Contextualizing AI-Supported Emotion Regulation Through Sport and Exercise in Higher Education: A Conceptual Reframing

Synthesis: Zhang, Hadier, Liu and Guo advance a conceptual reframing rather than an empirical study: the paper collects no participant data and tests no intervention. Using the five Stanford AI Index Reports (2021–2025, 1813 pages) as a bound AI discourse corpus, the authors ran BERTopic over paragraph-level segments and read the resulting topic structure alongside sport and exercise psychology. They describe a three-layer discourse — a technology core, an application-expansion layer, and an Ethics-and-governance layer — in which health and education are visible but sport, exercise, and campus ecology appear only diffusely. From that reading they propose a human–technology–environment framework for AI-supported emotion regulation in higher education. The framework is a set of design principles awaiting empirical testing; the corpus cannot show that any AI-supported exercise intervention improves student outcomes.

Key Findings

• The contribution is conceptual and agenda-setting. The authors state that the study "does not test intervention effects" and offers "no direct empirical evidence that AI-supported sport or exercise interventions improve emotion regulation outcomes."
• The corpus is five reports totaling 1813 pages — the Stanford AI Index Reports from 2021 through 2025 — deliberately bounded as high-level AI discourse, not the full research literature on emotion, sport, or student mental health.
• Topic modeling produced 22 topics, of which 13 were retained for interpretation after clusters unrelated to sport, exercise, education, and health were excluded.
• Retained topics formed three layers: a technology core (training/data, computing disciplines, AI performance), an application-expansion layer (medicine/health, higher education, computer science education, school pathways, skills/jobs, economy/industry, IT/assistive design), and an ethics-and-governance layer (ethics, policy, research/government coordination).
• Health and education themes grew across the reports, with medicine/health rising from 14 to 105 and school pathways from 3 to 105 segment occurrences between 2021 and 2025; the authors note these are raw frequencies, so growth may partly reflect report length or added chapters.
• Sport, exercise, embodied activity, and campus support appeared only diffusely — distributed across health- and education-related fragments rather than forming a distinct topic.

What the AI Index corpus shows

The authors treat the reports as a structured reading of discourse rather than hypothesis testing. BERTopic with multilingual sentence embeddings (paraphrase-multilingual-MiniLM-L12-v2), UMAP, HDBSCAN, and class-based TF-IDF was applied to cleaned paragraph segments, then dynamic and hierarchical analyses described topic prominence by year and topic relationships. Medicine/health (172 segments, 5.8%) and higher education (82, 2.8%) were among the larger retained topics, while sport, exercise, peer interaction, and campus ecology appeared as a diffuse set of 67 segments (2.3%) rather than a coherent theme.

The authors are careful about what this can and cannot support: a medicine/health topic indicates health relevance, but not necessarily sport or student mental health, and an education topic does not automatically mean campus wellbeing. The absence they report is an absence within this corpus, not evidence that the wider literature ignores sport and exercise.

Why detection is not enough

The paper's conceptual core is that emotional data do not speak for themselves. A single expressive channel — text sentiment, facial movement, voice tone — cannot reliably identify distress, effort, embarrassment, fatigue, concentration, or strategic self-presentation without context. The authors name this "contextual under interpretation": treating an emotional signal as if its meaning were fixed independently of the person, task, culture, and situation that produced it. A student who looks anxious in a sport class may be responding to performance evaluation, peer comparison, unfamiliar movement demands, physical fatigue, or fear of injury, and an AI system that classifies the visible signal while ignoring these conditions risks recommendations that are technically plausible but pedagogically wrong.

Their answer is to reposition AI as a contextual support tool rather than an autonomous interpreter of students' inner states, distinguished by four senses of support: contextual (reading emotion against task, body, relations, place, and timing), embodied (regulation through action, arousal, and interoception, not only verbal reflection), socially integrated (peer climate, teacher–student interaction, campus services), and ecologically grounded (person–task–environment rather than an isolated user–system dyad).

