support.

Qualitative data revealed the complex nature of this engagement, characterized by a dynamic of “productive paranoia.” Students were highly motivated but channeled that energy into rigorous verification practices. Classroom observations detailed a common ritual: a student would generate a term from DeepSeek, then immediately open two other tabs—a specialized terminology database (e.g., Tmxmall) and a search engine to check usage in recent Chinese news articles. This wasn’t distrust, but a professionally relevant critical engagement.

A focus group with high-achieving students revealed a strategic approach:

We don’t see DeepSeek as an answer key, one explained. It’s the most advanced member of our team, but it’s also an intern who sometimes hallucinates. It’s our job, as the project manager, to fact-check its work. That’s the real skill we’re learning.

This reframing of the translator’s role from a passive recipient of information to an active manager and auditor of AI output is a significant and novel finding that moves beyond simple metrics of motivation or anxiety.

However, engagement was tempered by critical concerns that extended beyond the tool to the institutional system. Students expressed acute anxiety about the misalignment between this new pedagogy and traditional assessment methods. Student #41 articulated a common fear:

I love using this for practice, but what about the final exam? If we can’t use AI there, are we being set up to fail? Or will the exams just test our ability to use AI, which feels like a different skill altogether?

This points to a crucial, often-overlooked institutional challenge: the potential for a “pedagogy-assessment gap” that can undermine the perceived value of new skills and create learner anxiety, a finding regarding the systemic barriers to AI integration.

Discussion

Transformed Competence Tempered by Instrumental Dependence

First, the results demonstrate a significant yet contingent alignment with the PACTE model. The integration of AI did not merely augment but fundamentally reconfigured the development of instrumental sub-competence (PACTE Group, 2003), transforming it from a mechanical skill of resource location into “a dynamic process of critical validation” (Zhao et al., 2024). This stands in stark contrast to traditional DDL, where significant effort is expended on the manual compilation of corpora. Conversely, AI-enhanced DDL shifted the cognitive focus toward interrogation and evaluation, as empirically illustrated by Student #22’s meticulous translation of “fiduciary duty.” Here, the AI-generated corpora facilitated meta-cognitive validation strategies, catalyzing a critical shift from passive information consumption to active, critical investigation.

Furthermore, the observed development of strategic sub-competence emerged not from seamless AI performance, but conversely, from its failures. This presents a crucial dichotomy: the tool’s usability inherently risked promoting uncritical adoption, yet when its limitations were exposed within a pedagogical framework that explicitly framed the AI as fallible, these moments became potent catalysts for learning (Cheng, 2024; Kim, 2019 ). The conscious decision to override AI output, therefore, was not an inherent feature of the technology but was directly mediated by pedagogical intervention. This contrast underscores that the AI’s value lies not in its infallibility, but in its capacity to create opportunities for exercising human judgment (Spaulding, 2020), thereby reinforcing the indispensability of nuanced problem-solving in translation.

Second, the study offers robust empirical support for CLT as the principal mechanism behind this conditional efficacy. Generative AI tools like DeepSeek function by drastically reducing the extraneous cognitive load associated with manual corpus linguistics, quantified by an approximate 80% decrease in search time. This resource reallocation is the critical link explaining the superior performance in higher-order tasks; the experimental group’s marked improvement in cultural appropriateness scores (19% vs. 9%) provides quantitative evidence that freed cognitive resources were directed toward germane load (Sweller, 2010), such as cultural adaptation and strategic decision-making.

