Digital translation practices in university English learning: adoption, disciplinary variation, and AI-assisted revision

1 Introduction

Digital translation technologies have expanded the resources available to university students for reading, writing, and revising texts in English. Online machine translation provides readily accessible linguistic support, including outside formal instruction. However, its pedagogical value depends on how learners evaluate and use the output in particular tasks. Niño (2020) investigated online machine translation in independent language learning, highlighting learners’ judgments about its usefulness and its relationship with other language resources. From this perspective, learners’ choices are central: access to a translation tool does not establish that its use leads to language development.

Generative artificial intelligence (AI) extends this discussion because large language models can support translation, explanation, text generation, and correction within a single interaction. These capabilities blur the boundary between linguistic support and the delegation of writing. Kasneci et al. (2023) discuss the educational opportunities and limitations of large language models, emphasizing critical evaluation and human oversight. For university English education, this requires distinguishing which tools students use, how they engage with the output, and what evidence would be needed to demonstrate learning.

Research on writing illustrates the importance of these distinctions. Lee and Briggs (2021) examined revisions made by 58 Korean university students through comparisons between their original texts and machine-translation output. Errors decreased in most categories after revision, and students who initially made fewer errors tended to correct a greater proportion of them. These findings relate the usefulness of technological support to learners’ existing language resources and revision activity. However, improvement in an assisted text is a different outcome from subsequent independent language performance. A survey of reported practices therefore addresses a different question from an intervention assessing writing quality or language development.

The present study examines translation practices as a heterogeneous set of activities. It distinguishes general digital-tool use, tool preferences elicited by different questionnaire items, print-dictionary experience and endorsement, and AI-assisted revision of students’ own texts. We use the term translation ecology to describe the coexistence of these resources and practices within a shared educational setting. This is an organizing concept, not a validated construct or a claim that individual tool-switching sequences were observed. Technology acceptance, task–technology fit, and self-regulated learning provide complementary interpretive perspectives.

1.1 Technology acceptance and the distinction between endorsement and use

The Technology Acceptance Model (TAM) identifies perceived usefulness and perceived ease of use as central to technology acceptance (Davis, 1989). For translation resources, this perspective draws attention to the practical value users perceive and the effort involved in use. It motivates examining adoption without assuming that the most widely used resource produces the strongest educational outcomes

Li et al. (2024) tested an extended acceptance model incorporating computer self-efficacy and perceived enjoyment among 370 Chinese university students with experience of translation technologies. Their findings linked ease of use and enjoyment to attitudes and linked attitudes to behavioral intentions. That study provides a theoretical reference point; the present survey did not measure these constructs with validated scales or estimate an acceptance model.

For RQ1, the acceptance perspective informs the distinction between reported use and endorsement. The questionnaire separately records print-dictionary experience, frequency of use, and support for continued promotion. These indicators describe behavior and endorsement; they do not directly measure perceived usefulness, perceived ease of use, or behavioral intention. A student may support print dictionaries for educational purposes while relying on digital tools in everyday tasks.

1.2 Task–technology fit and contextual variation

Task–technology fit offers a perspective on why tool choices may depend on the activity. Goodhue and Thompson (1995) emphasize that technology use and the correspondence between technological capabilities and task requirements both matter for performance. Applied here, the framework directs attention to the relationship between the support sought by students and the functions associated with a resource.

For RQ2, the questionnaire provides tool-preference distributions under general-use, application/website, and AI-focused wording. These are question contexts rather than observed tasks or direct measures of fit. Differences between distributions may suggest contextual variation in how respondents identify tools, but cannot establish tool-switching over time or superior performance on a controlled assignment.

For RQ3, faculty membership provides an observed grouping variable through which to explore variation in reported practices. Students in different programs may encounter different disciplinary genres, vocabulary, writing demands, and expectations for English use. These are plausible contextual explanations requiring direct measurement; they are not mechanisms established by faculty membership. RQ4 likewise examines associations with measured student characteristics and reported practices rather than testing task–technology fit.

