Critical thinking (CT) is a core competence that schools are expected to cultivate, yet whether it reaches classrooms depends on teachers’ readiness to teach it. Using an explanatory sequential mixed methods design, we surveyed 312 primary English-as-a-foreign-language teachers in Shandong Province, China, across seven dimensions of readiness and interviewed five teachers. Latent profile analysis identified four profiles organized by two weakly related dimensions—the capacity teachers bring to CT teaching and the constraints they perceive around them: Empowered-Engaged (15%), Willing-but-Constrained (28%), Moderate-Unobstructed (20%), and Committed-but-Under-resourced (37%). Endorsement of CT was near-universal, so the profiles differed in capacity and circumstance rather than in commitment. The central finding is a disconnection between what teachers believe and what they report doing: Willing-but-Constrained teachers wanted further professional learning as much as their Empowered-Engaged peers did, yet reported teaching CT less often. The divergence held under methods that account for classification uncertainty and under alternative measurement specifications, though not under all of them, and profile membership was largely unrelated to teachers’ background characteristics. The interview findings help explain why positive beliefs about CT did not always correspond to classroom practice. Teachers reported limited CT-specific training, examination pressure, time constraints, and norms that discouraged questioning, despite generally recognizing the value of CT. This suggests that the main barriers to CT instruction may lie not in teachers’ attitudes toward CT, but in their opportunities, preparation, and working conditions for implementing it.

Cognitive load and foreign-language anxiety are often examined separately in second-language learning, although both may arise when limited processing resources are strained. This meta-analysis examined whether the two constructs covary and why their association matters for models of L2 performance. A PRISMA-guided review identified eleven eligible studies (N = 1250), which were synthesised using a random-effects model. Two independent machine coders repeated the full-text eligibility assessment, effect-size extraction, and quality appraisal; two of the authors then verified every coding against the source reports and resolved all discrepancies. This process recovered one wrongly excluded study and corrected one misextracted coefficient. Greater load was associated with greater anxiety, r = 0.41, 95% CI [0.29, 0.52], and the estimate was r = 0.37, 95% CI [0.27, 0.46], after the most influential study was removed. All included estimates were positive. A specification analysis that substituted every available alternative component, wave, subscale, subgroup, and path yielded pooled estimates from 0.38 to 0.45; setting all eleven studies simultaneously to their least and most favourable alternatives widened the range to 0.29–0.52. Heterogeneity was high, I² = 80.1%, and the 95% prediction interval [0.03, 0.69] extended almost to zero. The pooled estimate therefore represents the centre of a dispersed literature rather than an expected result for a new study. An exploratory contrast between real-time and self-paced tasks was not significant and was confounded with language skill. The review also identified a reporting gap: thirteen additional reports measured both constructs but provided no statistic linking them. Conventional publication-bias diagnostics cannot address this form of selective non-reporting, which means that the pooled estimate is best regarded as an upper bound. This first construct-specific synthesis of the load–anxiety association connects cognitive architecture and attentional control with research on L2 anxiety and indicates why instructional studies should assess cognitive and affective outcomes together. Because the evidence is concurrent, predominantly self-reported, and drawn almost entirely from Chinese-speaking settings, it cannot establish that either construct causes the other or that both reflect a single mechanism.

Metacognition has informed research on second and foreign language (L2) learning for three decades, while AI tools have supported such learning for nearly as long. Yet, no review has examined how the field conceptualises and measures metacognition under AI mediation. This systematic review identified 60 eligible studies published between 2016 and May 2026. Each study was coded on fourteen dimensions anchored in established taxonomies and appraised with the Mixed Methods Appraisal Tool. Only four eligible studies fall within the pre-ChatGPT portion of the review window, from January 2016 to November 2022; 56 were published during the subsequent three and a half years. Three findings emerge. First, measurement is narrowly concentrated: 98% of the 56 post-ChatGPT studies measured self-reported metacognitive regulation, and 66% relied on self-report alone. Only one study in the full corpus measured metacognitive accuracy. Five post-ChatGPT studies combined an AI-absent comparison with process-level data, and only one of these also measured judgement accuracy. Second, 41% of post-ChatGPT studies used tools that combined two or more functional orientations within a single interface, a configuration not represented by the field’s inherited four-category taxonomy of AI applications. Third, none of the 18 contrast-based studies assigned the regulatory operation to the system in place of the learner. In every design that specified the division of labour, the learner retained that operation; three designs did not specify it. The condition most in need of testing therefore remains untested. Reported findings skew positive, but correlational and qualitative designs account for a disproportionate share of that evidence. The review concludes with a research agenda centred on calibration and process data.