The integration of artificial intelligence into higher education classrooms has triggered intense academic debate regarding how automated systems alter student learning behaviors, cognitive load, and academic dependency. A recent nine-week randomized, active-controlled study conducted with Chinese English as a Foreign Language (EFL) undergraduates has provided new empirical data on these dynamics. The research evaluated whether a personalized generative AI (GenAI) coaching package—incorporating situational observations, personalized rationales, and open-ended questions—produces different developmental trajectories in self-regulated learning (SRL) and GenAI dependency compared to standard template-based adaptive metacognitive prompting (AMP). Background Context and Technological Framework As generative AI tools become permanent fixtures in second language (L2) writing environments, educators and instructional designers face a persistent paradox. While these systems successfully assist students in real-time drafting, editing, and lexical enhancement, growing concerns suggest that heavy reliance on automated output can displace critical cognitive tasks such as independent planning, self-monitoring, and evaluation. To mitigate these risks, adaptive scaffolding systems have emerged, designed to detect operational bottlenecks during writing sessions and issue targeted metacognitive prompts. To test the efficacy of advanced scaffolding versus static templates, the study utilized the Doubao Large Language Model (LLM) as the underlying technical foundation for both experimental conditions. Ninety-two Chinese EFL undergraduates successfully completed the entire nine-week longitudinal trial out of an initial cohort of one hundred participants, representing an 8% attrition rate. All participants had previously passed China’s College English Test Band 4 (CET-4) with a minimum score of 425, ensuring a baseline homogeneity in English language proficiency verified prior to random assignment. Chronology of the Nine-Week Trial The empirical investigation was meticulously structured across three distinct operational phases to capture both immediate learning outcomes and the durability of any observed behavioral shifts: Week 1 (Pretest Phase): Participants completed an independent baseline argumentative writing task under a standardized 60-minute, 300-word protocol without GenAI assistance. Following this assessment, students completed baseline questionnaires measuring self-regulated learning dimensions and GenAI dependency levels. Subsequently, participants were individually randomized via a computer-generated sequence into either the experimental AMP-Agent group or the control AMP group. Weeks 2 through 5 (Training Phase): Over a four-week span, participants independently completed four weekly GenAI-assisted argumentative writing tasks adapted from Cambridge IELTS Academic Writing Task 2 prompts. Working asynchronously within a seven-day window each week, students drafted essays while the underlying system monitored for predefined behavioral triggers (such as word count stagnation or prolonged inactivity) and content triggers (such as weak claims or off-topic drift). When triggered, the system delivered the designated intervention package. Weeks 6 and 9 (Posttest and Delayed Posttest Phases): Week 6 marked the immediate posttest (T2), where participants completed an unaided writing assessment paired with full questionnaire readministrations. Week 9 served as the delayed posttest (T3), utilizing a new writing task and identical psychological instruments. This three-week temporal gap between T2 and T3 allowed researchers to evaluate whether any group differences persisted after the active intervention concluded. Comparative Intervention Design While both study groups shared the exact same user interface, domain knowledge base, trigger infrastructure, and underlying LLM architecture, the actual content of the delivered prompts differed significantly: The Control Condition (AMP): When a behavioral or content trigger activated, the system delivered pre-written template prompts directing learners toward general metacognitive strategies, such as reviewing thesis statements or checking paragraph coherence. These prompts did not dynamically connect to specific local sentences in the student’s current draft, nor did they incorporate open-ended prompts tailored to the immediate writing obstacle. The Experimental Condition (AMP-Agent): Trigger activations initiated a structured, constrained LLM call utilizing the live essay text and interaction history. The generated response consisted of three specific elements: a non-evaluative situational observation of the draft, a personalized rationale explaining why a specific metacognitive strategy was relevant to that exact drafting issue, and an open-ended reflective question. Stringent constraint-validation gates evaluated length, prohibited direct corrections, and prevented verbatim text quotes before output release. Both groups concluded prompted interactions with a standardized reflective query regarding perceived helpfulness. Quantitative Findings on Self-Regulation and Dependency The data analyses revealed significant Group by Time interaction effects for overall self-regulated learning and GenAI dependency, surviving retrospective Holm sensitivity corrections. For self-regulated learning, measured across dimensions including goal setting, task strategy, time management, self-evaluation, environment structuring, and help-seeking, the mixed ANOVA demonstrated a significant interaction. Group means for SRL totals shifted from 3.14 (T1) to 3.56 (T2) and 3.83 (T3) in the AMP-Agent condition, compared to 3.15, 3.34, and 3.62 in the control group. However, follow-up Bonferroni-adjusted contrasts at individual time points did not retain statistical significance at the strict threshold, indicating that while overall developmental trajectories diverged, a definitive total-score advantage at specific post-intervention intervals requires cautious interpretation. Dimension-level analyses suggested notable exploratory shifts in environment structuring and help-seeking behaviors. In contrast, measurements of GenAI dependency—operationalized through cognitive-behavioral, functional, and emotional subscales—yielded clear between-group divergence. Self-reported dependency scores were significantly lower in the AMP-Agent group at both the immediate posttest and the delayed posttest assessments after adjusting for multiple comparisons. Total dependency means for the experimental cohort moved from 3.01 down to 2.67 and 2.53, whereas the control group scored 3.10, 2.90, and 2.75 across the three assessment periods. Writing Performance and Behavioral Sequence Analysis Writing performance was evaluated using the Jacobs ESL Composition Profile across content, organization, vocabulary, language use, and mechanics. Although both groups demonstrated proficiency gains from pretest to posttest, the primary Group by Time interaction for total writing scores was not statistically significant. While an exploratory between-group contrast at T2 favored the experimental group, this advantage was not preserved at the T3 delayed posttest after accounting for multiplicity corrections. Consequently, the study found insufficient evidence to claim a sustained, differential improvement in unaided writing performance resulting exclusively from the personalized rationale package. To examine micro-level interactions during the writing process, Log-Sequential Analysis (LSA) tracked behavioral transitions among planning, monitoring, revision, passive GenAI utilization, and reflection states. Although descriptive trends indicated fewer immediate recursive loops into passive AI reliance in the coaching condition, aggregate transition probabilities did not survive stringent familywise error corrections across the 20-cell comparison matrix. Implications for Educational Technology Design The findings underscore important considerations for educators and developers deploying generative artificial intelligence in second-language instruction. The evidence demonstrates that prompt design directly influences students’ self-reported psychological relationship with AI tools and metacognitive awareness. Providing structured rationales linked to immediate situational feedback correlates with lower reported dependency scores over time, suggesting that learners rely less unquestioningly on automated suggestions when prompted to cognitively engage with the underlying purpose of a revision strategy. At the same time, the research highlights the methodological necessity of decoupling self-reported attitudes from direct measures of independent academic performance. Because unaided writing score trajectories did not significantly diverge between the cohorts, instructional designers are cautioned against assuming that enhancements in metacognitive prompting automatically translate into immediate, transferable writing gains. Future research is encouraged to employ multi-method component designs that isolate the individual contributions of personalization, rationales, and interface characteristics across diverse international student populations. Post navigation General Exploitation Theory
