Exploring the Role of AI-Enabled Personalized Learning Analytics in Enhancing EFL Grammar Acquisition: The Mediating Effects of Learner Engagement and the Moderating Role of Emotional Intelligence

Theoretical Foundation

The current research is based on the DeLone and McLean (2003) Information Systems (IS) Success Model, which assumes the influence of the perceived quality of information system and its outputs on the outcomes of the users. Quality, in the context of an AI-enabled personalized learning analytics (AI-PLA) is defensibly considered a multidimensional perception of what the learners get and experience when engaging with the analytics context. In line with this, we theorize AI-PLA quality as the perceptions of (a) information quality, (b) system quality, (c) feedback/service quality, and (d) personalization quality by the learners. It is assumed that these quality perceptions will go hand in hand with enhanced learner engagement since more transparent, more trustworthy, and more practical analytics cause less friction and make grammar-centered learning more objective (DeLone & McLean, 2003).

In this IS-success framing, the role of the learner as an intermediate process between perceived quality in analytics and learning outcomes is placed. Engagement elicits long term behavior investment, mental investment, and good participation in grammar-related study by the learners. In instances where AI-PLA is high quality, the learners have a better chance to stay attentive to recommendations, continue working through error-driven feedback, and convert analytics cues into purposeful practice, which ought to be associated with higher grammar acquisition achievements.

In order to state an empirically testable bound condition that is reflected in what we measure and analyze, we appeal to Control-Value Theory (CVT) of achievement emotions (Pekrun, 2006). CVT insists that situational contexts of achievement have the ability to trigger emotion-valenced appraisals (e.g., confidence, anxiety, frustration), thereby influencing engagement. Demonstrating an error or a gap in progress in the form of a dashboard in analytics-based grammar learning may invoke affective responses to reinforce or even discourage further action. Emotional intelligence (EI), or the ability to recognize, interpret, and manage feelings thus, should be a humble moderator: students with higher EI can be somewhat more efficient in remaining engaged in the situations when performance shortages are brought to the fore by analytics (Pekrun, 2006; Wong and Law, 2002).

Overall, the IS Success Model suggests the essential explanatory framework of why the perceived AI-PLA quality must be related to engagement and grammar acquisition and CVT allows EI to be a theoretically appropriate limit to this. Based on this framework, we test a directional (but cross-sectional and thus correlational) mediated-moderated model where the quality of perceived AI-PLA is directly and indirectly predictive of grammar acquisition through engagement, and EI moderates the direct AI-PLA quality engagement relationship.

Hypotheses

AI-enabled personalized learning analytics (AI-PLA) synthesize learner trace data into targeted recommendations and just-in-time feedback that can close persistent gaps in EFL grammar—such as tense/aspect, articles, and prepositions—by directing attention to high-value practice opportunities. From an information systems perspective, technology exerts effects on outcomes when its information quality (accuracy, relevance, completeness), system quality (usability, reliability, responsiveness), and personalization are salient to users (DeLone & McLean, 2003; Komiak & Benbasat, 2006). EFL analytics dashboards and recommendation agents that effectively visualize the patterns of errors and provide valuable corrective feedback in a timely fashion can help to shorten the feedback loop and reduce uncertainty about what to study next and scaffold effective rehearsal of the hard-to-learn grammar forms (Rets et al., 2021). This is particularly consequential in the big programs of China which are primarily exam-oriented where personalized feedback on teachers is minimal and students are left to wonder on their own between lessons.

There is a conceptual rationale, independent of motivational mechanisms, in having the direct positive link between AI-PLA and achievement. The high-quality analytics operationalize the enhancement of input and targeted output by (a) surfacing input contingencies of form/meaning/use that underlie grammatical competence and (b) directing intentional, criterion-based practice that brings morphosyntactic knowledge to procedural control (Purpura, 2004). With analytics that are correct, timely, and tailored, learners can devote study time to their most influential types of errors, they can be guided on corrective measures in real-time, and they can measure their progress in fine-grained terms: all of which are likely to produce measurable effects in grammar performance tests. In line with IS Success logic, such benefits in net form when the quality attributes of technology are strong enough to yield reliable outcome relevant signals (DeLone & McLean, 2003). Thus, the research hypothesis is that AI-PLA quality has a main-effect relationship to EFL grammar acquisition without (but complemented by) motivational processes.

