AI in education: boosting productivity or challenging cognitive effort?

Artificial intelligence (AI) has rapidly transformed higher education, reshaping administration, instruction and learning (Chen et al., 2020) in ways that were unimaginable a decade ago. Early research established that AI enables more effective grading, personalised content delivery and improved learner retention, with Hannan and Liu (2023) identifying student learning experience, student support and enrolment management as the primary operational areas in which institutions gained competitive advantage through AI adoption. As generative AI tools proliferated, the scope and pace of integration accelerated sharply. Crompton and Burke (2023) identified five dominant applications: assessment and evaluation, prediction, AI assistant functions, intelligent tutoring systems and managing student learning. Bond et al. (2024) confirmed that adaptive systems and personalisation dominate the field. On the opportunities side, Rahman and Watanobe (2023) demonstrated that tools such as ChatGPT support personalised feedback, interactive learning and accessibility for non-native speakers. These findings are further supported by Jaboob et al. (2026), who confirmed that personalised learning, automated feedback and grading and enhanced accessibility are among the most consistently documented benefits of AI integration in higher education.

Recent studies have shown that AI can improve academic performance, reduce time spent on assignments and streamline the writing process by providing structural, linguistic and content feedback (Bhullar et al., 2024; Dai et al., 2023; Wang, 2025). Specifically, Wang (2025) found that students achieved better writing quality with the use of AI, regardless of whether they were writing in their native language or another language. Firdaus et al. (2025) and Bećirović et al. (2025) emphasised that the use of AI correlates with better grades and greater academic self-efficacy. These findings are consistent with broader trends suggesting that AI helps to streamline learning by personalising the process and reducing the need to perform repetitive tasks (Jian, 2023).

In Asian countries, considerable attention is being paid to linking education and technology, particularly in Southeast Asia, where more than half of the population is under 30 years of age and highly tech-savvy (Sedik, 2018). The rapid advancement of AI has transformed not only students’ technical skills but also their cognitive and social development (Qolamani and Mohammed, 2023). AI-based platforms, particularly tools to support language learning and writing, have been shown to improve learning efficiency and academic outcomes among diverse groups of students, including those for whom English is not their first language (Wang, 2025; Micheni et al., 2024). Personalised AI-based learning environments offer tailored feedback, improve comprehension and increase student engagement (Jian, 2023). This study contributes contextual value by examining these dynamics at the student level in Vietnamese higher education, where the integration of AI is already changing the institutional and pedagogical practice (e.g. Furze et al., 2024).

However, experts warn of the potential negative impacts of using AI, particularly in relation to critical thinking. They fear that heavy reliance on AI tools may weaken students’ critical thinking, self-control and problem-solving skills (Tamrin et al., 2024; Chan and Hu, 2023). As Zhang et al. (2024) noted, excessive personalisation of AI may limit intellectual growth if students use it primarily to supplement their already acquired knowledge instead of using it to discover new areas of interest. These concerns have sparked discussions about whether AI supports learning and knowledge acquisition or merely facilitates task completion. These concerns thus raise questions about the cognitive effort students invest when using AI. Cognitive effort might be referred to as mental resources individuals allocate and actively use while performing a task (Chen et al., 2025). Nasr et al. (2026) noted that students may use generative AI to complete academic tasks while avoiding the cognitive effort required for critical thinking. Critical thinking is conceptually distinct. In the context of generative AI use, it is reflected in activities such as setting clear goals, refining prompts, assessing AI-generated content against specific criteria and standards and verifying outputs against external sources of one’s own expertise (Lee et al., 2025). The two concepts are thus related because critical thinking activities require cognitive effort.

Melweth et al. (2023) emphasised the need for AI adoption that minimises the potential threats of AI dependency, and suggest teamwork, digital literacy and critical thinking requiring tasks as key controls of AI adoption process. Similarly, Wang (2025) pointed out that students who actively work with AI – i.e. re-evaluate, criticise and revise the generated content – achieve better results compared to those who only passively copy its outputs without further modification. Passive use of AI is associated with reduced problem-solving skills and reduced cognitive performance (Zhai et al., 2024) and evolving dependency on AI-generated output.

Despite a growing number of studies on effects of AI adoption among university students, scholarship has focused either on productivity-related outcomes or on cognitive risks (Cotton et al., 2024). A study that would reconcile the AI-led productivity gains and risks – particularly self-perceived cognitive effort and dependency on AI tools, and more importantly, the assumption that the individual productivity experience of university students might be shaped not just by perceptions of gains but simultaneous individual perceptions of changes in cognitive effort or emerging dependency on AI tools appears to skip scholarly attention. Furthermore, the number of studies on AI-led productivity remains still limited or exploratory in certain aspect of AI adoption, particularly in Southeast Asia, where AI adoption is expanding rapidly (Pham et al., 2023). Research in Vietnam has, for example, already addressed related aspects of generative AI use, including course-level AI-assisted learning (e.g. Pham et al., 2023) and institution-level changes in assessment and pedagogical practice (e.g. Furze et al., 2024). In a course-level study at RMIT University, Pham et al. (2023) showed that ChatGPT has potential as an AI-assisted learning tool, while its responses may require refinement and verification. The authors therefore emphasised the importance of guiding students in the effective use of ChatGPT and in critically evaluating and verifying AI-generated answers. However, the relationships among productivity, dependency and perceived changes in critical thinking at the student level remain less directly examined.