However, the most compelling argument derived from CLT is that this efficacy is not universal but is critically moderated by access and scaffolding. This reveals a fundamental contrast in user experience: for urban students with prior exposure, the AI immediately functioned as a load-reducing tool, amplifying their advantages. In direct contrast, for their rural peers, the novel interface initially functioned as a load-inducing barrier, potentially exacerbating existing inequities. Therefore, the ultimate measure of this method’s success is not its peak performance for advanced users, but its capacity to engineer “equity through scalable pedagogy” (Zacamy & Roschelle, 2022). The empirical finding that structured scaffolding reduced the rural-urban proficiency gap from 14% to 5% is paramount. This convergence powerfully argues for a paradigm shift away from a deficit model (which attributes outcomes to student background) and toward a model of institutional responsibility (which actively constructs equity through design). Qualitatively, this shift is embodied in the rural students’ trajectory from initial disorientation (“lost”) to proficient command, signifying the acquisition of AI literacy that levels the epistemological playing field (cf. Tenório et al., 2023).

Democratization Tempered by Digital Divides and Cultural Bias

The most critical finding, however, is that this efficacy was not a given for all students. The study starkly reveals that generative AI’s promise to democratize access to specialized translation resources is initially constrained by pre-existing socioeconomic and digital divides. For rural students, the initial encounter with AI tools often had the opposite effect on cognitive load: it dramatically increased extraneous load. The cognitive effort required to understand the tool’s interface, learn effective prompt engineering from scratch, and troubleshoot basic technical issues competed for the same finite cognitive resources needed for the translation task itself. This initial disparity, driven by gaps in internet access, device availability, and prior digital literacy (Song, 2023), vividly illustrates Afzal et al.’s (2023) concept of the multifaceted digital divide. This finding challenges narratives of AI as an inherently democratizing force (Prinsloo & Khalil, 2024) by showing how it can initially exacerbate cognitive inequities. These regional inequities make China a critical case study in how generative AI can, without targeted intervention, perpetuate and even amplify existing educational disparities.

However, the crucial finding that structured pedagogical scaffolding—such as providing prompt engineering templates and step-by-step technical guides—significantly narrowed this gap offers a powerful blueprint for equitable implementation. This demonstrates that access alone is insufficient; meaningful access requires developing specific AI literacy skills (Walter, 2024). The convergence of performance trajectories by Week 12 suggests that targeted interventions can effectively build the necessary cultural capital for students from underserved backgrounds to leverage AI effectively. Nevertheless, residual inequities persisted, manifested in the disproportionate time rural students spent troubleshooting technical issues and the disadvantage faced by those with weaker English proficiency when navigating the AI’s Western-centric biases.

The latter point is critical: DeepSeek’s training data, dominated by English and Western perspectives (Singh, 2024), introduced culturally inappropriate suggestions that were harder for less proficient students to detect. This highlights a fundamental tension—while AI can generate Chinese translations, its underlying logic and biases often reflect its Western training origins (Yang & Wang, 2023). Thus, achieving true equity requires not only bridging the digital access and skills gap but also actively developing students’ critical AI literacy to identify, challenge, and compensate for these biases, potentially through cross-referencing with local resources as students began doing spontaneously.

Heightened Motivation Juxtaposed with Critical Ambivalence

The quantitative and qualitative data consistently pointed towards significantly heightened learner engagement within the AI-enhanced DDL environment. Increased motivation, self-efficacy, and autonomy reported by the experimental group align with findings on AI’s potential to foster learner-centered exploration (Mizumoto, 2023; Zhou & Hou, 2024). Students perceived DeepSeek as an “always-available tutor,” enabling iterative practice, experimentation with complex tasks earlier in their training, and fostering a sense of ownership over their learning (Lewis, 2014)—core tenets of DDL. This shift toward greater autonomy represents a potential disruption to traditional teacher-centered Confucian-heritage pedagogies (Tan, 2015), aligning with broader “AI + Education” goals in China (Knox, 2020).

However, this enthusiasm was tempered by a significant undercurrent of critical ambivalence. Learners readily acknowledged the efficiency gains but expressed profound anxieties about over-reliance potentially eroding their critical thinking and strategic decision making skills—competencies central to professional translatorial competence (PACTE Group, 2003). This mirrors global concerns about AI deskilling learners (Zábojník & Hromada, 2024; Zhai et al., 2024).