The theorized form of high-quality AI-powered personalized learning analytics (AI-PLA) is believed to bring vitality to the engagement of learners by simplifying the process of study choices, accelerating feedback systems, and personalizing the significance of goals. Through the information systems perspective, the perceptions of the quality of information (accuracy, relevance, completeness), system quality (usability, reliability, responsiveness), and meaningful personalization are the proximal variables of user reaction to information systems including use and satisfaction (DeLone & McLean, 2003; Komiak & Benbasat, 2006). In language learning, they are the responses that affect the engagements in the form of sustained behavioral effort, more profound strategic processing, and positive affect when performing the tasks the three facets of engagement most directly related to the achievement (Hiver et al., 2020). Demonstrating the pattern of errors and featuring just-in-time and personalized recommendations, learning analytics dashboards minimize confusion regarding what is the next step to take, which is a primary precondition of persistence and strategic control (Rets et al., 2021). Self-Determination Theory also postulates that the contexts that promote competence and autonomy give rise to high-quality engagement; the analytics that offers diagnostic feedback, in addition to having mastery-oriented advice and a choice of practice paths directly address these needs (Deci & Ryan, 2000; Ryan et al., 2021). Within large, exam-based EFL programs in China, where individual feedback between teachers is limited, AI-PLA can replace granular cues (responses) that maintain focus and effort during inter-class time with granular, timely cues that proffer action during inter-class time as choice architectures. Aligning correct signals to individual pathways, AI-PLA enhances the chances of learners to invest in grammar work cognitively (use of strategies, self-monitoring), behaviorally (time-on-task, persistence), and affectively (interest, less frustration; Hiver et al., 2020; Rets et al., 2021). Therefore, the research will assume a direct motivation channel between the quality of technology and engagement.

Learner engagement is a proximal driver of language acquisition since it focuses attention, effort and strategic control into the specific forms and functions that support grammatical competence. Engagement (behavioral, time-on-task, persistence), cognitive (deep strategy use, self-monitoring), and affective (interest, enjoyment) aspects of engagement are generally described in L2 research; the consistent congruity of these aspects is always linked to improved-quality practice and performance (Harper & Quaye, 2009; Hiver et al., 2020; Teravainen-Goff, 2023). In grammar, in particular, the improvement is based on repeated cycles of intensive practice and feedback that enhance the mapping between form, meaning and use; active learners have a higher likelihood to (a) identify recurring patterns of errors, (b) implement corrective measures, and (c) persevere until procedural control is achieved (Purpura, 2004). Contexts that facilitate autonomy and competence, viewed through the motivational perspective, are likely to generate high-quality engagement, or deep processing and persistence, which, in turn, provides performance gains (Ryan et al., 2021). The emotion processes also play a role: learners put in more effort and regulate themselves better in case they have activating, controllable emotions (e.g., constructive challenge), which also increases the engagement-achievement relationship (Pekrun, 2017). Students in China with large, exam oriented EFL programs, where individualization is seldom provided through the teacher, are particularly determined to maintain deliberate feedback-based practice between classes; students who are behaviorally and cognitively engaged can more easily consolidate morphosyntactic knowledge and transfer it to tests. Integrating such strands, engagement is not an outcome product but the instant process in which study behaviors build up to proficiency.