To address this gap, this study conceptualises student productivity as a multidimensional construct encompassing time efficiency, academic performance, academic self-efficacy, changes in cognitive effort and AI dependency. Rather than forming a causal sequence, these dimensions are examined as simultaneous aspects of AI-assisted productivity. Students may simultaneously perceive greater time efficiency, improved academic outcomes, stronger academic self-efficacy and changes in the effort they put into problem-solving and critical thinking when completing tasks (cognitive effort change), which together shape their overall perception of productivity and may encourage more extensive AI use. The conceptual framework for our study is close to the one proposed by Chinga-Montero and Zambrano-Solórzano (2025). Time efficiency and academic performance capture observable outputs of AI adoption, while academic self-efficacy, defined as students’ confidence in their ability to understand and successfully complete academic tasks (Zimmerman, 2000), reflects students’ confidence in their own academic capabilities. Zhang and Xu (2025) demonstrated that this relationship is paradoxical: enhanced AI use significantly increases students’ confidence and learning efficiency, yet simultaneously intensifies technological dependence, creating a tension between short-term performance gains and longer-term autonomous learning capacity.

This tension motivates the inclusion of AI dependency in the analytical framework. Rather than constituting an additional dimension of productivity, AI dependency captures the extent to which students rely on AI while achieving these outcomes. Previous research shows that AI dependency is related to academic self-efficacy, academic stress and performance expectations, with academic stress and performance expectations contributing to the indirect relationship between self-efficacy and dependency (Zhang et al., 2024). Students in the same study also identified reduced creativity and critical thinking among the perceived negative consequences of AI dependency. In AI-assisted learning, a reduction in cognitive effort required for one part of a task does not necessarily imply an overall reduction in mental activity. Rather, generative AI may change both the amount of effort required for particular parts of a task and the activities towards which this effort is directed. Chen et al. (2025), for example, distinguished cognitive effort from task performance and emphasise that analytical writing requires substantial cognitive effort in connection with the development of critical thinking.

The distinction between cognitive effort and critical thinking is therefore important. Cognitive effort concerns how much mental resource is invested in a task, whereas critical thinking concerns what those resources are used to do, such as analysing, evaluating, verifying and making judgements. Rather than assuming that generative AI reduces critical thinking, Lee et al. (2025) showed that it may shift where critical-thinking effort is directed. Critical-thinking effort can move from information gathering to information verification, from direct problem-solving to working with AI-generated responses and from task execution to monitoring and evaluating AI-supported work. In academic tasks, this means that less effort may be needed to find and organise information, while more attention is required to check whether AI-generated information is accurate, relevant, supported by the original source and correctly interpreted. However, self-reported changes in critical thinking should be interpreted cautiously. Lee et al. (2025) noted that reduced effort may go in four different ways:

1. reflect the same amount of critical thinking being performed with greater AI support;

2. the delegation of some critical thinking work to AI;

3. less critical thinking overall; or

4. confusion between reduced cognitive effort in general and reduced critical thinking specifically.

Therefore, perceived changes should not be interpreted as direct evidence of either improvement or reduction in critical thinking.

The paper is structured as follows: Section 2 presents the research design and methodology. Section 3 presents the key findings, first building up on the descriptive analysis of survey data and second on the multivariate analysis. Section 4 discusses the results with prior research as well as highlights the limitations of our study, and the final Section 5 summarises the key findings and recommendations for practice.

This study analyses the role of AI for the productivity and changes in cognitive effort amongst university students in Vietnam. We adopted an exploratory cross-sectional research design. Data were collected through a structured questionnaire survey (see  Appendix 1) administered to students at the National Economics University in Hanoi, one of Vietnam’s largest public universities, enrolling approximately 45,000 students across a broad range of academic disciplines. Most students live in urban areas, come from middle-income households and have regular access to digital technologies, information technology resources and emerging AI tools. While the sample is not nationally representative, it provides valuable insight into the perceptions and behaviours of a digitally connected and technologically experienced student population that is likely to contribute to Vietnam’s future knowledge-based economy and digital transformation.

A non-probability convenience sampling strategy combined with snowball distribution was used. The questionnaire was disseminated between September and December 2024 through student representatives, student organisations and university-related social media channels. Participation was voluntary and anonymous and informed consent was obtained electronically at the beginning of the questionnaire before respondents proceeded to the survey. No personally identifying information was collected. From the received completed questionnaires, 103 were retained for statistical analysis. Missing responses were handled through listwise deletion in the regression analyses. No imputation procedures were required.

