Artificial intelligence (AI)-mediated speaking applications increasingly extend English as a foreign language (EFL) oral practice beyond limited classroom time, yet evidence remains scarce on whether GPT-powered tools such as Talkpal.AI offer advantages over established instructional approaches. Grounded in self-regulated learning (SRL) theory, this quasi-experimental pretest–posttest study compared Talkpal.AI-mediated instruction, task-based instruction, and traditional instruction among 81 Chinese university EFL students in three intact classes: Talkpal.AI (n = 27), task-based (n = 26), and control (n = 28). Uncertainty-related speaking anxiety and metacognitive awareness in speaking were assessed using validated measures. Data were analyzed using MANCOVA with pretest scores as covariates, followed by univariate ANCOVAs and Bonferroni-adjusted pairwise comparisons. Results showed a significant overall group effect and significant between-group differences for both outcomes, with large effects. The Talkpal.AI group showed the most favorable adjusted posttest profile, followed by the task-based group; both groups showed more favorable outcomes than the control group. From an SRL perspective, structured AI-mediated speaking cycles involving goal setting, repeated practice, monitoring, and reflection may reduce speaking anxiety while strengthening metacognitive awareness. Pedagogically, guided Talkpal.AI rehearsal can be integrated as out-of-class preparation for classroom speaking tasks, supported by brief reflection activities that reinforce strategic regulation and reduce evaluative pressure.
Academic thriving reflects students’ continued learning and vitality. However, little is known about the psychological conditions needed for students to thrive when AI supports university learning. This study examined whether autonomy, competence, and relatedness were necessary conditions for academic thriving among 812 Chinese undergraduates who had used AI in their studies. Necessary condition analysis and bottleneck analysis were applied to overall thriving and its two dimensions. All three needs were necessary for overall thriving. Autonomy and competence showed similar medium effects, while relatedness had a slightly smaller effect. Competence was the main necessary condition for learning, whereas autonomy was the strongest constraint on vitality. Relatedness became increasingly important at the highest levels of vitality and overall thriving. These findings indicate that AI-supported education should strengthen competence, support meaningful choice, and maintain supportive relationships with teachers and peers.
Prospective memory often improves through repeated target encounters, but whether contextual reminders facilitate this learning or instead increase reliance on reminder contexts remains unclear. Guided by the Preparatory Attentional and Memory theory, we examined how contextual reminders influence the prospective and retrospective components of prospective memory. One hundred university students were randomly assigned to a Known, Unknown, or Control group and completed a dual-task paradigm. During Blocks 1–7, prospective memory targets were accompanied by contextual reminders in the Known and Unknown groups, but only the Known group was informed of their function. In the eighth block, every frame turned white, so the red reminder was no longer presented on prospective memory targets; the Control group underwent the same change. Behavioral measures and beta multinomial processing tree modeling assessed prospective memory performance and its components. Both reminder groups showed higher prospective memory accuracy than the Control group throughout learning, and their prospective memory responses became faster across blocks. After the reminder was removed, prospective memory accuracy declined in both reminder groups, and the Known group showed slower prospective memory and ongoing-task responses. Modeling showed that reminders enhanced the prospective component during learning, which decreased after the reminder was removed, whereas the retrospective component was selectively enhanced in the Known group. These findings suggest that contextual reminders facilitate prospective memory learning mainly through prospective processing, with explicit knowledge additionally supporting retrospective processing. However, reminder-supported benefits appear context-bound and may decline after context change.
Artificial intelligence (AI) is increasingly reshaping language learning in higher education, yet students’ engagement in AI-assisted language learning depends on more than access to technological tools. Drawing on Self-Determination Theory (SDT) as an overarching theoretical perspective, this study examined how emotional intelligence and metacognition were associated with engagement directly and indirectly through enjoyment and self-efficacy, while also evaluating their predictive value. Survey data from 847 university students in China were analyzed using structural equation modeling and random forest analysis with repeated nested cross-validation. The SEM results showed that emotional intelligence and metacognition were positively associated with engagement, enjoyment, and self-efficacy, while enjoyment and self-efficacy were also positively related to engagement. All four indirect effects were significant. The extended random forest model achieved an out-of-sample R² of .472, improving R² by .081 over the baseline random forest and by .069 over the extended linear model. Emotional intelligence showed the highest permutation importance and revealed potentially nonlinear predictive patterns across individual learners. These findings highlight the complementary value of explanatory and predictive approaches for understanding engagement in AI-assisted language learning.
Waldum and Sahakyan (2012) reported that directed forgetting (DF) of political statements depends on congeniality: participants selectively forgot politically incongruent statements, while congruent statements were unaffected. The present study tested whether this pattern replicates in a contemporary political context. In Experiment 1, 79 U.S. adults studied two lists of Democratic- and Republican-framed statements and received either a forget or remember instruction after List 1. Congruency was defined by participants’ political affiliation. In Experiment 2, 84 U.S. adults completed the same paradigm but rated their personal agreement with each statement, allowing recall to be examined as a function of both party-based congruency and subjective endorsement. Across both experiments, recall did not differ between congruent and incongruent statements in the remember condition, and the forget cue did not selectively reduce recall of incongruent statements. A pooled analysis (N = 163, >2.5 times the original sample) had approximately 97% power to detect a selective DF effect of the size originally reported, and the original effect size fell outside the confidence interval for the present estimate. Together, these findings do not replicate the selective DF pattern reported previously and are inconsistent with an effect of the magnitude originally observed. One possibility is that repeated exposure to both congenial and anti-congenial political content in contemporary media environments has reduced the contextual distinctiveness on which the original effect depends. Alternative explanations, including differences in materials and online administration, are discussed.
