AI coach for marathon training

A marathon plan is just a long chain of small decisions: how far to run on Sunday, how hard to push on Tuesday, when to back off before your knees start complaining. Bloomberg reporter Derek Wallbank handed those decisions to ChatGPT for six months before the April 12, 2026 Paris Marathon. (bloomberg.com) He started on October 18, 2025 after moving from Singapore to California, gaining weight, and realizing his April race was getting close fast. He told the chatbot to act as a running coach and nutritionist and gave it years of Strava workout logs, scale readings, diet details, stress triggers, and injury history. (bloomberg.com) The machine’s first job was not magic. It did the kind of sorting a decent spreadsheet can’t quite do on its own: it pulled through 288 Strava activities from the prior two years and turned them into a weekly routine with gym work on Monday and Friday, intervals on Tuesday, a Parkrun five kilometer benchmark on Saturday, and long runs on Sunday. (bloomberg.com) That is the part artificial intelligence is good at in endurance sports. A large language model can turn a messy pile of logs, dates, and goals into a neat calendar in seconds, which is why outlets like Outside have found these tools fast, cheap, and convenient compared with one-on-one coaching. (outsideonline.com) The catch is that marathon training is not just calendar math. A March 2026 review in the British Medical Bulletin said marathon plans need evidence-based pacing, training volume, and periodization, and it asked whether artificial intelligence can reliably prescribe those pieces for demanding endurance events. (academic.oup.com) That tension showed up in Wallbank’s results. He wrote that the system helped him lose 20 pounds in six months and sharpened his routine, but it still needed constant human steering because prompts had to be refined and advice had to be checked against real life. (bloomberg.com) Human coaches do something the chatbot still struggles with: they notice the stuff that never makes it cleanly into a prompt. Outside’s panel of runners and coaches said artificial intelligence still falls short on adaptability, emotional support, and understanding the full picture of an athlete’s life when work stress, bad sleep, or a sore ankle starts changing the plan. (outsideonline.com) That is why the running app market is splitting into two layers. Strava said on April 17, 2025 that it would acquire Runna, a British app built around personalized training plans, because runners want software that does more than store miles after the run is over. (strava.com) Wallbank’s experiment lands in the middle of that shift. The chatbot worked best as a tireless planner that could digest logs, suggest structure, and keep nudging him back toward the next workout, and it worked worst at replacing the judgment of a coach who knows when a plan on paper has stopped matching the body on the road. (bloomberg.com) So the mixed verdict was not that artificial intelligence failed. It was that twenty-six point two miles is still a stubbornly physical problem, and software can organize the training better than most people can organize themselves, but it still cannot feel your legs for you. (bloomberg.com)