The questionnaire was organised around the key concepts of interest and involved five sections:

1. Demographic and background characteristics

2. AI adoption and usage patterns

3. Perceived effects of AI on productivity

4. Academic self-efficacy and learning outcomes

5. Changes in cognitive effort and AI dependency

The first two sections provided relevant background data, some of which were subsequently included as control variables in the empirical analysis. The third section focused on measures of perceived productivity gains with the key outcome Q39, alongside more specific items capturing students’ perceptions of AI-led time efficiency and changes in academic performance. Following Bećirović et al. (2025); the fourth section included items related to academic self-efficacy, capturing students’ perceived academic confidence and understanding of study material. The fifth section addressed the original component of our conceptual framework, namely, changes in cognitive effort and emerging AI dependency. Cognitive effort was operationalised as self-reported changes in the effort students put into problem-solving and critical thinking while completing academic tasks. Open-ended responses concerning inaccuracies encountered in AI-generated content were additionally classified into three categories: no identified inaccuracies, primarily technical or minor limitations and substantive content inaccuracies or fabricated information.

The original questionnaire was prepared in English and subsequently translated into Vietnamese by a native speaker. To ensure linguistic and conceptual equivalence, both language versions were reviewed before distribution. Most items were measured using ordinal response categories or Likert-type scales. Open questions at the end of the questionnaire were added to collect additional insights into the respondents’ experience with AI-led cognitive effort (changes in the effort they put into critical thinking while completing tasks). Individual questionnaire items were coded numerically, with higher values indicating a stronger presence of the measured construct (e.g. greater productivity gains, higher self-efficacy, stronger AI dependency or greater changes in cognitive effort). Multi-item constructs were grouped into theoretically informed dimensions based on prior literature. Where applicable, questionnaire items included standard non-response options, such as “prefer not to say”, as well as item-specific options indicating that the question was not applicable to the respondent. The raw data was cleared first for non-response or omitted responses reported for the key outcome variable Q39 – the productivity perceptions, and subsequently for the non-response responses recorded for the key explanatory variables. As a result, 103 responses were retained for analysis.

The productivity perceptions (PP) were measured as a binary variable oscillating between 0 (some productivity gains) and 1 (significant productivity gains). Originally, the data set for the question Q39 as the key underlying question for our outcome variable included also one response for the “No significant gains”, yet this one has been dropped from the data set. PP should be interpreted as reported likelihood of reporting significant rather than some productivity gains among AI-using students, and not as an objective measure of productivity or a measure distinguishing between productivity gains from productivity losses.

Time efficiency (TE) was measured using two approaches. First, participants estimated the time needed to complete a ten-page seminar paper with the help of AI tools (Q43) and without the use of AI tools (Q44). Second, students estimated the number of hours they spend studying each day without (Q50) and with the help of AI (Q49). Both approaches followed the logic of some earlier studies (e.g. Lozano and Blanco Fontao, 2023; Chan and Hu, 2023).

Academic performance (AP) was proxied using question Q45 which captures respondents’ perceived improvement in academic results, namely grades. The measure is based solely on students’ retrospective self-assessment and academic records were not independently verified. Thus, it shall not be interpreted as objective measure of academic achievement.

Academic self-efficacy (SE) refers to students’ confidence in their ability to understand academic tasks, engage in solving them and succeed in solving them (Bandura, 1997; Zimmerman, 2000). In our study, this dimension is proxied by Q51 which asks students whether the use of AI tools improved their understanding of study materials. The questionnaire also included a related item on perceived self-confidence in completing assignments with AI assistance (Q54). To maintain a parsimonious model and avoid including closely related self-reported measures in the same specification, the main analysis uses Q51 only. Accordingly, SE should be interpreted as capturing the perceived understanding dimensions associated with academic self-efficacy, rather than as a comprehensive measure of the broader construct.

Cognitive effort (Eff) refers to the amount of mental resources individuals actively invest while performing a task (Chen et al., 2025). In our study, cognitive effort was proxied using the perception of changes in the effort students put into problem-solving or critical thinking when completing their tasks (Q55). We acknowledge that this measure captures the broader concept of cognitive effort only to a limited extent and per se it should not be interpreted as a direct measure of the total amount of cognitive effort. Nor does it indicate the direction of the cognitive effort change since it reports only the changes in the amount of effort invested into critical thinking.

AI dependency builds up on the concept of Zhang et al. (2024) who refers to AI dependency as an excessive reliance on AI technologies and applications across various aspects of life, including academic studies, daily routines and social interactions. In our study, AI dependency was measured using Q58, which asked students whether they felt dependent on AI tools for studying and completing assignments. The measure should therefore be interpreted as self-perceived AI dependency, and not as an objective measure of reliance.

Following up on earlier studies (e.g. Chan and Hu, 2023) that acknowledge the role of AI use intensity, we introduce AI use into the model as well. For the main specification of the model, we opt for the inclusion of the number of hours students interact daily with AI. For that purpose, we use question Q30. The measure does not reflect cognitive engagement with AI.

We introduce also some control variables into the model following Zhai et al. (2024) and Melweth et al. (2023) who emphasise teamwork and digital literacy (next to individual effort to critically assess AI-generated output) as key controls of AI adoption process.

With regard to the binary outcome variable our model takes a probit form. Let $PPi*$ denote an unobserved latent propensity to report significant productivity gains:

where $εi∼N01$⁠. The observed binary outcome is defined as $PPi=1$ when $PPi*>0$ and $PPi=0$ otherwise. Accordingly:

where $Φ⋅$ denotes the cumulative distribution function of the standard normal distribution. The TE, AP, SE and Eff denote the AI-led self-reported time efficiency, academic performance self-efficacy and cognitive effort change as explained in Table 1 with the estimated coefficients $β1$ through $β4$⁠, respectively. The estimated coefficients $β5$ and $β6$relate to the productivity effects of AI daily use and AI dependency, respectively. Z is the vector of control variables with estimated parameters γ′.