1.3 Self-regulated learning and AI-assisted revision

Self-regulated learning concerns how students direct, monitor, and evaluate their learning (Zimmerman, 2002). It offers a perspective on the quality of engagement with translation and revision tools. The questionnaire records whether AI-assisted revision occurred, as considered in RQ1, RQ3, and RQ4, but does not measure goal setting, monitoring, verification, or reflection. This framework therefore identifies what remains unknown about reported revision rather than providing an explanation tested by the survey.

AI-assisted revision may involve checking a word or grammatical choice, comparing alternative formulations, or requesting substantial rewriting. These activities imply different degrees of learner agency, linguistic processing, and delegation. A single affirmative response cannot distinguish them or establish that the learner critically evaluated the output. The operational definition below retains this behavioral scope.

Together, these perspectives distinguish adoption and endorsement, the context of tool selection, and the quality of learner engagement. The first two inform the organization of the measured indicators; the third identifies processes that require additional measurement. None is treated as a validated latent-variable model in this study.

1.4 Research gap and the present study

The literature considered above addresses related but distinct questions: independent use of translation resources, revision of written products, acceptance of translation technologies, and the educational implications of generative AI. These contributions do not establish whether adoption, tool preferences, print-resource use, and AI-assisted revision follow similar patterns across faculties within the same university. Examining these indicators together can provide an institutional account that complements studies focused on a particular platform, writing task, or specialist student population.

The present study contributes an account of this translation ecology through a cross-sectional survey of 435 students from 37 degree programs and seven faculties at a public university in southern Peru. Its contribution extends beyond documenting high adoption: it examines whether widespread digital-tool use coexists with context-dependent tool preferences, differing levels of AI-assisted revision across faculties, and continued endorsement of print resources. Bringing these indicators together within one institution helps distinguish access and reported use from forms of engagement that remain to be investigated. The Peruvian setting contextualizes the evidence without establishing representativeness beyond the surveyed students.

Four research questions guide the analysis: These research questions define a descriptive and exploratory investigation. The theoretical perspectives organize interpretation but are not tested as latent-variable models because the questionnaire did not operationalize their constructs using validated scales. The translation-ecology perspective connects the measured indicators while preserving the distinction between reported use and demonstrated learning.

• RQ1. How prevalent are translation application or website use, AI or automated-translator use, AI-assisted revision, and print-dictionary experience, and how widely is the continued promotion of print dictionaries supported?
• RQ2. How do reported tool preferences differ across general-use, application/website, and AI-focused questionnaire items?
• RQ3. How do translation-related practices vary across faculties?
• RQ4. Which measured student characteristics and reported practices are associated with AI-assisted revision of students’ own English writing or translations?

2 Materials and methods

2.1 Study design and setting

This observational cross-sectional survey was conducted at Universidad Nacional Jorge Basadre Grohmann, a public university in Tacna, Peru, during the 2025 academic year. Reporting was guided by the STROBE recommendations for cross-sectional studies (von Elm et al., 2007). The source project reported an enrolled student population of 9,032. The analytical dataset comprised 435 questionnaires from students in all 37 degree programs and seven faculties.

All 9,032 enrolled students were invited to participate. The 435 respondents represented a response rate of 4.8% (435/9,032). Participation was voluntary and no incentives were offered, resulting in a nonprobability sample based on self-selection. The university-wide invitation and coverage of all faculties and degree programs do not establish population representativeness. Confidence intervals describe uncertainty under the statistical assumptions used; they do not correct for selection or nonresponse bias.

2.2 Participants and survey administration

The 435 respondents were university students aged 18 years or older who participated voluntarily and completed the questionnaire. Recorded characteristics were age, gender, year of study, degree program, faculty, and language-center enrollment. Valid denominators for questionnaire items ranged from 382 to 434 because of missing responses or conditional questions. No record was removed solely because one or more item responses were missing. Table 1 defines the faculty abbreviations used throughout the manuscript.