Building on information systems and motivational theories, the relationship between AI-enabled personalized learning analytics (AI-PLA) and grammar outcomes is best understood as operating through learner engagement. The IS Success Model posits that information and system quality influence individual benefits primarily via proximal user responses (use, satisfaction), implying that technology quality exerts its effects through intermediary states rather than solely by direct transmission to outcomes (DeLone & McLean, 2003). In language learning, these proximal responses manifest as engagement—sustained behavioral effort, deep cognitive strategy use, and positive affect during tasks—rather than generic “use” (Hiver et al., 2020). When analytics are accurate, timely, and personalized, they clarify what to practice, reduce search costs, and supply diagnostic feedback, thereby satisfying competence and autonomy needs that energize high-quality engagement (Deci & Ryan, 2000; Ryan et al., 2021). Engagement, in turn, is the engine converting guidance into repeated, feedback-informed cycles of rehearsal that consolidate the form–meaning–use mappings underlying grammatical competence (Purpura, 2004). Empirically, recent work links stronger L2 engagement with deeper processing and better performance, suggesting that the benefits of technology-enabled interventions materialize through learners’ ongoing investment of attention and effort (Hiver et al., 2020; Teravainen-Goff, 2023). In China’s large, exam-oriented programs—where individualized teacher feedback is limited—such an engagement pathway is especially plausible: analytics can scale guidance, but only engaged learners translate signals into deliberate practice that yields measurable grammar gains. Accordingly, the model predicts a positive indirect effect from AI-PLA quality to EFL grammar acquisition via learner engagement.

When it highlights the instances of constant grammar errors, it may cause anxiety, frustration, or even shame, especially when it is used in the context of high-stakes coaching, such as in the Chinese EFL program. Control-Value Theory predicts that such emotions influence the engagement through the appraisals of control (can I fix this?) and value (does it matter?) by learners (Pekrun, 2017). Emotional intelligence (EI) also prepares students to evaluate feedback correctly, use activating emotions and down-regulate debilitating ones, maintain attention, strategic processing, and persistence against corrective signaling. Operationalized using Wong and Law Emotional Intelligence Scale (WLEIS), EI includes self- and other-emotion appraisal, use of emotion to facilitate performance, and emotion regulation, which have been established in Chinese university students and have been associated with adaptive coping and academic functioning (Kong et al., 2013; Shi & Wang, 2007; Wong & Law, 2002). Provided that the AI-enabled personalized learning analytics (AI-PLA) provide the correct, timely, and personalized feedback, high-EI learners are more prone to decoding the errors as active information, maintain motivation, and transform the recommendations into the intentional practice. On the other hand, lower EI learners have the threat appraisals that reduce engagement despite the high quality of analytics. Therefore, EI cannot simply increase engagement: it must enhance the beneficial impact of the quality of AI-PLA on engagement by converting the emotional connotation of feedback by switching the ego-threat emotion to the mastery challenge one. The model describes the heterogeneity in willingness to act on analytics as a result of technology quality by identifying EI as a boundary condition to the technology-engagement relationship. In this framework, emotional intelligence functions as a boundary condition on how learners respond to potentially ego-threatening information. Learners who can monitor and regulate their emotions are more likely to interpret red flags in dashboards as informative cues for adjustment rather than as global judgments of ability. They may thus remain engaged with AI-PLA even when confronted with frequent error signals, whereas learners with lower emotional intelligence may disengage or avoid analytics. Accordingly, we hypothesize a positive interaction between EI and AI-PLA quality in predicting engagement.

Research Site and Participants

We surveyed EFL learners enrolled in undergraduate English courses at four tertiary institutions in mainland China: two public universities, one vocational college, and one private language institute. After data screening (see Section “Sample Size Planning”), the final sample comprised 472 learners (58.9% female; see Table 1). Most participants were between 18 and 23 years old (45.3% aged 18–20, 44.5% aged 21–23), with a smaller group aged 24 or above. Year-of-study and major distributions reflected the mixed EFL provision at these institutions: 43.6% were English or applied linguistics majors and 56.4% were non-English majors taking compulsory EFL courses. Self-reported prior proficiency, benchmarked against CEFR levels, ranged from A2 to B2+, with A2 (25.0%), B1 (50.0%), and B2+ (25.0%) bands all represented. Weekly AI-PLA use also varied, with almost half of the sample reporting 1 to 3 hr/week and smaller groups at lower and higher usage bands. Together, this heterogeneity allowed us to examine AI-PLA, engagement, and grammar acquisition across a realistic cross-section of Chinese university EFL learners. Table 1 summarizes the socio-demographic profile and organizational segments. Briefly, the sample was geographically balanced across the four cities, skewed slightly female, and primarily at intermediate proficiency (B1). Most respondents were enrolled in public universities, with meaningful representation from private universities, vocational colleges, and private language institutes—mirroring the heterogeneous EFL delivery landscape examined in this study.