Table 1 summarises the key survey variables selected for model development and analysis.

The collected data were first systematically organised and cross-tabulated using tables and figures to map the associations across the variables of interest. Pairwise associations among the explanatory variables and variance inflation factors (VIFs) were calculated to assess potential multicollinearity. These diagnostics were used to assess statistical overlap among the explanatory variables; they do not, however, eliminate the possibility of conceptual overlap or common-method variance arising from the self-reported nature of the survey measures.

The survey data indicate a high level of AI use among respondents. As many as 98.1% of respondents actively use AI tools in their academic work. The most frequently used tool was ChatGPT (94.2%), followed by Google Bard (36.5%) and Microsoft Copilot (23.1%). Most respondents (76.9%) had been using AI tools for more than a year, 16.3% had been using them for several months, while 2.9% had only recently started using AI and 3.8% could be considered complete novices. Respondents have integrated AI into a wide range of activities, including research and study (87.5%), academic writing (75%) and personal use (62.5%). A significant proportion of respondents also used AI for translation tasks (57.7%) and to explore new areas (45.2%). Most respondents did not receive any formal training in AI, although 52.9% of them reported interest to get some. Students’ primary motivation to use AI was reportedly related to time efficiency (85.6%), improving the quality of outputs (76%) and support to generate new ideas (63.5%).

Our survey data shows that 38.5% of respondents reported a reduction in the total time spent studying. Conversely, 20.2% of respondents reported spending more time studying while using AI. The largest group of respondents (41.3%) did not observe any significant changes in the time spent studying. These mixed responses provide no clear descriptive pattern regarding perceived time savings from AI use. To examine time efficiency in greater detail, we therefore compare respondents’ reported daily study time with and without AI assistance (Figure 1). Before using AI, most respondents reported studying 1–2 h per day, with the second-largest group studying less than one hour per day. After introducing AI into their study routine, the number of students in the 1–2 h per day category increased further. At the same time, the number of students studying 3–4 h or more per day decreased. However, the proportion of students studying less than one hour per day also decreased, consistent with the fact that most students did not experience any time save after AI adoption. The data thus provides no clear indication that AI use is associated with an overall reduction in study time. This does not, however, preclude time efficiency from being an important dimension of students’ perceived productivity gains. We examine this relationship further in the multivariate analysis.

Academic performance represents another potentially relevant dimension, captured in our survey data through students’ self-reported improvement in academic results following AI adoption. Figure 2 shows that students who already had excellent study results before the introduction of AI reported further improvements. Among average students, we observe a slight improvement in grades, whereas among C-scoring students, most respondents reported none or only some AI-led improvement in grades. This raises the question of whether AI may exacerbate differences in academic performance among university students – where high-achieving students improve further while weaker students do not experience comparable gains.

Unlike time efficiency and perceived academic performance, which are examined directly through the preceding descriptive comparisons, the self-efficacy-related dimension is represented in our empirical analysis by two closely related indicators: perceived improvement in understanding study materials (Q51) and increased confidence in completing academic tasks (Q54). We therefore cross-tabulate both indicators with students’ reported perception to gain significantly rather than moderately (some) from AI to examine whether they display comparable descriptive patterns before selecting one for the main multivariate specification.

Table 2 shows that both indicators display a similar descriptive association with students’ general productivity perception: students reporting greater improvement in understanding or confidence also tend to report stronger perceived productivity gains. Given their conceptual proximity, including both could increase overlap among the explanatory variables. We therefore use perceived improvement in understanding as the primary indicator and subsequently replace it with self-confidence as an alternative operationalisation in the sensitivity analysis.

As students experience AI-led time savings, they are motivated to use AI tools in other areas as well, including academic writing, research, translation tasks and personal and professional activities (Table 3). Daily and weekly users report using AI in several areas simultaneously, with both daily and weekly users involving AI gradually for personal purposes. This indicates a shift from functional use towards more integrative patterns of engagement.

Beyond areas of AI Use, Figure 3 examines whether daily users engage with AI also more hours per day. Weekly users represent the largest group (44.2%), followed by daily users (40.4%), monthly users (8.7%) and occasional users (5.8%). There were significant differences between the two largest groups (daily and weekly users) in the estimated daily time spent working with AI. While most daily users report using AI tools for 1–2 h a day, most weekly users spend less than 1 h a day with AI. At the same time, eight daily users reported using AI for more than 3 h a day, while only one weekly user reported daily intensity exceeding 3 h. In other words, our survey data confirm the pattern that the more frequently students integrate AI into their academic routines, the more time they devote to it each day.

Taken together, the descriptive data indicate that greater frequency of AI use is also associated with greater daily intensity of use: students who integrate AI more frequently into their academic routines reported spending more time interacting with it on the days they use it. This mechanism raises a question concerning the implications of more intensive engagement with AI. Students’ interaction with AI tools for more hours and more frequently create conditions under which reliance on these tools becomes a more salient feature of their academic routines. Simultaneously, it raises the question of whether more intensive AI engagement is accompanied by changes in the cognitive effort students report putting into problem-solving and critical thinking in their tasks.

