Participants were second-year undergraduates enrolled in a traditional PE elective. The trial compared real-time AI pose feedback with conventional instruction over 12 weeks. Students in the AI condition received a superimposed skeleton on a screen, with deviation cues drawn at key joints, while the control group practiced the same movements with instructor feedback alone.

The AI system processed each frame in roughly 120 milliseconds, drawing deviation vectors at the shoulders, hips, and knees. Students saw their own skeleton overlaid on an expert template, with red lines marking where their pose diverged. The system updated on every frame, so students could adjust their posture in real time rather than waiting for instructor correction.

The trial measured competence, autonomy, and motor skill at baseline and after 12 weeks. Students in the AI condition showed significantly greater gains in all three measures. The competence pathway—students’ sense that they could master the movements—explained a substantial share of the motor skill improvement. Autonomy gains, students’ sense that they could direct their own practice, also contributed.

The findings suggest that AI feedback works not by replacing instructors but by giving students a visible, immediate record of their own progress. The skeleton overlay turns an abstract instruction like “keep your back straight” into a concrete visual target. Students can see exactly where their pose diverges from the expert template, and they can adjust in the moment rather than waiting for the next correction.

The study also found that the AI condition did not reduce students’ sense of connection to their instructor. Students in both conditions reported similar levels of relatedness, suggesting that AI feedback can supplement rather than supplant the human relationship in PE.

The trial has implications for the design of AI feedback systems in PE and beyond. The competence pathway suggests that feedback should be visible, immediate, and specific. The autonomy pathway suggests that students should be able to direct their own practice, with the AI serving as a record of progress rather than a director of behavior. The relatedness finding suggests that AI feedback can supplement rather than supplant the human relationship.

The study is among the first to test the motivational mechanisms behind AI feedback in traditional PE. The findings suggest that AI feedback works by building students’ sense of competence and autonomy, with the competence pathway explaining a substantial share of the motor skill gains. The relatedness finding suggests that AI feedback can supplement rather than supplant the human relationship in PE.

The study’s design deserves closer attention because it addresses a common weakness in educational technology research. Rather than randomly assigning individual students, the researchers randomized twelve intact classes, which preserved the natural structure of real teaching. Three instructors each taught two classes per condition, meaning that any instructor-specific effects—teaching style, enthusiasm, or experience—were distributed across both groups. This cluster-level randomization, combined with negligible intraclass correlations reported in the analysis, strengthens the case that the observed differences in competence, autonomy, and motor skill reflect the feedback intervention itself rather than differences in who was teaching or how classes happened to be composed.

The choice of Zhuang Daoyin as the learning context is also scientifically meaningful. Traditional ethnic sports place unusual demands on learners because their movements coordinate posture, breathing, rhythm, and expressive qualities simultaneously. Unlike standardized exercises where a single biomechanical error may be obvious, these practices require learners to detect subtle discrepancies between a demonstrated form and their own execution. Conventional instruction handles this through a demonstration–imitation–correction cycle in which one teacher observes many students and comments after errors have already occurred. The AI system’s contribution was to compress that correction loop from minutes or entire sessions down to roughly a tenth of a second, allowing adjustment while the movement was still being performed.

The theoretical framing draws on self-determination theory, a widely tested account of motivation that distinguishes between controlled forms of regulation—acting from external pressure or internal guilt—and autonomous forms, in which behavior is self-endorsed and personally valued. The relative autonomy index used in the study summarizes a learner’s position along this continuum. Decades of research in physical education have shown that environments satisfying the basic psychological needs of autonomy, competence, and relatedness tend to foster autonomous motivation, which in turn predicts effort, engagement, and learning. What this study adds is a test of which specific need an AI feedback tool engages, rather than treating the technology as an undifferentiated intervention.

The mediation results are the study’s most distinctive contribution. The statistical model showed that the effect of AI feedback on motor skill traveled entirely through a two-step sequence: the tool first strengthened students’ sense of competence, and that competence gain in turn shifted motivational regulation toward more autonomous forms, which then predicted skill improvement. The direct path from the AI condition to motor skill, once this sequence was accounted for, was not statistically significant. This pattern matters because it rules out a simpler explanation—that the technology improved skill through some non-motivational channel such as merely increasing practice volume—and instead supports the claim that the psychological experience of visible progress is the active ingredient.

Equally informative is what did not change. Relatedness, the sense of mutual care and belonging with instructors and peers, showed no difference between conditions. From a measurement standpoint, this is discriminant evidence: if the AI condition had simply made students more enthusiastic about the course in general, one would expect ripples across all three psychological needs. Instead, the effects were confined to the needs most directly tied to the tool’s affordances—competence, because the feedback makes mastery visible, and autonomy, because students can self-correct without waiting for an authority. The untouched relatedness score suggests the effects are specific rather than diffuse.

