Synthesizing the Expert: A Validated Multimodal Dataset for Trustworthy AI-Assisted Swimming Coaching

AI coaching systems need grounding in expert knowledge: The paper demonstrates that AI-generated coaching advice is more trustworthy when connected to authoritative domain expertise rather than relying solely on general language models. The researchers validated 1,864 question-answer pairs against 12 physiological soundness rules to ensure accuracy.

Ethical data challenges require creative solutions: Rather than collecting sensitive biometric data directly from athletes (which raises privacy concerns), the researchers synthesized training data from multiple sources including physiological literature and sensor data, addressing a key tension between AI development and athlete privacy protection.

Validation against expert standards is essential: The paper emphasizes that AI systems for high-stakes domains like sports coaching require rigorous evaluation—their dataset went through multiple rounds of expert review, with nearly 1,914 drafts evaluated to produce 1,864 validated outputs, highlighting the importance of quality control over quantity.

Multimodal data improves AI reliability: By combining four different data types (physiological data, literature, sensor readings, and expert knowledge), the system produces more contextually appropriate and technically sound recommendations than systems relying on single data sources.

Responsible AI requires domain expertise integration: The research shows that trustworthy AI in specialized fields like sports science cannot be achieved through generic AI models alone—it requires deliberate collaboration with subject matter experts to ensure generated advice is both credible and safe for real-world application.