Can ChatGPT Train Me for the Boston Marathon?

Editor’s Note: the featured image was made by AI tool Magic Media, by Canva.

Preparing for a marathon is as much a science as it is an art form. Each training buildup is like a unique case study that includes data, timelines, and variable factors such as fatigue, recovery, fueling, dietary changes, and weather conditions. Much like in a laboratory setting, a study’s conclusion (or in the instance of marathon prep, the most effective training strategies) are revealed over time through trial and error.

Due to the complexity and intensity of marathon training, elite athletes and committed age-groupers typically hire coaches, exercise physiologists, and nutritionists to oversee all the preparation, from collecting data to monitoring and adjusting workouts throughout the process. While this results in a fine-tuned regime, the average, everyday runner can’t always afford to have a team at their beck and call.

As an exercise physiologist, I am one of those team members who would be spending countless hours deeply involved in developing data-driven training plans. One day, I was asked if I thought Artificial Intelligence (AI) would eventually threaten my job—a thought that elicits a mix of intrigue and apprehension. The advent of AI in performance training is undeniably transformative, particularly in data analytics. Traditional coaching platforms and professional organizations like Barcelona FC are also adopting AI to analyze player performance data. Even Team USA, U.S. Soccer, and USA Surfing are incorporating AI to prepare for the Paris 2024 Olympics.

However, the democratization of AI-enabled training tools raises questions about accessibility and customization for non-elite athletes who want the training plans without the time and cost barriers. While wearable devices and apps like Strava offer a wealth of personal data, translating this information into actionable insights remains challenging. The key lies in generating high-quality training plans and delivering precise coaching cues and adjustments tailored to individual needs.

Before I turn my job over to artificial intelligence, I decided to dip a toe in these automated waters by engaging a free and widely accessible platform such as OpenAI’s ChatGPT. I tasked it to craft a 20-week program covering endurance training, resistance workouts, flexibility exercises, and nutritional guidance in my own preparation for the 2024 Boston Marathon.

Defining Artificial Intelligence

Before we dive into the results of my experiment, it’s important to define exactly how AI works. Shaun Wallace is a Human-Computer Interaction Systems researcher at my alma mater, the University of Rhode Island. He sheds light on the trajectory of AI platforms explaining that AI has significantly reshaped human work processes across all fields, not just athletics. 

“Tools like ChatGPT will continue to evolve and grow over time,” he explains. “Current research has shown AI can generate an abundance of examples rapidly. The challenge lies in generating the right examples.”

Wallace says that the AI we interact with today has caught up after learning from countless contributions across the web over decades. In other words, AI is a digital apprentice standing on the shoulders of human creators and collaborators, not merely a machine at work but a fusion of human knowledge and digital intelligence.

AI’s involvement in a marathon training program is a great example of how this works, as the machines draw heavily from established training techniques from endurance coaching legends Jack Daniels and Matt Fitzgerald, in particular their pacing and weekly training design.

The Parameters: An AI Experiment Begins

I begin my AI journey by giving ChatGPT the necessary ingredients like the situation, parameters, and goals that make up the launching pad for any solid training regimen. The challenge is clear:

Design a training program for the 2024 Boston Marathon. I need a complete nutrition plan, flexibility, and strength training to reduce the risk of injury. I can commit to 5-7 days of training per week, each session lasting about an hour and a half (disregard for long run days).

And in just 30 seconds, ChatGPT produces a marathon training program spanning more than 10 pages. Impressive, right? 

The Results: The 20-Week Boston Marathon AI Running Plan

It’s important to note that although I got heaps of information in mere seconds, it took several tries and a few corrective prompts to get ChatGPT to produce an adequate training plan. The first regime generated by the AI platform was a general one that provided only basic instructions and limits. Given access to the vast expanse of the internet, it is not surprising that it offered very basic guidance on training, including the widely accepted principle of not increasing the volume of training by more than 10 percent per week. Here is a boiled down version of the AI output:

Weeks 1-4: Base Building

• Running: Gradually increase weekly mileage by 10-15 percent. Aim for 4-5 weekly running days, including easy, tempo, and long runs. Start with a long run of 6-8 miles.
• Strength: Focus on full-body strength workouts twice weekly, incorporating squats, lunges, deadlifts, push-ups, and rows.
• Flexibility: Include dynamic stretches before running and static stretches after each workout.

Weeks 5-10: Building Endurance

• Running: Gradually increase the long run by 1-2 miles weekly. Add a weekly interval session for speed and endurance.
• Strength: Continue with two weekly strength sessions, increasing weights gradually. Emphasize core exercises.
• Flexibility: Consider adding a weekly yoga or Pilates session for enhanced flexibility.

Weeks 11-16: Marathon-Specific Training

• Running: Increase the long run to a peak of 18-20 miles. Add hill training to simulate the Boston Marathon’s course.
• Strength: Maintain two strength training sessions, focusing on exercises that support running mechanics.
• Flexibility: Prioritize regular stretching and foam rolling to prevent muscle tightness.