Table 1. Faculty names and abbreviations.

Abbreviations retain the institution's Spanish faculty acronyms; English translations are provided for readability.

2.3 Questionnaire and operational definitions

The study-specific questionnaire contained 20 questions organized into four topic domains: general use of translation aids in English learning; translation applications and websites; print dictionaries; and AI or automated translation. Response formats included binary, nominal, ordinal, multiple-response, and open-ended items. The domains organized the questionnaire content and were not treated as psychometric subscales. No total score was produced, and no global internal-consistency coefficient was interpreted because the items represented heterogeneous behaviors and preferences rather than interchangeable indicators of a single construct.

Throughout the manuscript, machine translation refers to automated translation between languages, whereas generative AI refers to systems that can generate or reformulate text in response to prompts. The questionnaire category “AI or automated translators” combines these possibilities and cannot isolate generative-AI use. “Translation applications or websites” describes a mode of access rather than a distinct technological class. These two indicators are not mutually exclusive in interpretation: a tool such as Google Translate may be reported under both, and a respondent may answer yes to both items. The original item labels are retained when reporting results, but their percentages must not be added to estimate unique users.

Before administration, the questionnaire was reviewed by six experts, all of whom held doctorates in Educational Sciences.

Five indicators were selected for comparison: (1) use of translation applications or websites, (2) use of AI or automated translators, (3) AI-assisted revision of one's own English writing or translations, (4) ever having used a print dictionary, and (5) support for continued promotion of traditional bilingual dictionaries in professional education. AI-assisted revision was defined as an affirmative response to item 19.1: “Have you used AI to correct your own writing or translations in English?” (English translation of the Spanish item). The category may encompass a lexical or grammatical check, comparison of alternatives, or substantial rewriting. The item did not distinguish the purpose or extent of revision, verification of suggestions, retention of student authorship, or degree of delegation. It therefore measures reported behavior, not learner agency, writing quality, English proficiency, or learning gains.

The questionnaire also included open-ended questions about perceived accuracy, benefits, disadvantages, learning autonomy, and skills needed for effective AI use. The retained dataset did not contain a sufficiently complete qualitative corpus: most fields were blank or contained placeholders, with only isolated substantive responses. Consequently, no thematic analysis was undertaken and no qualitative themes or participant quotations are reported.

2.4 Statistical analysis

Analyses used available cases for each indicator, with denominators reported for each analysis. Categorical variables were summarized as frequencies and valid percentages. Key prevalence estimates are accompanied by 95% Wilson confidence intervals. For multiple-response items, percentages indicate the proportion of valid respondents selecting each option and need not sum to 100%.

Faculty-specific percentages and 95% Wilson confidence intervals are descriptive, especially where samples are small. The omnibus faculty tests assess the overall association and do not establish differences between particular faculty pairs. These intervals are unadjusted for multiple comparisons, and neither they nor the contingency-table tests adjust for dependence within degree programs.

Paired exact McNemar tests compared reported use of AI or automated translators, translation applications or websites, and print dictionaries within respondents. Matched-pair odds ratios and 95% confidence intervals summarized these contrasts. Associations between categorical variables were assessed using chi-square or Fisher's exact tests, as appropriate. Faculty comparisons for the five indicators used Monte Carlo p values based on 200,000 simulated tables with fixed margins. The Benjamini–Hochberg procedure controlled the false discovery rate within families of exploratory tests (Benjamini and Hochberg, 1995). Association strength was summarized using bias-corrected Cramér's V.