Furthermore, engagement was actively shaped by encounters with the AI’s cultural limitations. Frustration with Western-centric outputs, such as politically loaded translations or culturally insensitive metaphors, prompted students to develop compensatory critical evaluation strategies. This active negotiation, where learners neither blindly accepted nor wholly rejected AI suggestions but engaged in critical scrutiny, represents a sophisticated form of engagement crucial for future translators operating in an AI-mediated landscape. It underscores that engagement with AI in education must encompass critical awareness and evaluative judgment (Bearman et al., 2024), not just enthusiastic usage.

Toward Pedagogically Scaffolded and Equitable Integration

This study underscores that generative AI is not merely a tool but a transformative agent in translator education, demanding careful pedagogical orchestration. Its power lies in generating dynamic, personalized learning resources that enhance instrumental and strategic competences (Talgatov et al., 2024), particularly in specialized domains, while fostering learner autonomy. However, this transformation is not automatic or universally beneficial. Its success hinges on proactively addressing the equity gaps it can initially exacerbate and fostering the critical literacy needed to navigate its inherent biases. The findings advocate for a pedagogically scaffolded integration approach:

Balanced pedagogy: Position AI as a powerful supplement to, not a replacement for, human expertise and critical thinking. Emphasize its role in exploration, pattern identification, and generating options, while reserving final judgment, cultural nuance refinement, and creative adaptation for the learner guided by the instructor. For instance, instructors can design exercises where students must compare AI-generated translations with human-produced ones, justifying their preference based on cultural nuance and stylistic appropriateness.

Mandatory AI literacy scaffolding: Implement structured, mandatory training modules focusing on prompt engineering tailored to translator needs, critical evaluation of AI outputs (especially for cultural bias and factual accuracy), and understanding AI limitations. This is essential for leveling the playing field and empowering all students. A concrete strategy is to conduct workshops where students practice deconstructing ineffective prompts and collaboratively build effective ones for specific translation scenarios.

Proactive equity measures: Institutions must implement policies ensuring equitable access to technology (hardware, software, connectivity) and prioritize AI literacy training for students from disadvantaged backgrounds. Peer mentoring programs, as observed organically in this study, should be formally encouraged. An actionable example is creating a “digital buddy” (Kwan et al., 2023) system that pairs less-experienced students with advanced peers for the first few weeks of AI tool usage.

Culturally responsive AI development and use: Educators should collaborate with developers to advocate for and utilize more linguistically and culturally diverse AI training data. Within the classroom, explicitly teach strategies to identify and mitigate Western-centric biases, integrating AI use with local terminological databases and critical cultural analysis. A practical classroom activity involves students using AI to translate culturally-loaded concepts (e.g., “individualism”) and then using local news media and academic databases to critique and correct the AI’s output.

In a nutshell, generative AI like DeepSeek offers a compelling pathway to revitalize DDL within Chinese translator education, aligning it with the demands of the digital age and national “AI + Education” initiatives. It demonstrably enhances specific competencies and engagement when implemented thoughtfully. However, its integration must be guided by a commitment to pedagogical equity and critical awareness.

Conclusion

This study demonstrates that the question of AI’s efficacy in education is inseparable from the question of equity. Our findings show that generative AI’s potential is highly conditional, and its primary condition is the presence of pedagogical scaffolding designed to mitigate pre-existing socioeconomic disparities, hinging on pedagogical and institutional support to navigate its promises and pitfalls. While AI-enhanced DDL facilitated measurable gains in terminology accuracy and cultural adaptation—aligning with the development of instrumental and strategic sub-competences—its benefits were not uniformly distributed. A significant rural-urban divide initially mediated outcomes, underscoring that without intervention, technology can exacerbate existing inequities.