OPEN IN VIEWER

Table 1.

Sample Socio-Demographic Characteristics (N = 472).

Variable	Category	n	%
Gender	Female	278	58.9
	Male	186	39.4
	Other/Prefer not to say	8	1.7
Age	18–20	214	45.3
	21–23	210	44.5
	24+	48	10.2
Year of study	Year 1	122	25.8
	Year 2	150	31.8
	Year 3	130	27.5
	Year 4	70	14.8
Major	English/Applied Linguistics	206	43.6
	Non-English majors (taking EFL)	266	56.4
Prior proficiency (CEFR, self-report)	A2	118	25.0
	B1	236	50.0
	B2+	118	25.0
Weekly AI-PLA use (hr)	<1	106	22.5
	1–3	224	47.5
	4–6	98	20.8
	>6	44	9.3
Organization type	Public university	248	52.5
	Private university	96	20.3
	Vocational college	84	17.8
	Private language institute	44	9.3
City	Wuhan	128	27.1
	Xi’an	116	24.6
	Harbin	112	23.7
	Kunming	116	24.6

Note. Percentages may not total 100 due to rounding.

An invitation was disseminated by course coordinators in every institution that characterized the study as an assessment of AI-assisted grammar study. In each of the participating classes, the first author or a trained research assistant came during regular EFL sessions, briefly outlined the study, and asked all students present to take part in the study on an anonymous basis. Students who were interested accessed the survey and grammar test using their own devices following a QR code and only a few filled out paper-based forms which were later scanned. It was voluntary and did not require any course credit or grade incentives.

Sample Size Planning

Because our structural model includes a reflective–reflective higher-order construct (AI-PLA quality) and a latent interaction term (AI-PLA quality × emotional intelligence), we followed recent sample-size recommendations for PLS-SEM models with hierarchical components and interactions (J. Hair & Alamer, 2022; Wang et al., 2025). With four first-order indicators of AI-PLA quality, one second-order engagement construct, and the interaction term, the maximum number of arrows pointing at a single endogenous construct was three. Monte Carlo simulations reported in the PLS-SEM literature suggest that, for small-to-medium effect sizes (f² ≈ 0.05–0.15), a sample of around 400 to 450 cases yields power above 0.80 for such models; our N = 472 therefore provides adequate power for testing the hypothesized paths.

Instruments for Data Collection

All focal constructs were measured with established, validated scales and adapted for the present context using forward–back translation and cognitive interviewing. Items were anchored on five-point Likert scales (1 = strongly disagree, 5 = strongly agree) unless otherwise noted.

AI-PLA Quality

Perceived AI-PLA quality was operationalized through four first-order reflective dimensions—information quality, system quality, feedback quality/timeliness, and perceived personalization—adapted from information systems success and learning analytics research (DeLone & McLean, 2003; Rets et al., 2022; Susnjak et al., 2022). Each dimension was measured with 3 to 4 items capturing the clarity and usefulness of dashboard information (information quality), reliability and ease of use of the platform (system quality), timeliness and specificity of feedback on grammar tasks (feedback quality/timeliness), and the extent to which recommendations were tailored to learners’ weaknesses (perceived personalization). Table 2 reports reliability and validity statistics, and the second-order AI-PLA quality construct is modeled as a formative combination of these four dimensions.