Binary logistic regression explored associations with AI-assisted revision of students’ own English writing or translations. Predictors were year of study, weekly translator-use time, faculty, use of AI or automated translators, language-center enrollment, gender, and print-dictionary experience. The model used complete cases (n = 410) and standard errors clustered by 37 degree programs. Adjusted odds ratios (aORs), 95% confidence intervals, McFadden's pseudo-R², and the area under the receiver operating characteristic curve (AUC) are reported. The AI or automated-translator use indicator is conceptually close to AI-assisted revision and is interpreted as a check on consistency between related responses rather than an independent explanatory determinant. The model is exploratory and no coefficient is interpreted causally.

Clustering by degree program allows for possible dependence among students exposed to shared curricular and instructional contexts, while assuming independence between programs. The 37 programs contained 1–31 respondents in the full sample (median = 6) and 1–29 in the complete-case model (median = 6); one program had a single respondent in both datasets. The cluster-robust covariance used the CR1 finite-sample multiplier G/(G-1) × (N-1)/(N-k), with G = 37, N = 410, and k = 17 estimated coefficients including the intercept. The primary intervals and p values use a normal reference distribution. This correction does not guarantee reliable inference with limited, unequal clusters (Cameron and Miller, 2015), and clustering does not correct for nonprobability recruitment.

Two sensitivity checks evaluated the regression results. First, a logistic model with a normally distributed random intercept for degree program used the same 410 cases and fixed-effect predictors. Maximum marginal likelihood was evaluated using 81-point Gauss-Hermite quadrature, and model-based Wald intervals used the observed information matrix. Second, the primary CR1 standard errors were retained while replacing the normal reference distribution with an approximate t distribution with 36 degrees of freedom (G-1). The latter assesses sensitivity to the reference distribution; it is not a guarantee of valid inference with unequal clusters.

3 Results

3.1 Participant characteristics

The 435 respondents represented all 37 degree programs and seven faculties. Valid responses showed that 54.6% identified as women and 45.4% as men. The sample was concentrated in the first year of study (46.2%) and in the youngest recorded age category (55.4%). Students in the second, third, fourth, and fifth years accounted for 19.0%, 25.8%, 4.2%, and 4.9%, respectively. Approximately one quarter (26.0%) reported studying at a language center (Table 2). The Faculty of Education, Communication and Humanities contributed 162 participants (37.2%), followed by Legal and Business Sciences (n = 74, 17.0%) and Engineering (n = 62, 14.3%). The remaining four faculties contributed between 15 and 46 participants each.

Table 2. Participant characteristics.

Valid n was 431 for year of study, 434 for gender, 433 for age, and 430 for language-center study. Original questionnaire age-category labels are retained. All participants were adults; the 17–19-year category therefore includes only respondents aged 18–19 years.

3.2 Prevalence and context of translation practices

Digital translation was nearly ubiquitous. Of 430 valid respondents, 407 (94.7%, 95% CI 92.1%–96.4%) reported using AI or automated translators. Translation applications or websites were used by 400 of 433 respondents (92.4%, 95% CI 89.5%–94.5%). AI-assisted revision of students’ own writing or translations was reported by 318 of 429 respondents (74.1%, 95% CI 69.8%–78.0%). By comparison, 282 of 431 respondents (65.4%, 95% CI 60.8%–69.8%) had used a print dictionary (Table 3; Figure 1).

Table 3. Prevalence and context of translation practices.

Context items are single-choice categories in the source questionnaire.

CI, Wilson confidence interval.

Within-person contrasts supported the predominance of digital tools. Use of AI or automated translators was more common than print-dictionary experience (matched-pair OR = 9.86, 95% CI 5.69–17.08, exact p < .001). Translation application or website use was also more common than print-dictionary experience (matched-pair OR = 7.88, 95% CI 4.76–13.06, exact p < .001). The contrast between the two digital-use indicators was smaller and not statistically significant (matched-pair OR = 1.56, 95% CI 0.86–2.81, p = .184).