Our primary contribution is a dual theoretical extension. First, we empirically apply CLT to AI-mediated learning, revealing the digital divide not just as an access issue but fundamentally as a disparity in managing extraneous cognitive load. This extends CLT by demonstrating that a tool’s cognitive impact is not intrinsic but is mediated by user familiarity, positioning digital literacy as a key factor in instructional design for complex cognitive tasks. Second, we advance the PACTE model by testing it in an AI-augmented environment, showing its sub-competences are developable but contingent on scaffolded pedagogy that teaches critical tool use and counters AI’s cultural biases. This contributes to the model by explicitly integrating ‘critical AI literacy’ as a core component of modern instrumental and strategic sub-competence, essential for translators in an AI-saturated profession. Thus, efficacy is not inherent to the technology but is a function of equitable instructional design. While this study was conducted in China, the mediating role of the rural-urban divide and the critical importance of pedagogical scaffolding are likely to be relevant in any context where socioeconomic and digital inequities exist. China’s experience offers a framework for understanding and addressing these universal challenges in technology integration.

This study has several limitations that caution against overgeneralization. First, the single-institution sample limits generalizability to eastern Chinese universities; future research should include central-western institutions to account for regional disparities in digital infrastructure. This geographic limitation means the observed power of our pedagogical scaffolding must be tested in regions with potentially more severe infrastructural constraints, where its efficacy might be diminished. Second, DeepSeek’s Western-centric training data may not represent other AI tools (e.g., ChatGPT or ERNIE); replication with diverse, and especially multilingual, models is needed. Consequently, our findings regarding cultural bias and the need for critical evaluation may be more pronounced when using Western-trained models compared to those trained on more balanced or Sinocentric datasets. Finally, the focus on written translation in technical and legal domains leaves its efficacy in literary or interpreting specializations an open question. The observed gains in cultural adaptation may not translate directly to these more creative and real-time domains, where AI’s current limitations are more acute.

These limitations define clear pathways for future work. Crucially, longitudinal studies (1+ years) could explore AI’s impact on professional translation outcomes, such as employer ratings of graduates’ competence, moving beyond academic metrics to assess real-world efficacy.

Ultimately, this research argues for a critically optimistic approach: generative AI can revitalize translator education, but only if its integration is guided by a steadfast commitment to scalable equity, positioning pedagogical scaffolding as the non-negotiable prerequisite for ethical and effective implementation.

Footnotes

None.

ORCID iD

Zhong Lin

Author Contributions

Yiding Guo conceptualized,curated data,and drafted the manuscript;Zhong Lin revised the manuscript.

Funding

The authors received no financial support for the research,authorship,and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

Data Availability Statement

Data is available if requested.

References

Abou-Khalil

Flanagan

Ogata

(2021). Personal vocabulary recommendation to support real life needs. In Roll

McNamara

Sosnovsky

Luckin

Dimitrova

(Eds.), Lecture notes in computer science (pp. 18–23). Springer International Publishing.

Afzal

Khan

Daud

Ahmad

Butt

(2023). Addressing the digital divide: Access and use of technology in education. Journal of Social Sciences Review, 3(2), 883–895. https://doi.org/10.54183/jssr.v3i2.326

Al-Obaydi

L. H.

Pikhart

(2020). The relationship between artificial intelligence and reducing the cognitive load of learners. Zanco Journal of Humanity Sciences, 24(5), 1–10. https://doi.org/10.21271/zjhs.24.s5.34

Bearman

Tai

Dawson

Boud

Ajjawi

(2024). Developing evaluative judgement for a time of generative artificial intelligence. Assessment & Evaluation in Higher Education, 49(6), 893–905. https://doi.org/10.1080/02602938.2024.2335321

Pek

L. S.

Cong

Tiannan

Krishnasamy

H. N.

Ne’matullah

K. F.

Arar

(2025). Transforming translation education: A bibliometric analysis of artificial intelligence’s role in fostering sustainable development. International Journal of Learning, Teaching and Educational Research, 24(3), 166–190. https://doi.org/10.26803/ijlter.24.3.9