The autonomy finding warrants some nuance. One might expect algorithmic feedback to feel like surveillance, undermining volition, yet students in the AI condition reported greater autonomy. A plausible reading is that the skeleton overlay functioned as an impersonal information source rather than a commanding one. When a visual display shows where a pose diverges from a template, the learner retains the decision about whether and how to adjust, whereas verbal correction from an instructor can carry implicit evaluative weight. This aligns with self-determination theory’s distinction between controlling and informational feedback, and suggests that the presentation format of automated guidance—not just its content—shapes motivational outcomes.

Several limitations frame how far these conclusions extend. The sample consisted of second-year undergraduates at a single institution taking an elective course, and the intervention lasted twelve weeks, so durability beyond a semester remains untested. The outcome measures relied on self-report for the psychological constructs, which are subject to demand characteristics even in cluster-randomized designs. The study also examined one specific system built on one pose-estimation architecture for one movement discipline; systems that score and rank performance, or that generate training plans between sessions, engage learners through different mechanisms and may produce different motivational profiles. Finally, the effect sizes, while statistically reliable, were modest in magnitude, which is typical for educational interventions delivered at the classroom level.

For practitioners, the human–AI division of labor implied by the results is perhaps the most actionable takeaway. Because relatedness was unaffected and teachers remained central to instruction in both conditions, the findings support a complementary model: automated systems handle the high-frequency, precision-oriented corrections during practice, while instructors concentrate on relational support, cultural context, and the motivational scaffolding that machines cannot provide. This division is especially relevant in settings where qualified instructors of traditional movement arts are scarce and transmission has historically depended on face-to-face apprenticeship. A tool that makes progress legible to learners—showing them, frame by frame, that their form is converging on a target—may allow scarce expert attention to be allocated where it matters most.

More broadly, the study illustrates a shift in how educational technology should be evaluated. Comparing post-intervention outcomes between a tool group and a control group answers whether a technology works, but not why it works, and the distinction has design consequences. A system that raises technical scores through external evaluation and social comparison might achieve short-term performance gains while eroding autonomous motivation once the novelty fades. Mechanism-level evidence of the kind reported here—specifying which psychological needs a tool satisfies and which it leaves untouched—offers a template for designing and auditing AI feedback features so that they improve both performance and the quality of the learning experience over time.

Subject of Research: AI Feedback Coach: How Real-Time Pose Detection Builds Motivation and Motor Skill in Traditional PE

Article Title: Effects of AI-mediated real-time feedback on motivation and motor skill learning in traditional ethnic sports education: testing self-determination theory mechanisms

Article References: Qin, Y., Yang, Y., Chen, G., Huang, H., & Wei, D. (2026). Effects of AI-mediated real-time feedback on motivation and motor skill learning in traditional ethnic sports education: testing self-determination theory mechanisms. Frontiers in Psychology, 17, Article 1958348. https://doi.org/10.3389/fpsyg.2026.1958348

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1958348

Keywords: artificial intelligence, real-time feedback, motor skill, motivation, traditional PE, competence, autonomy, Effects, AI-mediated, real-time, feedback, motor

The research team, led by Jingwen Xiao and Xiaowei Zhou of Wuhan Sports University alongside colleagues from Central China Normal University, set out to settle debates that have simmered in the sport psychology literature for decades. Previous meta-analyses had suggested that talking to oneself helps, but they disagreed on the details — on whether instructional phrases beat motivational ones, on whether saying “I” or “you” to yourself matters, on whether the words should be spoken aloud or kept silent, and on how the demands of the motor task shape which kind of self-talk works best. The new analysis, drawing on randomized controlled trials identified through searches of seven electronic databases, was designed specifically to test these emerging and previously established moderator variables and to clarify how they interact with one another.

The scale of the evidence base is meaningful if not enormous. After applying predefined inclusion and exclusion criteria following PRISMA guidelines, the team retained 33 eligible randomized controlled trials encompassing 56 separate effect sizes and a pooled sample of 1,848 participants. That these were randomized controlled trials — the gold standard for causal inference — matters enormously. Participants were assigned to self-talk interventions or control conditions by chance, which means the observed differences in performance cannot be easily explained by pre-existing differences in motivation, skill, or confidence. The analysis was conducted with a three-level random-effects model using the metafor package in R, a statistically sophisticated approach that acknowledges a crucial structural reality of meta-analysis: multiple effect sizes drawn from the same study are not independent of one another, and ignoring that nesting inflates the apparent certainty of the findings.

The headline result is striking in its clarity. Across all included trials, self-talk improved sport performance with a pooled effect size of Hedges’ g equal to 0.52, with a 95 percent confidence interval spanning 0.35 to 0.68 and a p-value below 0.001. In practical terms, an effect of this magnitude means that the average athlete receiving a self-talk intervention performed better than roughly 70 percent of athletes in control conditions — a difference comparable to many physical training interventions that athletes and teams spend considerable money pursuing. Hedges’ g is a corrected version of the standardized mean difference, adjusting for small-sample bias, which makes this estimate more trustworthy than a raw Cohen’s d would be in a literature populated by modestly sized experiments.