Translation was reported most often when students needed to learn something new (42.8%) or complete academic tasks (39.5%). Most respondents reported less than one hour of translator use per week (69.5%); 26.5% reported one to three hours, and 4.0% reported more than three hours. In the multiple-response activity item, assignments (36.2%), academic texts (29.7%), and everyday phrases (28.7%) were the most frequent selections. These patterns describe brief, task-linked use, although the categories do not establish whether the tool replaced or supported independent language work.

3.3 Tool profiles across questionnaire contexts

Google Translate was the most frequently named tool in all three questionnaire contexts, accounting for 76.3% of valid responses to the general-use item, 71.4% to the application/website item, and 53.1% to the AI-focused item. ChatGPT accounted for 7.7% in the general-use item and 34.0% in the AI-focused item, while DeepL ranged from 5.1% to 7.9% across the three items (Figure 2). These are descriptive distributions elicited by differently worded questions with item-specific denominators. Because the categories overlap, they do not identify mutually exclusive user groups, demonstrate an individual transition between tools, or show that respondents consistently distinguished machine translation from generative AI.

3.4 Print-dictionary use and the attitude-practice gap

Although 65.4% had used a print dictionary, current use was infrequent: 49.5% reported using one rarely and 25.5% never, compared with 22.9% sometimes and 2.1% always. Nevertheless, 333 of 427 respondents (78.0%, 95% CI 73.8%–81.7%) supported continued promotion of traditional dictionaries in professional education. Support was higher among students who had used a print dictionary (82.8%) than among those who had not (68.3%; OR = 2.24, 95% CI 1.40–3.57, q = .007). Bilingual dictionaries were preferred by 71.5% of respondents who stated a preference, compared with 25.5% for monolingual dictionaries. Thus, endorsement remained high despite infrequent routine use; the questionnaire did not establish respondents’ reasons for that endorsement.

3.5 Faculty variation

Use of translation applications or websites and use of AI or automated translators were high across faculties, ranging from 86.7% to 100.0% and from 91.1% to 100.0%, respectively (Table 4; Figure 3). After false-discovery-rate adjustment, no statistically significant faculty association was detected for application or website use, AI or automated-translator use, print-dictionary experience, or support for promoting print dictionaries (all q ≥ .172). These results do not establish equivalence across faculties.

Table 4. Translation-related indicators by faculty.

Each indicator cell reports yes/valid n, followed by the percentage and its 95% Wilson confidence interval. The n column gives the total faculty sample. AI/MT use denotes the combined category AI or automated translators. Intervals are descriptive, unadjusted for clustering and multiple comparisons, and do not correct for nonprobability recruitment. Precision is limited, particularly for Sciences (n = 15). The omnibus association does not establish which faculty pairs differ.

AI-assisted revision showed a different pattern. It ranged from 52.2% in Civil Engineering, Architecture and Geotechnics to 96.8% in Engineering. The global faculty association was statistically significant (Monte Carlo test based on χ²(6) = 32.27, p < .001, q < .001) with a small-to-moderate effect (V = .274). Degree-program estimates were not tested inferentially because several programs had very small and unequal samples.

Faculty-specific estimates should be interpreted descriptively, particularly for Sciences (n = 15). The omnibus result does not support ranking faculties or asserting pairwise differences without appropriate additional analysis.

3.6 Bivariate associations

Eight exploratory associations remained statistically significant after false-discovery-rate correction (Table 5). The strongest was the expected relationship between ever using a print dictionary and reported frequency of print-dictionary use (V = .673). Year of study was associated with language-center enrollment (V = .309); reported enrollment ranged from 16.7% in the first year to 76.2% in the fifth year. Use of AI or automated translators was associated with AI-assisted revision (V = .268): 76.8% of those reporting such use reported revision, compared with 22.7% of those not reporting it (OR = 11.25, 95% CI 4.04–31.30). This association primarily confirms the logical proximity of the two survey items.

Table 5. Paired contrasts and exploratory bivariate associations.