OPEN IN VIEWER

Table 2.

Reliability, Validity, and Bias/Collinearity Diagnostics (N = 72).

Construct	α	ρA	CR	AVE	√AVE	Max r with others	HTMT	Full coll. VIF	Notes
Information quality (IQ)	.89	.90	.93	.72	.85	.62	.74	2.1	Reflective
System quality (SQ)	.88	.89	.92	.68	.82	.58	.71	2.0	Reflective
Feedback quality/Timeliness (FQ/T)	.87	.88	.91	.69	.83	.63	.73	2.0	Reflective
Perceived personalization (PP)	.90	.91	.94	.77	.88	.64	.79	2.2	Reflective
Learner engagement (LE)	.92	.93	.95	.74	.86	.60	.78	2.6	Reflective (second-order R → R)
Emotional intelligence (EI)	.91	.92	.94	.71	.84	.42	.55	1.9	Reflective (WLEIS total)
Grammar acquisition (OPT)	—	—	—	—	—	.55	.66	2.4	Single-indicator performance DV
AI-PLA quality (HOC)	—	—	—	—	—	—	—	1.8–2.5	Formative weights (bootstrapped): IQ = .36***, SQ = .27***, FQ/T = .21**, PP = .29***

Note.α = Cronbach’s alpha; ρA = rho_A; CR = composite reliability; AVE = average variance extracted. √AVE should exceed the construct’s highest correlation (“Max r with others”); HTMT_max = highest heterotrait–monotrait ratio with any other construct (all < .85; 95% CIs <0.90); “Full coll. VIF” are Kock’s full collinearity VIFs used as an omnibus check for common method bias (all <3.3). The second-order AI-PLA quality construct was estimated using a disjoint two-stage approach with its four first-order reflective dimensions as formative indicators in stage two; reported are formative weights and their significance (*p < .05, **p < .01, ***p < .001) and formative VIF range. Fornell–Larcker criterion was assessed for reflective constructs (DV excluded because it is single-indicator). SRMR = .058; NFI = .91. All thresholds met or exceeded.

Data collection process

To control the data collection among institutions, the protocols and instruments were identical over all the four sites within a 4-week period during the middle of the semester when students should have used the AI-PLA dashboard a few weeks earlier, but still worked on the target grammar units. Since the proficiency of the language may influence both the engagement and performance, the self-reported CEFR band A2, B1, B2+ (see Table 1) of learners were recorded, and the grammar test was developed to address intermediate constructions that are found in the core syllabi of each of the given levels. This minimized chances of advanced learners being bored by the test as well as lower-performing learners experiencing pervasive floor effects. In its interpretation, we consequently focus AI-PLA, engagement and grammar acquisition as acting in a context of moderate to a high degree of proficiency as opposed to extreme beginner or extreme advanced proficiency.

Data Analysis and Model Specification

We used partial least squares structural equation modeling (PLS-SEM) in SmartPLS 4 to estimate both the measurement and structural models. AI-PLA quality was modeled as a second-order formative construct composed of four first-order reflective dimensions: information quality, system quality, feedback quality/timeliness, and perceived personalization. Learner engagement was specified as a reflective higher-order factor indicated by behavioral, cognitive, and emotional engagement subscales. Emotional intelligence and grammar acquisition were modeled as first-order constructs, with grammar acquisition represented by a single observed test score. Following the two-stage approach for hierarchical component models, we first obtained latent-variable scores for the first-order constructs and then used these scores to estimate the higher-order AI-PLA quality construct and the latent interaction term (AI-PLA quality × emotional intelligence) via the product-indicator method.

All analyses were done using SmartPLS 4, which is highly appropriate with hierarchical component model (HCM) of the current study, latent interaction to moderate, non-normal indicators, and the objectives of the study are prediction-oriented. PLS-SEM allows disjointed two-stage estimation on our second-order AI-PLA construct, consistent PLS estimation on reflective blocks, and nonparametric bootstrapping of inference-strengths, which are consistent with our sample size and model complexity.