Braun

Clarke

(2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa

Clark

Gorski

(2001). Multicultural education and the digital divide: Focus on race, language, socioeconomic class, sex, and disability. Multicultural Perspectives, 3(3), 39–44. https://doi.org/10.1207/s15327892mcp0303_7

Cheng

(2024). Artificial intelligence and language education: A state-of-the-art review. Language Teaching, 57(2), 161–189. https://doi.org/10.1017/S0261444823000295

Chung

T. T.

Crosthwaite

Cao

C. T. H.

de Carvalho

C. T.

(2024). Walking the walk? (Mis)alignment of EFL teachers’ self-reported corpus literacy skills and their competence in planning and implementing corpus-based language pedagogy. TESOL Quarterly, 58(3), 1046–1080. https://doi.org/10.1002/tesq.3299

10.

Crosthwaite

Baisa

(2023). Generative AI and the end of corpus-assisted data driven learning? Not so fast! Applied Corpus Linguistics, 3(3), Article 100066. https://doi.org/10.1016/j.acorp.2023.100066

11.

Crosthwaite

Steeples

(2022). Pedagogy and equity in the age of generative AI: Addressing the digital divide in language education. ReCALL, 34(3), 265–281. https://doi.org/10.1017/S095834402200011X

12.

Díaz

Nussbaum

(2024). Artificial intelligence for teaching and learning in schools: The need for pedagogical intelligence. Computers & Education, 217, Article 105071. https://doi.org/10.1016/j.compedu.2024.105071

13.

Dung

(2024). A study on how globalization influences translation practices and the demand for translators in various industries. International Journal of Language and Linguistics, 12(5), 171–181. https://doi.org/10.11648/j.ijll.20241205.11

14.

Farroni

(2024). Korpora im DaF-Unterricht: Einsatz im Schulbereich. Linguistik Online, 126(2), 83–98. https://doi.org/10.13092/lo.126.11044

15.

Franzoni

(2023). From black box to glass box: Advancing transparency in artificial intelligence systems for ethical and trustworthy AI. In Esposito

(Ed.), Advances in artificial intelligence (LNCS Vol. 14123, pp. 118–130). Springer.

16.

Gonzales

A. L.

McCrory Calarco

Lynch

(2018). Technology problems and student achievement gaps: A validation and extension of the technology maintenance construct. Communication Research, 47(5), 750–770. https://doi.org/10.1177/0093650218796366

17.

Chen

Yang

C.-H. H.

Zhang

Chen

Chng

E. S.

(2024, February 10). GenTranslate: Large language models are generative multilingual speech and machine translators. arXiv, ArXiv:2402.06894. https://arxiv.org/abs/2402.06894

18.

Hwang

G.-J.

Chang

C.-Y.

(2023). A review of opportunities and challenges of AI in language education. Interactive Learning Environments, 31(9), 5599–5616. https://doi.org/10.1080/10494820.2021.2023436

19.

Kim

(2019). A critical literature review on the data-driven language learning approach. The Journal of Linguistics Science, 90, 23–45. https://doi.org/10.21296/jls.2019.9.90.23

20.

Knox

(2020). Artificial intelligence and education in China. Learning, Media and Technology, 45(3), 298–311. https://doi.org/10.1080/17439884.2020.1754236

21.

Kshetri

(2024). Linguistic challenges in generative artificial intelligence: Implications for low-resource languages in the developing world. Journal of Global Information Technology Management, 27(2), 95–99. https://doi.org/10.1080/1097198X.2024.2341496

22.

Kwan

R. Y. C.

Lai

Wong

Chan

(2023). The effects of Digital Buddy programme on older adults’ mental well-being: Study protocol for a multi-centre, cluster randomized controlled trial. Trials, 24(1), Article 95 https://doi.org/10.1186/s13063-023-07130-5

23.