But the true novelty of this analysis lies in what it found — and did not find — among the moderators. The most consequential moderator to survive statistical scrutiny was self-talk content itself, which significantly influenced the size of the performance benefit at p below 0.05. This distinction maps onto a classic framework in the field: instructional self-talk, consisting of cue words that direct attention to technique and movement execution, such as “elbow up” or “follow through,” versus motivational self-talk, consisting of phrases designed to energize, build confidence, and sustain effort, such as “I can do this” or “give it everything.” The significant moderation effect confirms that what you say to yourself is not interchangeable — the words must be matched to the psychological function they are meant to serve, and some content strategies reliably outperform others.

Perhaps more surprising is the list of variables that failed to reach significance. The moderating effects of motor skill level, motor demands, overtness of the self-talk, personal pronoun use, application situation, and intervention duration were all statistically non-significant at p above 0.05. This is a genuine challenge to received wisdom. The influential matching hypothesis in sport psychology has long held that instructional self-talk works better for fine, precision-based tasks while motivational self-talk works better for gross, strength- and endurance-based tasks. The new analysis tested the interaction between motor demand and content directly and found it did not reach significance. Similarly, the interaction between motor demand and overtness — whether the self-talk was spoken aloud or performed silently — also failed to achieve statistical significance. The message is that the broad strokes of self-talk may be more forgiving than theorists have assumed: athletes appear to benefit across skill levels, task types, and delivery formats, so long as the content is appropriately constructed.

Methodological rigor was a central selling point of the update. Study quality was assessed using the Risk of Bias 2.0 tool, the current standard from the Cochrane collaboration for randomized trials, which evaluates bias arising from the randomization process, deviations from intended interventions, missing outcome data, measurement of outcomes, and selective reporting. The certainty of the evidence was further evaluated using GRADEpro, the grading framework that rates confidence in effect estimates across a body of evidence. The authors explicitly frame their contribution as providing improved methodological rigor relative to earlier syntheses, and the three-level modeling with restricted maximum likelihood estimation represents a technical step forward in handling the dependence among effect sizes. A likelihood ratio test framework was employed in the moderator analyses, ensuring that comparisons between subgroups accounted for the hierarchical structure of the data.

Yet the analysis is not without an important caveat, and the authors are commendably upfront about it. Publication bias diagnostics revealed a possible small-study effect, with Egger’s test returning a p-value below 0.001. In plain language, this pattern suggests that smaller studies in the self-talk literature may be more likely to report larger effects — a signature consistent with publication bias, in which null or weak findings languish in file drawers while positive results reach print. If that is the case, the pooled estimate of 0.52 may represent something of an optimistic ceiling, and the true effect could be somewhat smaller. The authors explicitly caution that the results should be interpreted with this in mind, a level of transparency that strengthens rather than weakens confidence in the overall enterprise.

The practical implications ripple outward from the laboratory to the field, the gym, and the starting line. Because the moderators that many practitioners obsess over — pronoun choice, overtness, duration, skill level — showed no significant effects, the barrier to entry for self-talk training appears low. An athlete does not need a precisely calibrated script or weeks of graduated practice to benefit; what matters most is choosing content that fits the intended function, whether that is refining technique with instructional cues or summoning effort and confidence with motivational phrases. For coaches, this suggests that embedding brief self-talk cues into technical instruction and pre-performance routines is an evidence-supported practice, not a superstitious flourish. For recreational athletes, the finding that benefits emerge across skill levels is quietly democratizing: the mental tool is not reserved for elites.

The authors close with a call that is both standard and necessary: future research should include more high-quality studies to further examine how relevant moderators shape the effect size of self-talk interventions. The unresolved questions are real — the non-significant interaction tests may reflect genuine null effects or simply insufficient statistical power within the available trial pool, and the small-study effect flagged by Egger’s test demands larger, preregistered, rigorously blinded trials to confirm the true magnitude of the benefit. Still, the core conclusion stands on firm ground. With 33 randomized controlled trials and nearly two thousand participants pointing the same direction, the science now says what athletes have intuited all along: the voice in your head, given the right words, is a genuine performance enhancer — free, portable, and always available on demand.

Subject of Research: Effects of self-talk interventions on sports performance and their moderator variables

Article Title: Emerging moderators in effects of self-talk on sports performance: an updated meta-analysis of randomized controlled trials

Article References: Xiao, J., Zhou, X., Wang, N., Li, J., Zhu, J., Li, J., & Hong, X. (2026). Emerging moderators in effects of self-talk on sports performance: an updated meta-analysis of randomized controlled trials. BMC Psychology. https://doi.org/10.1186/s40359-026-05674-0

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05674-0

Keywords: self-talk, sports performance, meta-analysis, sport psychology, randomized controlled trials, instructional self-talk, motivational self-talk, moderators, publication bias, motor skill, Hedges g, BMC Psychology