The structural model included a hypothesis test in each research question. The direct path of perceived AI-PLA quality to grammar acquisition tested H1 (RQ1). H2 (RQ2) was tested using the direct route of perceived AI-PLA quality to engagement of the learner. The H3 (RQ3) hypothesis was tested using the route between learning engagement and acquisition of grammar. H4 (RQ4) was tested using the bootstrapped indirect effect of the quality of AI-PLA and grammar acquisition through engagement. H5 (RQ5) was tested through interaction term (AI-PLA quality × emotional intelligence) to predict interaction, and then simple-slope probing at ±1 SD of EI.

Ethics

The guidelines of Helsinki Declarations were followed. The protocol received institutional ethics approval from Academic Ethics Committee of the host university. Participation was voluntary with no course credit or monetary incentives. Before the first item, participants read an informed consent statement describing risks, benefits, confidentiality, data retention (encrypted storage; de-identified exports), and the right to withdraw at any time without penalty. Only age (to confirm ≥18), city, institution type, and study-related variables were recorded. Under-18 students were not eligible.

Fitness of the Data

Data screening on the 505 returns produced N = 472 valid cases after removing submissions that failed attention checks, exhibited extreme straight-lining, had implausibly short completion times, or exceeded missing-data thresholds. Reflective measurement quality was strong: all standardized loadings exceeded .70, item reliabilities exceeded .50, and internal consistency indices were high (Cronbach’s α, ρA, and composite reliability ranged .87–.95). Convergent validity was established with AVE ≥ .64 for all reflective constructs. Discriminant validity held by both Fornell–Larcker (√AVE for each construct exceeded its largest inter-construct correlation) and HTMT (all HTMT ratios <.85; 95% bootstrap CIs stayed below 0.90). The second-order formative AI-PLA construct showed significant indicator weights for its four first-order dimensions with acceptable collinearity (VIFs < 2.6).

Evidence against common method bias was multi-pronged. Procedurally, we used temporal separation for the objective OPT grammar score, anonymity, mixed scale formats, and attention checks. Statistically, Harman’s single-factor accounted for 33.7% of variance, the full collinearity VIFs for all latent variables were <3.3 (range 1.8–2.8), and a common latent factor captured ≤24% additional variance; together these suggest CMB is unlikely to bias estimates. Multicollinearity was not a concern at either the outer- or inner-model level (outer VIFs < 3.0; inner VIFs for endogenous blocks ≤2.7). Model fit indices were acceptable for PLS-SEM: SRMR = .058 and NFI = .91, supporting adequate model-implied correlations.

Table 2 summarizes reliability, validity, and collinearity diagnostics. In brief, all reflective constructs met or exceeded conventional thresholds (α/CR ≥ .70; AVE ≥ .50; HTMT < .85), the second-order formative AI-PLA exhibited significant weights with low collinearity, and omnibus CMB screens were negative—collectively indicating a healthy measurement model suitable for hypothesis testing.

Hypotheses Testing

Structural paths were estimated in SmartPLS 4 using disjoint two-stage modeling for the higher-order AI-PLA construct and 5,000 bias-corrected bootstrap resamples (two-tailed; see Table 3). Standardized coefficients (β), t values, p values, and 95% BCa confidence intervals are reported. Inner-VIFs were ≤2.7, indicating no multicollinearity issues. The model explained substantial variance in Learner Engagement (R² = .56) and Grammar Acquisition (R² = .44), supporting the predictive focus of PLS-SEM.

OPEN IN VIEWER

Table 3.

Bootstrapped Path Estimates and Hypothesis Decisions (N = 472; 5,000 Resamples).