Lampou

(2023). The integration of artificial intelligence in education: Opportunities and challenges. Review of Artificial Intelligence in Education, 4, Article e015. https://doi.org/10.37497/rev.artif.intell.educ.v4i00.15

24.

Lewis

(2014). Learner autonomy and the theory of sociality. In Murray

(Ed.), Social dimensions of autonomy in language learning (pp. 37–59). Palgrave Macmillan.

25.

(2024). Cognitive load and translation accuracy in technology assisted simultaneous interpreting enabled by artificial intelligence. English Language Teaching and Linguistics Studies, 6(3), Article 23. https://doi.org/10.22158/eltls.v6n3p23

26.

Lusta

Demirel

Ö.

Mohammadzadeh

(2023). Language corpus and data driven learning (DDL) in language classrooms: A systematic review. Heliyon, 9(12), Article e22731. https://doi.org/10.1016/j.heliyon.2023.e22731

27.

McGuire

(2019). Toward written error correction with a Japanese-English parallel corpus: Data-driven learning in the Japanese EFL classroom. Vocabulary Learning and Instruction, 8(1), 17–28. https://doi.org/10.7820/vli.v08.1.mcguire

28.

Mizumoto

(2023). Data-driven learning meets generative AI: Introducing the framework of metacognitive resource use. Applied Corpus Linguistics, 3(3), Article 100074. https://doi.org/10.1016/j.acorp.2023.100074

29.

Moreno

Park

(2010). Cognitive load theory: Historical development and relation to other theories. In Plass

J. L.

Moreno

Brünken

(Eds.), Cognitive load theory (pp. 9–28). Cambridge University Press.

30.

Pack

Maloney

(2024). Using artificial intelligence in TESOL: Some ethical and pedagogical considerations. TESOL Quarterly, 58(2), 1007–1018. https://doi.org/10.1002/tesq.3320

31.

PACTE Group. (2003). Building a translation competence model. In Alves

(Ed.), Triangulating translation: Perspectives in process oriented research (pp. 43–66). John Benjamins.

32.

Prinsloo

Khalil

(2024). The democratisation of AI in, for and through education: A position paper. Ubiquity Proceedings, 8, 1–14. https://doi.org/10.5334/uproc.130

33.

Rotgans

J. I.

Schmidt

H. G.

(2010). The motivated strategies for learning questionnaire: A measure for students’ general motivational beliefs and learning strategies? The Asia-Pacific Education Researcher, 19(2), 463–486. https://doi.org/10.3860/taper.v19i2.1603

34.

Singh

(2024). Diverse yet biased: Towards mitigating biases in generative AI [Student abstract]. Proceedings of the AAAI Conference on Artificial Intelligence, 38(21), 23653–23654. https://doi.org/10.1609/aaai.v38i21.30512

35.

Song

(2023). Disparity in educational resources between urban and rural areas in China. Journal of Advanced Research in Education, 2(5), 64–69. https://doi.org/10.56397/jare.2023.09.06

36.

Spaulding

(2020). Artificial intelligence and machine learning as business tools: A framework for diagnosing value destruction potential. Business Horizons, 63(2), 183–193.

37.

Sweller

(2010). Cognitive load theory: Recent theoretical advances. In Plass

J. L.

Moreno

Brünken

(Eds.), Cognitive load theory (pp. 29–47). Cambridge University Press.

38.

Talgatov

Kassymova

G. K.

Nurtanto

(2024). AI in the classroom: A boon or a threat to pedagogical practices? Journal of Educational Computing Research. Advance online publication. https://doi.org/10.1177/07356331231226345

39.

Tan

(2015). Teacher-directed and learner-engaged: Exploring a Confucian conception of education. Ethics and Education, 10(3), 302–312. https://doi.org/10.1080/17449642.2015.1101229

40.

Tenório

Bittencourt

I. I.

Silva

A. P.

(2023). Epistemic equity in AI-enhanced learning environments. Computers & Education, 190, Article 104623. https://doi.org/10.1016/j.compedu.2023.104623

41.