Some analyses contained expected cell counts below five and should be interpreted with the effect sizes and confidence intervals. AI use denotes the combined item AI or automated translators; apps/web denotes translation applications or websites. The two digital-use indicators overlap.

q, Benjamini-Hochberg adjusted p value; V, bias-corrected Cramér's V; OR, odds ratio.

Smaller associations were found between print-dictionary frequency and support for promotion (V = .258), year of study and AI-assisted revision (V = .200), translation context and weekly use time (V = .150), print-dictionary experience and support for promotion (V = .159), and weekly use time and AI-assisted revision (V = .142). AI-assisted revision was reported by 69.4% of respondents reporting less than one hour per week and by 84.7% of those reporting one to three hours.

3.7 Exploratory model of AI-assisted revision

The complete-case logistic model included 410 respondents, with McFadden's pseudo-R² = .167 and AUC = .760. Use of AI or automated translators had the largest coefficient (aOR = 11.24, 95% CI 4.53–27.91, p < .001), consistent with its conceptual proximity to AI-assisted revision. Relative to the Faculty of Education, Communication and Humanities, Engineering had higher estimated adjusted odds (aOR = 8.57, 95% CI 1.29–57.10, p = .026). The wide interval indicates limited precision. Reporting one to three hours of translator use also had higher estimated odds than reporting less than one hour (aOR = 1.96, 95% CI 1.01–3.81, p = .046). The remaining coefficients in the primary model had p values above.05 (Table 6; Figure 4). These estimates are exploratory. With 37 unequal degree-program clusters, conventional cluster-robust intervals and p values may overstate precision; findings near conventional significance thresholds require particular caution.

Table 6. Primary logistic regression and random-intercept sensitivity model for AI-assisted revision.

Both models use the same 410 complete cases and 37 degree programs. aOR, adjusted odds ratio; CI, confidence interval. Primary-model intervals and p values use CR1 cluster-robust standard errors and a normal reference distribution; McFadden's pseudo-R² = .167 and AUC = .760. The sensitivity model includes a normally distributed program random intercept and uses model-based Wald intervals. Its random-intercept variance is 0.049 (latent-scale ICC = .015). FECH is the reference faculty. Estimates remain exploratory; agreement between models does not establish protection against all small-cluster biases.

The random-intercept model yielded similar point estimates for AI or automated-translator use, Engineering, and one to three hours of use, although interval widths differed (Table 6). Under the additional t(36) reference check, the one-to-three-hour association had a 95% CI of 0.99–3.90 and p = .054, compared with p = .046 under the primary normal approximation. The corresponding t-based intervals remained above 1 for Engineering (1.20–61.00; p = .033) and AI or automated-translator use (4.39–28.80; p < .001). Thus, the time-use association is sensitive to the inferential specification and should not be treated as a stable threshold-based finding.

4 Discussion

4.1 Principal findings

The findings describe a heterogeneous translation ecology within the surveyed university. Widespread digital-tool adoption coexisted with differing tool-preference profiles across questionnaire contexts, substantial faculty variation in AI-assisted revision, and continued endorsement of print dictionaries despite infrequent use. The central contribution is the distinction among these practices within a shared institutional setting, rather than the prevalence of any single technology.

Two distinctions organize this interpretation. First, frequent tool use does not establish either benefit or harm to English learning: the survey measured reported behaviors and preferences, not proficiency, writing quality, retention, or change over time. Second, translation technology does not represent a uniform learner experience. Google Translate, ChatGPT, DeepL, and print dictionaries occupied different positions in the response profiles. The translation-ecology perspective makes this heterogeneity visible, while the survey cannot determine the sequence in which individual learners combined resources or how they evaluated the output.

4.2 Digital translation as an ordinary learning practice

The widespread use of digital translation is consistent with the literature reviewed by Deng and Yu (2022) and Jolley and Maimone (2022). For teaching, this raises the practical question of how to design activities when students can readily access translation and generative AI. Considered alongside previous research, the present findings motivate making assisted work visible through prompts, source texts, comparisons of alternatives, and explanations of revisions. These are pedagogical implications to be evaluated, not effects demonstrated by this survey.

