Artificial intelligence coaches on the sidelines: Evaluating readability and quality of soccer training plans from six generative models

Abstract

Background and Objectives: Generative Artificial Intelligence shows increasing promise for developing personalized training programs in sports science. While previous research has demonstrated its utility in aerobic and resistance training, its effectiveness in generating structured, sport-specific plans for technically demanding, high-contact team sports like soccer remains underexplored. This study aims to assess the readability and quality of soccer training programs generated by six leading generative AI (GAI) models— GPT-4o, GPT-4.5, GPT-o1, GPT-o3-mini, and DeepSeek-R1 and DeepSeek-V3—to assess their feasibility for practical use. Methods: Each model was prompted to create a 30-day soccer training plan following exercise prescription principles. Three expert raters assessed its quality using a ten-point custom rubric covering key exercise prescription components. Readability was evaluated using Flesch-Kincaid metrics. Visualizations were generated in RStudio, and inter-rater reliability was assessed via intraclass correlation coefficients (ICCs). Results: All models produced structured, soccer-specific programs with varying narrative styles and recovery protocols. DeepSeek-R1 yielded the most accessible text (Grade Level 8), while ChatGPT-o3-mini produced the most complex (Grade Level 12). ChatGPT-o1 was the most verbose. Quality scores from raters demonstrated consistency (ICC = 0.79), ranging from 6.00 to 7.00. ChatGPT-o1 received the highest rating, followed by DeepSeek-R1. All models shared common limitations, including the absence of citations and lack of individualized health screening protocols. Conclusion: GAI models demonstrate strong potential for generating soccer training plans, though trade-offs exist. However, the outputs require human oversight to ensure safety, individualization, and scientific rigor. Future research should explore context-aware prompt engineering, multilingual applications, and real-world implementation through coach- and athlete-centered trials.