To explore the mechanism further, we cross-tabulate daily AI use with the students’ reported perceptions on AI dependency. Figure 4 shows that students using AI for 1–2 h per day reported at least some degree of dependency more frequently than those using AI for less than 1 h per day. Among students using AI for 3–4 h per day, all respondents reported either some or significant dependency. Significant dependency, however, is not reported consistently across all categories. Almost 50% of students who reported significant reliance on AI reported using AI less than an hour per day. We examine these descriptive patterns further; yet, to maintain manuscript concise, the additional cross-tabulations are reported in  Appendix 2.

Nevertheless, AI dependency appears to become more common with greater daily AI use. Increasing reliance on AI, however, also raises the question of whether students experience changes in the effort they devote to problem-solving or critical thinking in their tasks. Greater reliance on AI may be accompanied by either reduced or, conversely, greater changes in effort devoted to problem-solving and critically engaging with AI-generated content while completing the tasks. To explore how these dimensions coexist and how they relate to students’ perceived productivity gains, we use cross-tabulations (Table 4).

Following up on Table 4, students reporting significant rather than some productivity gains also tend to report higher AI dependency and larger cognitive effort change, while the relationship of daily AI use with the both appears less uniform. These patterns suggest that productivity perception, AI dependency, cognitive effort change and daily AI use are interrelated in the descriptive data and motivate examining them jointly in the model estimation.

Of particular interest is the relationship between AI dependency and cognitive effort change. Significant cognitive effort change becomes more frequent across higher levels of reported AI dependency, increasing from 12.5% among students reporting no dependency to 28.8% among those reporting some dependency and 38.5% among those reporting significant dependency. This pattern suggests that greater perceived AI dependency may coexist with larger reported shifts in cognitive effort. We use this association as a basis for examining the interaction effects within the multivariate regression framework. To note, cognitive effort change captures the extent rather than the direction of the perceived change, thus it cannot be interpreted that greater dependency correlates with greater cognitive effort.

Building on these descriptive patterns, the following section examines these relationships within a multivariate framework.

Given the exploratory nature of the study and the relatively small sample, the empirical analysis follows a staged specification strategy designed to explore the main associations between the proposed dimensions of productivity perceptions while carefully considering model stability and risk of overfitting. The first family of specifications examines the association between the proposed dimensions of students’ AI-related academic experience and the probability of reporting significant rather than some perceived productivity gains. We first estimate a parsimonious baseline model and subsequently introduce two sets of background controls sequentially. This nested approach allows us to assess the stability of estimated coefficients while we treat the controls as variables of no substantive interest.

A second set of specifications explores whether the association between students’ reported change in cognitive effort and productivity perception varies with their level of AI dependency. AI dependency may coexist with greater perceived productivity while simultaneously being associated with differences in the effort students report putting into problem-solving or critical thinking in their tasks. We therefore estimate an interaction between cognitive effort change (Q55) and AI dependency (Q58), first in a parsimonious specification and subsequently with additional controls. This analysis is treated as exploratory and is not interpreted as establishing a causal moderation mechanism.

Finally, we assess the sensitivity of the results to the operationalisation of the self-efficacy dimension. Whereas the main specification uses perceived improvement in understanding as the focal indicator, the alternative specifications replace it with students’ reported self-confidence in completing academic tasks with AI assistance following more closely (Bandura, 1997). Given the limited sample size and the conceptual proximity among several self-reported measures, this sensitivity analysis uses a deliberately reduced set of covariates, retaining time efficiency, cognitive effort change and AI dependency while excluding daily AI use and perceived academic-performance improvement. The alternative specification is therefore intended as an operationalisation check rather than as a competing structural model.

The estimation results are summarised in Table 5.

A probit model yields parameters as changes in Z-scores (standard deviations) based on the cumulative normal distribution. These estimates cannot be interpreted as probabilities or impact parameters in our model. For interpretable results, we calculate average marginal effects that correspond to average of derivatives, i.e. partial derivatives of the probability with respect to each explanatory variable for every single observation in the sample using standard normal probability density function. In other words, we calculate effects of a 1-unit change in an independent variable on the probability of the outcome, calculated using each observation’s actual covariate values. The results are reported in Table 6.

Our estimations appear to show rather stable positive association of AI-led productivity perceptions with all proposed dimensions, namely, time efficiency (Q42), better grades (Q45), better understanding of study materials (Q51) as well as cognitive effort change (Q55) and AI dependency (Q58). It also appears that students correlate their productivity with time they spend daily with AI tools (Q30) and with use of AI tools in collaborative work (Q52).

