AI coach trial: 6 months

A Bloomberg reporter gave ChatGPT a six-month job on October 18: get him ready for the Paris Marathon on April 12, 2026, which is 26.2 miles. Six months later, he reported 20 pounds lost and a verdict that boiled down to pain, progress, and a lot of manual correction. (bloomberg.com) (schneiderelectricparismarathon.com) (britannica.com) The basic pitch of an artificial intelligence coach is simple: you type in your age, goal race, injuries, schedule, and workout history, and the software spits back a plan in seconds. That is cheaper and faster than hiring a human coach, but it also means the plan is only as good as the data and follow-up you feed it. (bloomberg.com) Marathon training is not just “run more.” The Centers for Disease Control and Prevention says adults need at least 150 minutes of moderate activity and 2 days of strength work each week, and marathon plans usually stack much more running stress on top of that baseline. (cdc.gov) That extra stress is where the experiment got useful. Bloomberg’s account says the system could build schedules and adjust workouts, but it still needed the runner to notice soreness, fatigue, and calendar problems the model could not reliably feel or foresee. (bloomberg.com) In practice, the artificial intelligence worked best like a very fast planner. It could organize mileage, suggest nutrition, and respond to new prompts quickly, which is helpful when a race is six months away and every missed week changes the math. (bloomberg.com) It worked worst at judgment. A human coach can hear strain in your voice, notice when you are hiding an injury, or tell when a “bad workout” is really illness, stress, or lack of sleep; a language model only sees what you type into the box. (bloomberg.com) That gap matters because marathon plans are built on recovery as much as mileage. Sports medicine guidance from UCLA Health says recovery after heavy running is part of the training itself, not an optional extra, which is exactly the kind of gray-area call software still struggles to make on its own. (uclahealth.org) The Bloomberg test also points to a more ordinary truth about artificial intelligence fitness tools: they are easier to trust when the task is structured. “Build me next week’s runs” is a cleaner problem than “tell me whether this knee pain is harmless,” and the second question is the one that can ruin a race. (bloomberg.com) So the six-month result was not a robot coach replacing a person. It was closer to a tireless assistant that can draft plans, summarize choices, and keep you moving, while the runner still has to supply the one thing the software does not have: a body that can tell when “push through” is smart and when it is a mistake. (bloomberg.com)