Abstract Background Marathon running has evolved into a global phenomenon, with rising participation across age and experience groups. Training for a marathon requires adherence to well-established principles involving pacing, training volume, and periodization. With the increasing integration of artificial intelligence (AI) into healthcare and fitness, it remains unclear whether AI can reliably prescribe evidence-based training programs for such demanding endurance events. Sources of data We conducted a descriptive study using outputs from leading AI models: Claude 3.5 Sonnet, Claude 3.5 Haiku (Free), ChatGPT 4.0 (o-model), ChatGPT 0.1, ChatGPT 4 (free), Gemini 2.0 Flash, Gemini 2.0 Flash Thinking, and DeepSeek R1. Each was prompted to generate a 6-month marathon training plan tailored to three athlete levels: Beginner, Intermediate, and Advanced. Outputs were compared with peer-reviewed literature on the determinants of marathon training. Areas of agreement Most AI systems identified key training components: weekly mileage progression, tapering, and intensity distribution (80% at low intensity), which aligns with current endurance training theory. Areas of controversy AI responses varied in accuracy and completeness. Some engines omitted key details (e.g. weekly mileage), failed to differentiate clearly between athlete levels (intermediate and advanced have been merged as if they were the same level), or offered inconsistent pacing data, especially for advanced runners. This descriptive analysis evaluated qualitative adherence to evidence-based training principles rather than quantitative outcomes requiring statistical inference. Growing points AI demonstrates strong potential in accessible, structured training content. When properly prompted, outputs often align with contemporary training principles, though significant limitations regarding personalization and professional oversight necessitate further validation before clinical implementation. Areas timely for developing research Future studies should evaluate the real-world outcomes of AI-generated programs in randomized trials including the integration of personal physiological data. Inizio moduloFine modulo.
Montaruli et al. (Fri,) studied this question.