AI-related time savings appear to be associated with approximately a 25% higher probability of reporting significant rather than some productivity gains and compared to the marginal effect estimates for other dimensions, it appears dominant. Grades and content understanding are associated with roughly half of the effect rendered for the time efficiency measure. Notably, self-confidence, which is conceptually more closely aligned with Bandura’s (1997) definition of self-efficacy, is associated with an approximately 12- and 11-percentage-point higher probability of reporting significant rather than some perceived productivity gains. This estimate is comparable to that obtained for perceived improvement in understanding in the main specifications, where the corresponding marginal effects range from approximately 11–14 percentage points. Cognitive effort change appears to be associated with approximately a 12% increase in the probability of reporting significant rather than some productivity gains from AI use. The estimation of interaction effects of cognitive effort change with AI dependency further suggests that the two are associated with the students’ likelihood to report only some rather than significant AI productivity gains.

The estimated coefficients and marginal effects of the focal variables remain broadly stable across all the specifications in Table 6, suggesting that the coefficients generated in the equation (1) specification of the model are not substantially altered by the inclusion of the additional observed characteristics.

Probit specification uses cumulative standard normal distribution to model the data. To assess whether the data follows rather logistic distribution, we re-estimated the same model using a logit specification, The logit estimates are substantively consistent with the baseline probit results, and are reported in  Appendix 3.

The findings support the proposed multidimensional view of student productivity. Time efficiency emerged as the strongest and most consistent factor associated with perceived productivity, which is consistent with Chinga-Montero and Zambrano-Solórzano (2025), who identify improved time management and greater efficiency in completing academic tasks among the main benefits of AI use. Wang (2025) similarly showed that generative AI can make different stages of academic work more efficient. Academic performance and academic self-efficacy are also associated with perceived productivity, consistent with previous findings by Bhullar et al. (2024) and Bećirović et al. (2025).

More importantly, our results suggest that students associate their perceived AI-led productivity with changes in the effort they put into problem-solving and critical thinking in their tasks. This finding adds an important cognitive dimension to the productivity framework. Chinga-Montero and Zambrano-Solórzano (2025) primarily emphasise time management, efficiency and academic performance, while also raising concerns about excessive reliance on AI and its possible implications for critical skills. Our study brings these aspects into the same model by examining change in cognitive effort alongside time efficiency, academic performance, academic self-efficacy and AI dependency. The results are also consistent with Lee et al. (2025), who show that generative AI can shift critical-thinking effort from information gathering towards information verification, from direct problem-solving towards AI-response integration and from task execution towards monitoring and evaluating AI-supported work. In our data, students who reported stronger productivity gains were also more likely to report changes in their effort they put into problem-solving or critical thinking in their tasks.

The relationship between productivity and AI dependency adds another layer to this pattern. Perceived AI dependency was positively associated with perceived productivity, indicating that productivity gains and greater reliance on AI coexist. This is consistent with Fan et al. (2025), and with Zhang and Xu (2025), who identify a tension between greater learning efficiency and increasing technological dependence. In our study, this association is explored alongside the four dimensions of student productivity both as an additional regressor and in interaction with changes in cognitive effort. The results suggest that the benefits students perceive from AI may be accompanied by stronger reliance on the same tools that support those benefits.

The results also point to potential redistributive effects within the academic environment. Students with stronger prior academic performance reported greater improvements, whereas lower-performing students reported more limited benefits. This suggests that AI may reinforce existing differences in academic performance rather than benefiting all students equally. Together with the observed differences in cognitive effort change and AI dependency, these findings support the need for AI literacy that emphasises source verification, evaluation of AI-generated outputs and responsible use (Chan and Hu, 2023; Miao and Holmes, 2023).

Despite these contributions, the findings should be interpreted with consideration of several limitations. The results are based on a relatively small sample drawn from a single university in Hanoi. This limits the broader generalisability of the findings and makes it difficult to determine whether the observed patterns would remain consistent across other universities in Vietnam or other Asian countries. The sample consisted predominantly of Vietnamese students with relatively similar academic backgrounds and was drawn from an urban environment characterised by access to digital tools and stable internet. Students in different regions or educational systems may therefore experience AI use differently. Future research should extend the analysis to additional universities in Vietnam or other countries to assess whether the observed associations remain consistent across different cultural, institutional and technological settings. A longitudinal study could further contribute to a more comprehensive understanding how impact of AI on student productivity develops over time. Our results reflect AI tools available in 2024, and the findings should therefore be interpreted accordingly.

More importantly, our results suggest that students associate their AI-led productivity with changes in the effort they put into problem-solving or critical thinking in their tasks. This finding adds to the productivity framework proposed by Chinga-Montero and Zambrano-Solórzano (2025) as well as Lee et al. (2025), who show that generative AI can shift critical-thinking effort from information gathering towards information verification. We find also significant association between AI dependency and productivity perceptions. This is consistent with Fan et al. (2025) and with Zhang and Xu (2025), who identify a tension between greater learning efficiency and increasing technological dependence. Yet, our study goes further and finds significant interaction between AI dependency and changes in cognitive effort. Regardless the exploratory nature of the study and limited sample, this interaction might provide additional insights into acceleration of AI adoption.

This study provides exploratory evidence that university students tend to associate their AI-led productivity with several distinct dimensions of their academic experience, including perceived time efficiency, improvement in academic performance, academic self-efficacy and changes in cognitive effort. Among these dimensions, time efficiency emerges as the strongest and most consistent correlate of self-reported productivity gains. Perceived improvements in academic performance and understanding of study materials are also positively associated with productivity perceptions, although their estimated marginal effects are smaller.