Vandenberg

Boulden

Catete

Min

Mott

(2023). Exploring the AI experiences of rural students and teachers: A sociocultural perspective. Journal of Research on Technology in Education, 55(5), 771–789. https://doi.org/10.1080/15391523.2023.2172447

42.

Walter

(2024). Embracing the future of artificial intelligence in the classroom: The relevance of AI literacy, prompt engineering, and critical thinking in modern education. International Journal of Educational Technology in Higher Education, 21(1), 1–25. https://doi.org/10.1186/s41239-024-00448-3

43.

Woo

D. J.

Wang

Yung

Guo

(2024). Effects of a prompt engineering intervention on undergraduate students’ AI self-efficacy, AI knowledge and prompt engineering ability: A mixed methods study. Computers in Human Behavior, 159, Article 108355. https://doi.org/10.1016/j.chb.2024.108355

44.

Yang

Wang

(2023). Cultural bias in machine translation: A case study of Chinese-English translation. Journal of Language and Cultural Education, 11(1), 45–62. https://doi.org/10.2478/jolace-2023-0004

45.

You

(2018). The seeming ‘round trip’ of learner-centred education: A ‘best practice’ derived from China’s New Curriculum Reform? Comparative Education, 55(1), 97–115. https://doi.org/10.1080/03050068.2018.1541662

46.

Zábojník

Hromada

(2024). The role of generative AI in empowering generation Z in higher education. Media & Marketing Identity, (2024), 758–776. https://doi.org/10.34135/mmidentity-2024-75

47.

Zacamy

Roschelle

(2022). Navigating the tensions: How could equity-relevant research also be agile, open, and scalable?Digital Promise.

48.

Zhai

Wibowo

L. D.

(2024). The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: A systematic review. Smart Learning Environments, 11(1), 1–18. https://doi.org/10.1186/s40561-024-00316-7

49.

Zhang

Tao

Huang

Lin

Chen

Feng

(2024). MC2: Towards transparent and culturally-aware NLP for minority languages in China. Transactions of the Association for Computational Linguistics, 12, 883–905. https://doi.org/10.1162/tacl_a_00679

50.

Zhang

(2021). Digital divide and educational equity in rural China: A multilevel analysis. Educational Technology Research and Development, 69(3), 1453–1471. https://doi.org/10.1007/s11423-021-09998-z

51.

Zhao

P.-J.

(2024). Reconceptualizing the link between validity and translation in qualitative research: Extending the conversation beyond equivalence. International Journal of Qualitative Methods, 23, 1–11. https://doi.org/10.1177/16094069241260134

52.

Zhou

Hou

(2024). Can AI empower L2 education? Exploring its influence on the behavioural, cognitive and emotional engagement of EFL teachers and language learners. Computer Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2024.2314261

Generative AI as a Catalyst for Data-Driven Learning: Efficacy,Equity,and Engagement in Translation Education

Abstract

Plain Language Summary

Keywords

Introduction

Literature Review

DDL and Translatorial Competence

Generative AI in Translator Education

Equity in AI-Enhanced Translator Education in China

Gap Synthesis and Present Study

Methodology

Participants

Intervention

Data Collection

Data Analysis

Ethical Considerations

Results

Efficacy: AI-Enhanced DDL Fosters Critical Triangulation and Reveals Domain-Specific Limits

Equity: The Rural-Urban Divide is a Critical Mediator of Efficacy

Engagement: Learners Report High Motivation Amidst Critical Concerns

Discussion

Transformed Competence Tempered by Instrumental Dependence

Democratization Tempered by Digital Divides and Cultural Bias

Heightened Motivation Juxtaposed with Critical Ambivalence

Toward Pedagogically Scaffolded and Equitable Integration

Conclusion

Footnotes

ORCID iD

Author Contributions

Funding

Declaration of Conflicting Interests

Data Availability Statement

References