The reported duration and contexts of use are consistent with brief, task-related engagement, although actual tool-use episodes were not observed. Such engagement may provide opportunities for linguistic comparison and revision, as examined by Lee (2020) and Lee and Briggs (2021). Accepting a complete replacement text without evaluating changes may involve a different degree of linguistic processing. The present survey cannot distinguish these pathways or establish whether either produced learning.

The tool-preference distributions also depended on questionnaire wording. Google Translate remained prominent, while ChatGPT had a larger share in the AI-focused item. One possible interpretation is that respondents associated these tools with different functions; however, differences in wording, overlapping categories, and item denominators also limit that interpretation. Research on generative AI in EFL writing (Yang and Lin, 2025) and AI-powered machine translation in L2 writing (Yu and Jiang, 2025) similarly draws attention to varied forms of language mediation. These distinct technologies should not be collapsed into a single learning process. For assessment, a related implication is to distinguish permitted assistance during practice from tasks intended to assess independent performance.

4.3 AI-assisted revision and disciplinary variation

AI-assisted revision varied more substantially across faculties than either general digital-use indicator. This suggests that widespread adoption alone does not describe how students engage with tools during revision. Faculty membership may reflect differences in disciplinary genres, assessment practices, the amount and nature of English-language production, teacher guidance, perceived legitimacy of AI assistance, and prior AI instruction. None of these potential explanations was measured. The higher point estimate for Engineering must therefore be interpreted in light of unequal group sizes, sparse nonuse, and substantial statistical uncertainty, without attributing the pattern to a causal disciplinary effect.

The association between use of AI or automated translators and AI-assisted revision was expected because the items refer to related behaviors. It offers limited evidence of consistency between responses, not a new independent explanation of revision or proof of measurement validity. A positive revision response can cover anything from checking a lexical choice to accepting an extensively rewritten text. Future instruments should distinguish purpose, verification, the extent of student involvement, and the degree to which AI replaces student work. Those distinctions would help identify whether similar reported use rates conceal different forms of learner agency.

Pedagogically, the faculty pattern motivates examining local writing and assessment practices before designing support. Future studies could record the genres and volume of English writing students complete, whether tasks are formative or summative, the guidance teachers provide, and students’ understanding of permitted assistance. These measures could be linked to revision logs and course-level observations to test contextual explanations directly. Acceptance-related constructs examined by Li et al. (2024) could also be assessed using appropriate measures, but they cannot explain the present pattern without such data. A multilevel design would help distinguish student, course, and degree-program variation.

4.4 The continuing role of print dictionaries

Print-dictionary responses indicate an attitude-practice gap: routine use was infrequent, whereas approximately four in five respondents supported continued promotion. This need not be contradictory, because endorsement of a resource and its use in daily tasks are different responses. The survey did not ask why students endorsed print dictionaries, so explanations involving trust, time pressure, or perceived learning value remain hypotheses. The preference for bilingual dictionaries likewise describes a reported choice without establishing its relation to proficiency or learning.

An evidence-informed implication is to teach dictionary consultation and evaluation across formats. Relevant practices include selecting the appropriate sense, checking grammatical information and usage examples, and rejecting translations that do not fit the context. These practices could also support evaluation of machine-generated output, but that transfer was not tested here. Work on pedagogical translation supports structured comparison and reflection rather than treating the mere presence of translation as evidence of learning (Asquerino Egoscozábal and Estrada Chichón, 2024; Pintado Gutiérrez, 2022).