The findings further reveal a more complex relationship between perceived productivity, changes in cognitive effort and AI dependency. Students reporting greater dependency on AI are also more likely to report significant productivity gains. At the same time, perceived changes in cognitive effort students put into problem-solving and critical thinking are positively associated with productivity perception. The exploratory interaction analysis further indicates that the association between perceived cognitive effort change, and productivity varies with the level of reported AI dependency. These findings do not establish that dependency increases cognitive effort, or that intensive AI use strengthens critical thinking. Rather, they suggest that greater reliance on AI can coexist with greater self-reported cognitive effort, challenging a simple interpretation in which increasing AI use necessarily corresponds to limited rather than significant changes in cognitive effort. Given the cross-sectional and self-reported nature of the data, the direction and mechanisms underlying these relationships remain an important question for longitudinal and experimental research.

The collected data additionally point to potentially uneven academic benefits from AI use. Higher-achieving students were more likely to report further improvements in academic performance, whereas students with weaker prior performance reported more limited gains.

Our results thus point to the need for structured AI approach in university education. The educational challenge is not simply to encourage or restrict AI use, but to equip students with the skills needed to translate its efficiency benefits into effective and critically engaged learning. Structured AI literacy should therefore combine effective prompt formulation and appropriate use of AI tools with source verification, critical evaluation of generated content and reflection on when reliance on AI may substitute for independent reasoning. Particular attention should be paid to students who may be less able to convert access to AI into academic gains, so that AI-assisted learning does not inadvertently reinforce existing differences in academic performance. Assessment practices can support this objective by requiring students to justify, critique, verify and revise AI-generated material rather than merely reproduce it.

Lucia Satkova is based at the Department of International Trade, Bratislava University of Economics and Business, Bratislava, Slovakia.

Paula Puskarova is based at the Department of International Trade, Bratislava University of Economics and Business, Bratislava, Slovakia.

The anonymised data set underlying the findings of this study will be deposited upon publication. During peer review, the data are available from the corresponding author upon reasonable request.

Q1. Gender

• Male

• Female

• Non-binary

• Prefer not to say

• Other

Q2. Age

• Under 18

• 18–24

• 25–34

• 35–44

• 45–54

• 55–64

• 65–74

• 75 and over

• Prefer not to say

Q3. Nationality

• Vietnamese

• Japanese

• Chinese

• Korean

• Taiwanese

• Indonesian

• Thai

• Other

Q4. Country of birth

• Vietnam

• Japan

• China

• South Korea

• Taiwan

• Indonesia

• Thailand

• Other

Q5. Native language

• Vietnamese

• Japanese

• Mandarin Chinese/Cantonese/Shanghainese

• Korean

• Taiwanese Hokkien

• Hakka

• Indonesian

• Javanese

• Thai

• Other

Q6. Other languages spoken (multiple responses allowed)

• English

• Mandarin Chinese

• French

• Russian

• German

• Spanish

• Other

Q7. Education level

• No formal education

• Primary education

• Secondary education

• High school diploma

• Bachelor’s degree

• Master’s degree

• Doctorate/PhD

Q8. Current occupation (multiple responses allowed)

• Student

• Employed (full-time)

• Employed (part-time)

• Self-employed

• Unemployed

• Homemaker/caregiver

• Retired

• Volunteer/community service

Q9. Marital status

• Single

• Married

• Divorced

• Widowed

• Prefer not to say

• Other

Q10. Number of children (if any)

• 0

• 1

• 2

• 3

• 4

• 5

• Other

Q11. Have you ever travelled abroad?

• Yes

• No

Q12. If yes, how long did you stay abroad?

• Less than one year

• One to three years

• Four to six years

• More than six years

• I haven’t travelled abroad

Q13. Have you ever studied abroad?

• Yes

• No

• No, but I plan to

Q14. Have you ever worked?

• Yes, I am currently employed

• Yes, I have worked in the past

• No, I have never worked

Q15. If yes, how long have you worked for?

• Less than one year

• One to three years

• Four to six years

• More than six years

• I do not have any previous experiences

Q16. Are your parents currently employed?

• Yes, both parents

• Yes, one parent

• No, neither parent

Q17. How would you classify your family’s income level?

• Low income

• Middle class

• High income

Q18. Did you grow up with a personal computer at home?

• Yes

• No

Q19. When did you first have contact with a computer?

• Before age 6

• Between ages 6–10

• Between ages 11–15

• After age 15

Q20. Did you have computer science classes during your studies?

• Yes, regularly

• Yes, occasionally

• No, never

Q21. When did you first encounter IT (information technology)?

• Before age 10

• Between ages 10–15

• After age 15

Q22. Where did this exposure take place?

• Formal setting (e.g. school)

• Informal setting (e.g. clubs, friends)

• Both formal and informal settings

Q23. How would you rate your IT skills?

• Beginner

• Intermediate

• Advanced

Q24. Do you use artificial intelligence (AI)?

• Yes

• No

Q25. Which specific AI tools or platforms do you use the most? (Multiple responses allowed)

• ChatGPT

• Google Bard

• Microsoft Copilot

• Other

Q26. Have you received any formal training or guidance on how to use AI tools effectively? If yes, please describe briefly below.