The human–technology–environment framework

The framework holds that students, teachers, coaches, counselors, and sport psychologists remain the primary agents of interpretation and care, while AI supports pattern recognition, documentation, visualization, and timely prompts, and the environment supplies task structures, peer relations, spaces, routines, and governance. It has five design pathways:

1. Multimodal AI contextual profiling — integrating self-report, participation, subjective exertion, workload, sleep, and contextual notes, limited to what is necessary and consented.
2. Autonomy-supportive task adaptation — adapting intensity, task complexity, group size, feedback frequency, and recovery options without turning exercise into algorithmic prescription or coercive monitoring.
3. Feedback–reflection–practice loops — using visualizations to help students notice patterns between activity, stress, mood, and coping, Scaffolding rather than replacing their judgment.
4. Peer and campus support integration — embedding tools in physical education, sport clubs, counseling services, and referral pathways instead of isolated apps.
5. Human-in-the-loop governance — keeping teachers, coaches, counselors, and sport psychologists responsible for interpreting outputs, spotting misclassification, and protecting students from harmful feedback.

The pathways are grounded in Self-Determination Theory (autonomy, competence, relatedness as conditions for sustained Motivation), ecological dynamics (behavior emerging from individual, task, and environmental constraints), and social-ecological accounts of campus life. The authors insist the pathways are design principles requiring empirical testing, not established causal mechanisms.

Where the risks sit

Because the area combines emotional data, student populations, educational authority, and behavior guidance, the authors foreground governance before technical optimization. They list seven issues: privacy (emotional, behavioral, and physiological data treated as sensitive), consent, data minimization, transparency about uncertainty, bias across gender, culture, disability, language, and Neurodiversity, autonomy (recommendations that do not become coercive exercise prescriptions), and human oversight. The failure mode they name is the transformation of support into surveillance — monitoring becomes coercive if emotional analytics are tied to attendance, performance grading, disciplinary decisions, or institutional risk management.

What this means for practice

• Keep the human in the loop. Treat AI output as a prompt for professional judgment — teachers, coaches, and counselors interpret emotional patterns and decide on referral, and no automated diagnosis should stand in for professional review.
• Design for Learner Agency and choice. Adapt tasks, group sizes, and recovery options in response to observed need, and offer alternative activity formats rather than uniform prescriptions when anxiety or fatigue appears.
• Put governance before features. Decide data minimization, consent, opt-out routes, and retention rules first; use aggregated wellbeing dashboards for service planning rather than individual surveillance.
• Embed tools in the campus ecology. Connect anything you deploy to physical education, sport clubs, counseling, and peer-led activity so it strengthens Well-Being through belonging and enjoyment rather than isolating students in an app.
• Treat the framework as a hypothesis. Adopt these pathways as design principles to be tested, not as evidence that AI-supported exercise already improves emotion regulation.

Limitations

• The corpus is narrow: five annual AI Index Reports represent a high-level AI discourse, not the full AI, sport and exercise psychology, student mental health, or intervention literature, and conclusions about absent themes refer to that corpus only.
• Topic modeling is exploratory and sensitive to segmentation, preprocessing, embedding model, dimensionality reduction, and clustering choices, and the dynamic analysis rests on five reports, so it is not a formal longitudinal trend analysis.
• The framework is interpretive: the five pathways were informed by the topic model but developed mainly through theory-informed synthesis.
• The study tests no intervention and provides no direct empirical evidence that AI-supported sport or exercise improves emotion regulation; the supporting literature was not reviewed systematically, so the discussion is conceptual integration rather than exhaustive evidence synthesis.

Citation

Zhang, Y., Hadier, S. G., Liu, Y., & Guo, Y. (2026). Contextualizing AI-Supported Emotion Regulation Through Sport and Exercise in Higher Education: A Conceptual Reframing. Behavioral Sciences, 16(7), 1173.