4.5 Implications for university teaching and policy

The following recommendations are evidence-informed pedagogical implications drawn from the present descriptive findings in conjunction with previous research; their effectiveness was not tested in this study. First, universities should specify permitted assistance in relation to assessment objectives, including when independent English production is required. Second, courses could teach verification and post-editing routines that involve checking meaning, terminology, grammar, and the reasons for accepting or rejecting suggestions. Third, assignments could make the revision process visible through annotated changes, version histories, or brief explanations. Fourth, guidance should take account of disciplinary genres and local assessment practices, informed by direct consultation with teachers and students.

These proposals draw on research concerning mediated revision and post-editing (Lee, 2020; Lee and Briggs, 2021; Omar and Salih, 2024), as well as concerns about dependency, authorship, and institutional expectations reported by academic staff and students (Groves and Mundt, 2021; Zhang et al., 2025). The intended educational aim is informed language judgment and opportunities for independent performance. Whether the proposed practices achieve that aim requires evaluation using process evidence and objective learning outcomes.

4.6 Strengths and limitations

The study includes students from 37 degree programs, reports item-specific denominators, uses paired comparisons for within-participant contrasts, adjusts families of exploratory tests for multiple comparisons, and presents effect sizes and interval estimates. Treating heterogeneous questionnaire items as separate indicators avoids an unsupported composite scale and keeps the claims aligned with the instrument.

Several limitations constrain interpretation. The nonprobability sample and low response rate limit generalizability and leave the magnitude and direction of nonresponse bias uncertain. First-year students and the Faculty of Education, Communication and Humanities accounted for large shares of the sample, limiting comparisons across groups. Self-report is subject to recall, interpretation, and social-desirability biases. The wording partially conflates machine translation and generative AI, and the same tool may contribute to more than one digital-use indicator. Cross-sectional data cannot establish temporal order or learning effects. The AI-assisted revision item does not measure the extent or quality of revision, verification, learner agency, or subsequent learning. The regression includes a predictor conceptually close to the outcome, and sparse outcomes contribute to wide confidence intervals. Course-level practices, objective English proficiency, digital access, prior AI instruction, and understanding of permitted assistance were not measured. Finally, the retained open-ended responses were insufficient for a qualitative analysis of the reasons underlying these patterns.

No separate pilot with students was documented, limiting assessment of item comprehensibility beyond the expert review. Small and unequal faculty samples limit the precision of descriptive estimates, and limited, unequal degree-program clusters can compromise conventional cluster-robust inference. The random-intercept sensitivity model produced similar main point estimates but relies on additional distributional assumptions. The association with one to three hours of use was sensitive to the reference distribution used for inference. These checks do not remove uncertainty from sparse data, selection bias, or the exploratory specification.

Future research should link documented tool use to objective English outcomes and indicators of the revision process. Longitudinal studies could examine changes in practices over the course of university study. Randomized or quasi-experimental designs could compare unassisted writing, direct machine translation, and guided AI post-editing using blinded assessments of writing and subsequent independent performance. Revised questionnaires should distinguish machine translation from generative AI, use coherent validated scales where latent constructs are intended, and directly assess verification and post-editing behavior. Interviews or think-aloud methods could clarify how students identify problems, evaluate alternatives, and decide whether to accept suggested changes.

5 Conclusion

The surveyed students reported a heterogeneous translation ecology characterized by widespread digital-tool use, context-dependent tool preferences, substantial variation in AI-assisted revision across faculties, and continued endorsement of print dictionaries. The contribution lies in showing that high adoption can coexist with different reported practices within the same institution. The findings describe use and its distribution; they do not demonstrate improved English proficiency, better independent writing, or effective self-regulation. Considered alongside previous research, they motivate pedagogical attention to verification, post-editing, and alignment between permitted assistance and assessment objectives. These implications require direct evaluation, while the observed disciplinary variation calls for research on genres, teaching guidance, assessment practices, and learner agency.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Noluntu Mpekoa, University of South Africa, South Africa

Reviewed by: Dana Leon-Henri, University of Franche-Comté, France

Dragana Bozic Lenard, Josip Juraj Strossmayer University of Osijek, Croatia