Weeks 17-20: Tapering and Fine-Tuning

• Running: Gradually decrease mileage for recovery and energy reserve building.
• Strength: Reduce strength training intensity and volume to prevent fatigue.
• Flexibility: Continue with gentle flexibility exercises while avoiding intense stretching to prevent muscle strain.

AI Insights and Personal Evaluation

Even if you are not familiar with the 26.2-mile Hopkinton to Boylston Street course, you’ve probably heard of Heartbreak Hill, a grueling mile-long, uphill stretch and the most punchy ascent of the infamous Newton Hills. It is just one of many elevation changes on a course that has 815 feet of elevation gain over 1,275 feet of loss. Thus, any decent training plan needs to include workouts that prepare runners for these steep inclines.

According to Christie Ward-Ritacco, Ph.D., a certified exercise physiologist and kinesiology professor at the University of Rhode Island, the program is versatile. It takes into account Boston’s hilly course’s unique characteristics. For instance, the workout on April 1 is an excellent example:

Hill Repeats:

• Warm-up jog for 10 minutes.
• Hill Repeats: 4-5 x 30-second uphill sprints with a slow jog back down for recovery.
• Cool down jog for 10 minutes.

“The nuances of a course [are] important to consider,” Ward-Ritacco says. “The downhill portion of the run actually results in a significant amount of muscle damage due to the eccentric nature of the contractions. Incorporating these aspects into your training can help with both race preparation and recovery.” 

As with any new technology, though, AI-powered platforms have limitations, and ChatGPT is no exception. Notably, the AI-program suggests workouts that aren’t, let’s say, realistic. For example, performing five sets of 10,000-meter repeats at a 4:17-minute/mile pace is an extraordinarily demanding workout that would challenge even elite athletes. Each repeat would need to be completed in approximately 42 minutes and 50 seconds to maintain that pace. Here’s a breakdown of the workout:

Set 1: Run 10,000 meters (or 6.2 miles) at a consistent 4:17-minute/mile pace. This requires exceptional cardiovascular endurance, muscular strength, and mental fortitude.

Set 2: After a brief rest period, repeat the 10,000-meter run at the same 4:17-minute/mile pace. The cumulative fatigue will significantly increase the difficulty of maintaining this intense speed.

Set 3: Continue with the third 10,000-meter repeat, focusing on maintaining proper form, hydration, and nutrition to sustain energy levels.

Set 4: By this point, fatigue will be setting in, making it crucial to listen to your body, adjust pacing if necessary, and prioritize safety and proper technique.

Set 5: The final 10,000-meter repeat demands everything you have left, both mentally and physically. Focus on maintaining a consistent pace, managing fatigue, and finishing strong.

Yes, ChatGPT did recognize that it would be difficult, but without proper guidance on alternatives, where does that leave the athlete? This example highlights Dr. Ward-Ritacco’s comments on the critical role of a trained coaching professional in creating a personalized plan for a specific runner’s abilities, monitoring changes during training, and making necessary adjustments—capabilities currently lacking in AI not designed for running.

“A seasoned and trained coaching professional is capable of recognizing changes in your attitude, times, and even gait and running mechanics and can make the necessary adjustments,” she says. “While an AI program may be useful in modifying plans, it is limited by the information you provide to the algorithm and model. On the other hand, an individual whom you interact with, even if it is virtual, can serve as a sounding board for discussing any necessary adjustments.”

Mark Hartman, PhD, CMPC, a professor of kinesiology at the University of Rhode Island, believes integrating artificial intelligence into training can indeed be beneficial. Hartman, a former collegiate track and field athlete turned trail ultrarunner, ran an impressive 2:44:27 in Boston this past spring. He thinks ChatGPT’s generated training plans could be improved if connected to the athlete’s smartwatch. The AI could monitor heart rate response and pace while considering environmental conditions and elevation changes during workouts. The AI could then track the athlete’s progression across the training plan’s duration and make real-time adjustments. 

“The efficacy of an AI training plan (even one that is monitoring cues and making real-time adjustments) is still limited by the data inputs and the algorithm that the AI is following when making ‘decisions’ on how to use the data to generate recommendations to the athlete,” he says. Hartman points to the program’s response to specific parameters, such as goal finish time and pace, is not always satisfactory. “AI models may still need to be designed and guided by trained professionals in exercise prescription and coaching.”

Expert Insights: Personalized Nutrition

Let’s start by picking through the massive (and mostly nonsensical) amount of nutritional guidance AI provided.

Jordan Hill, MCD, RD, CSSD, lead registered dietitian at Top Nutrition Coaching, offers crucial insights about why having personalized (perhaps human) nutrition is critical. According to Hill, emphasizing the depth of personalization required to design athletes’ nutrition plans is crucial. 

“There’s a significant amount of personalization that goes into designing nutrition plans for athletes,” she remarks.

