A participant who used ChatGPT as a marathon coach succeeded in losing 9 kg and setting a personal best, but also experienced significant side effects, drawing attention.

On April 13 local time, online media outlet Gigazine cited Bloomberg as saying Derek Wallbank, the person at the centre of the case, used ChatGPT for several months while preparing for this year’s Paris Marathon after setting it up like a training coach and nutritionist.

Wallbank decided to enter the Paris Marathon in April 2025, but after moving to California in the United States his training time fell sharply and he gained weight. With 6 months left until the marathon, he asked ChatGPT to act as both running coach and nutritionist and support him, setting his goal as “finishing the race without injury.”

The initial setup took about 1 hour. Wallbank entered records from the running app Strava, which he was already using, along with his weight, eating habits, stress factors and past injury history. Based on this, ChatGPT advised him to improve pace control, increase upper-body and core strength training, and combine long-distance training with injury management.

He then used ChatGPT to plan a weekly training schedule. It consisted of gym sessions on Mondays and Fridays, interval training on Tuesdays, rest on Wednesdays, other activities such as golf on Thursdays, a 5 km run on Saturdays, and long-distance running of at least 5 km on Sundays. Wallbank logged every workout on Strava and then entered his heart rate and pace data into ChatGPT, running separate chat windows for running, nutrition and weight management.

The results were clear. After he started using ChatGPT, Wallbank lost 9 kg and achieved personal bests in the 5 km and 10 km distances. “Thanks to ChatGPT, I can run farther and faster than before,” he said.

But the limits of an artificial intelligence coach were also clear. The biggest problem was the burden of manual input. Wallbank had to enter his diet and exercise details himself several times a day. He wrote down salad ingredients and amounts one by one, and sometimes even uploaded photos to estimate portion sizes by hand. The more information he entered, the more finely ChatGPT adjusted meals and training, but the user’s logging work grew as well.

After several weeks, the consistency of its output began to waver. For example, something he said once, such as “I don’t have time to go to the gym this Monday,” was reflected as if it were an “always-on rule.” Wallbank thought ChatGPT would keep remembering what he entered like a human coach, but in reality its memory was limited.

About 3 months into training, hallucinations also appeared. “At first it was about correcting errors, but it gradually turned into annoyance,” Wallbank said. When he asked about the day’s plan, it suggested interval training instead of the scheduled workout. At other times it changed his target weight on its own and praised him for managing his weight well even though he had not reached the goal.

ChatGPT was also used to set a pace-control strategy. Wallbank planned to run for 28 minutes to the first water station, then repeat 4 minutes running and 3 minutes walking 6 times, followed by 10 minutes running and 2 minutes walking 3 times. He tested the strategy in a half marathon, but said he had to revise the plan mid-race after encountering a steep uphill climb with an elevation difference of 183 metres. Even so, he finished that half marathon in under 3 hours and recorded his second-best time for the distance.

The case shows that generative AI can have some value as a personalised training support tool, but still has clear limits in fully replacing a human coach. Despite driving results such as weight loss and faster times, the burden of manual input, memory constraints and hallucinations increased fatigue from long-term use. It highlighted both the potential and the imperfections of an AI coach.