• Open-ended response

Q27. Would you be interested in attending workshops or training sessions to improve your understanding and use of AI tools?

• Yes

• No

• Maybe

Q28. How often do you use AI tools like ChatGPT?

• Daily

• Weekly

• Monthly

• Rarely

• Never

Q29. How long have you been using AI?

• For the past year

• For the past several months

• For the past month

• Less than a month

Q30. If you use AI, how many hours per day do you typically spend using AI tools?

• Less than 1 h

• 1–2 h

• 3–4 h

• More than 4 h

Q31. In what types of activities do you use AI? (Multiple responses allowed)

• Personal use (e.g. hobbies, daily tasks)

• Professional work-related tasks

• Educational purposes (e.g. research, studying)

• Writing papers (e.g. thesis, dissertations, term papers)

• Translation tasks

• Other

Q32. Have you ever used AI tools for collaborative projects or team tasks?

• Yes

• No

Q33. Do you consider ethical implications when using AI tools in your work or studies?

• Always

• Often

• Sometimes

• Rarely

• Never

Q34. How do you feel about the impact of AI tools like ChatGPT on your work or study efficiency?

• Very positive

• Positive

• Neutral

• Negative

• Very negative

Q35. What are the main benefits you’ve experienced from using AI tools like ChatGPT? (Multiple responses allowed)

• Improved efficiency

• Better decision-making

• Enhanced creativity

• More accurate information or data

• Stress reduction

• Other

Q36. How comfortable do you feel about using AI for sensitive or confidential tasks?

• Very comfortable

• Somewhat comfortable

• Neutral

• Somewhat uncomfortable

• Very uncomfortable

Q37. What are the primary reasons you use AI tools? (Multiple responses allowed)

• To save time

• To enhance the quality of work

• For inspiration or ideas

• To assist with translations or language tasks

• For learning and exploring new topics

• Other

Q38. Have you encountered any challenges or difficulties while using AI tools? If yes, please describe briefly.

• Open-ended response

Q39. Do you feel that AI tools have improved your productivity?

• Yes, significantly

• Yes, somewhat

• No, not really

• No, it has made no difference

• No, it has decreased my productivity

Q40. How would you rate your overall satisfaction with the AI tools you use?

• Very satisfied

• Satisfied

• Neutral

• Dissatisfied

• Very dissatisfied

Q41. Would you recommend AI tools like ChatGPT to your peers or colleagues?

• Yes, definitely

• Yes, with some reservations

• No

• Not sure

Q42. Has the use of AI tools like ChatGPT helped you complete your tasks more quickly?

• Yes, significantly faster

• Yes, somewhat faster

• No, there is no noticeable difference

• No, it takes the same amount of time

Q43. How long does it currently take you to write a ten-page seminar paper using AI tools?

• Less than a day

• One to two days

• Three to four days

• More than four days

• Other

Q44. How long did it take before you started using these tools?

• Less than a day

• One to two days

• Three to four days

• More than four days

• Other

Q45. Have your academic results improved since you started using AI tools like ChatGPT?

• Yes, I have better grades

• Yes, but only slightly

• No, my grades have stayed the same

• No, my grades have declined

Q46. What was your average grade over the last two years before using AI tools?

• Excellent (A)

• Good (B)

• Average (C)

• Below average (D)

• Poor (E/F)

Q47. Has your average study time changed since you started using AI tools?

• Yes, I spend less time studying now

• Yes, I spend more time studying now

• No, my study time has stayed the same

Q48. Do you find that you have more time for extracurricular activities since you started using AI tools like ChatGPT?

• Yes, a lot more time

• Yes, a little more time

• No, about the same amount of time

• No, I have less time

Q49. How much time do you currently spend studying each day?

• Less than 1 h

• 1–2 h

• 3–4 h

• More than 4 h

Q50. How much time did you spend before using AI tools?

• Less than 1 h

• 1–2 h

• 3–4 h

• More than 4 h

Q51. Do you feel that AI tools have helped you understand your study materials better?

• Yes, significantly

• Yes, somewhat

• No, it has made no difference

• No, it has made it harder to understand

Q52. Have you used AI tools to collaborate on group projects or study sessions with classmates?

• Yes, often

• Yes, occasionally

• No, never

Q53. Do you use AI tools primarily for specific subjects or across all your subjects?

• Only specific subjects

• Across all subjects

• It varies

Q54. Do you feel more confident in your assignments and study materials since using AI tools like ChatGPT?

• Yes, much more confident

• Yes, slightly more confident

• No, about the same

• No, less confident

Q55. Have AI tools changed the effort you put into problem-solving or critical thinking in your tasks?

• Yes, significantly

• Yes, somewhat

• No, not really

• No, not at all

Q56. Would you recommend the use of AI tools like ChatGPT to other students?

• Yes, definitely

• Yes, with some reservations

• No

• Not sure

Q57. Have you experienced any drawbacks or challenges when using AI tools for your studies? If so, please describe briefly.

• Open-ended response

Q58. Do you feel that using AI tools has made you dependent on them for studying and completing assignments?

• Yes, completely dependent

• Yes, somewhat dependent

• No, I use them as an occasional aid

• No, not dependent at all