The baseline knowledge of standard nutrition education that AI seems to have is just that—a baseline. The real work lies in the unique application of this knowledge, tailored to each athlete’s specific needs. Factors such as current health status, medication use, hydration needs, supplement needs, and overall lifestyle form the wide range of considerations that have to be accounted for.

My AI plan, however, falls short of such specificity, offering general recommendations without key details, such as what exactly a “handful of seeds” or “tropical berries” mean. Hill points out the absence of other important particulars. “While the sample menu includes consistent and balanced meals and snacks,” she says, “it fails to provide the portion sizes and suggested timing of consumption.” 

Jordan suggests that athletes leverage AI-generated nutritional guidance only as a starting point. She underscores the value of seeking professional consultation for a personalized nutrition plan. The expertise of a sports dietitian becomes essential  in ensuring that recommendations align precisely with individual needs and goals.

Cautious Optimism

With AI-powered training plans becoming more mainstream, experts worldwide have weighed in on their potential impact. The consensus? Cautious optimism. 

While artificial intelligence can provide valuable insights, it’s not a perfect solution. As Wallace points out, these tools may struggle to answer questions or provide helpful information on recent events and data. As of this publication, the free version of ChatGPT, for example, hasn’t been fully updated since January 2022, which means it’s not up-to-date with the latest research and protocols. This severely limits innovation by not having access to cutting-edge research and protocols to design the best training systems possible. 

Experts like Dr. Ward-Ritacco caution against relying solely on AI for any athletic goal.

“If you’re using an AI system, one might need to regularly augment the information put into the algorithm to get an appropriate guide (update mile times, mileage covered, muscle soreness reports),” she says. 

Bottom line? AI-generated plans can be helpful, but they can’t replace the human touch. Dr. Hartman agrees, pointing out that AI can’t monitor and respond to an individual’s subjective training experience. 

“Is the training plan something that the person finds enjoyable or motivating versus unpleasant, terrible, or unmotivating,” he says. “AI, as it currently exists, cannot help with that.” 

So, what’s the solution? According to Hill, live communication and human interaction is vital to creating and implementing a dynamic training program. 

“Communicating live with clients allows me to read body language, facial expression, and observe emotion—factors that help me individualize my approach, problem-solve challenges, and personalize care,” Hill says. 

All that being said, will I put the fate of my Boston Marathon dreams solely in ChatGPT’s servers? Not a chance. 

Next-Gen AI Coaching Is Already Here

The future promises a collaborative approach where AI technology enhances, not replaces, human experience, dedication, and creativity. New platforms like TrainAsONE, Athletica, AIEndurance, and Humango have been working on leveraging AI to analyze extensive data sets, generating next-level, cost-effective, and user-friendly training programs. 

The Humango platform, which launched two years ago, has incorporated several levels of AI assistance within its training mobile app platform that’s aimed at guiding, adapting, and motivating an individual to improve performance. It starts by inputting a user’s metabolic profile and availability for training and filtering it through a training strategy.

That’s the basic level that any level of AI or human coach can provide. But the next level of coaching, either from a human coach or technology, needs to be able to adapt to how an individual is progressing through the training—managing things like workout efficacy and fatigue, but also enthusiasm and confidence, says Eric Abecassis, founder and CEO of Humango.

By integrating machine-learning technology known as the Large Language Model (LLM), aimed at detecting patterns and responding, Humango has implemented a personal coaching assistant feature called Hugo that is capable of creating, suggesting custom adaptations, and even having conversational interactions with the user. Hugo can help schedule a day off, increase or reduce the intensity of workouts, adapt to an injury or sickness, and ultimately make the fitness more personable, something that Abecassis believes is a tool that can steer emotional and motivation energy in a positive direction.

Abecassis says that, while ChatGPT scrapes information from the Internet and averages out known training data to create a training plan, Hugo is capable of understanding who you are, what you are training for, how you are progressing, and answer questions with a curated base of knowledge.

“It can actually do a pretty decent job at telling stories, which up until now is something that a human coach could do best,” Abecassis says. “We’re trying to create kind of an emotional connection with the athlete to enhance motivation. It’s still going to be a machine and that will be a limit for certain people there, but it’s not purely robotic either. It represents a bit of smoothing and rounding the edges of the technology that is available and also how technology can be perceived by people.” 

The Future

Will AI-assisted training become more so prevalent it will make the coaching profession obsolete? Highly unlikely. But it will be able to help coaches become more accurate and proficient at how they monitor and guide their athletes.

As the world of running evolves, so too will technology. It’s up to runners to strike a delicate balance between integrating AI-generated plans with human guidance, just as they have needed to manage their use of carbon-plated supershoes, advanced nutrition supplements, and other aspects of running and racing that evolved in recent years. The goal is to continually push the boundaries and drive athletes to new heights and better performances, while properly implementing whatever tools, technologies, and resources are available as part of that process.