Hypothesis	Path	β	t	p	95% BCa CI	f 2	Decision
H1	AI-PLA → Grammar Acquisition	.19	4.02	<.001	[0.10, 0.28]	0.05	Supported
H2	AI-PLA → Learner Engagement	.48	11.30	<.001	[0.40, 0.55]	0.38	Supported
H3	Learner Engagement → Grammar Acquisition	.41	8.20	<.001	[0.31, 0.50]	0.21	Supported
H4	AI-PLA → Engagement → Grammar (indirect)	.20	6.40	<.001	[0.13, 0.27]	—	Supported
H5	AI-PLA × EI → Learner Engagement	.12	2.88	.004	[0.04, 0.20]	0.02	Supported

Note. BCa = bias-corrected and accelerated. Mediation evaluated via bootstrapped indirect effect (H4). Interaction probed via simple slopes at ±1 SD of EI (H5).

Answer to Research Questions

The fifth research question wanted to know whether emotional intelligence influences AI-PLA quality to engagement relationship. In line with H5, the interaction term was significant but the magnitude was low (β_interaction = 0.12, t = 2.88, p = .004) which corresponds to a low degree of strengthening the association between AI-PLA quality and engagement at increased levels of EI.

Theoretical Implications

The results are mostly an extension of the previous studies as they apply the existing models to the grammar-learning setting with the support of AI and PLA instead of creating a completely new theory. In line with the IS Success Model, perceived quality of AI-PLA among learners was similar and more so associated with greater engagement and, to a lesser degree, with better grammar results (DeLone & McLean, 2003). The outcomes of the mediation indicate the practical importance of considering engagement as an intermediary that learners can use to transfer analytics perceptions to learning outcomes. Secondly, the insignificance of the interaction effect indicates that emotional abilities can act as a weak boundary condition in analytics-intensive setting, which is congruent with the focus of CVT in processes of engagement related with emotion (Pekrun, 2006). On the whole, the research provides a logical empirical example of these connections in EFL grammar learning, and the longitudinal or experimental studies are required to be able to test the causality.

Practical Implications

Practically, the results offer cautious guidance for instructors, program leaders, and platform designers. First, because perceived information, feedback, and personalization qualities are strongly associated with engagement, it seems worthwhile to invest in dashboards that clearly visualize grammar progress at the feature level and that provide specific, timely recommendations for practice. Second, the small EI moderation effect suggests that simple emotion-aware routines—such as normalizing errors, encouraging reflective rather than punitive interpretations of analytics, or offering optional emotion-regulation tips—may help some learners stay engaged, though our data do not speak to the effectiveness of any particular training program. Importantly, without behavioral trace data on how learners actually clicked through dashboards or changed their study plans, we cannot recommend specific “choice architecture” features. Our recommendations should therefore be read as design hypotheses to be tested in future experimental or mixed-methods work rather than as direct prescriptions.

Limitations and Potential Research

This study has several limitations that qualify the conclusions. First, the design is cross-sectional and the questionnaire and grammar test were administered in a single sitting, separated only by a short break. Although PLS-SEM encourages directional modeling grounded in theory, the observed paths are correlational; we cannot rule out reverse or reciprocal influences (e.g., learners who are already strong in grammar may view AI-PLA more positively). Second, while we implemented multiple procedural and statistical remedies, common method bias cannot be definitively excluded because AI-PLA quality, engagement, and emotional intelligence were all self-reported in the same session. Third, our AI-PLA quality construct captures learners’ perceptions of information, system, feedback, and personalization quality rather than any objective system logs or algorithmic performance metrics. Conclusions about “high-quality analytics” should thus be understood as statements about students’ experiences with analytics, not about the underlying technology itself.

Future research should therefore combine perception measures with behavioral and system-level data, such as trace logs of how learners navigate dashboards, which recommendations they follow, and how their grammar performance evolves over time. Longitudinal or experimental designs that manipulate dashboard features, feedback framing, or emotion-regulation supports would allow stronger causal inferences about how AI-PLA, engagement, and emotions jointly shape grammar development. Qualitative work could also explore how learners interpret analytics visualizations and how teachers integrate AI-PLA into grammar instruction, extending the present focus on learners in mainstream Chinese tertiary EFL programs.