If you’ve spent any time on the internet in the last two to three years, you’ll be familiar with the four-word prompt of ChatGPT: "Where should we begin?".
Launched in November 2022, OpenAI’s artificial intelligence chatbot started an arms race in the world of generative AI, with Microsoft, Google and Meta all releasing their own iterations in the intervening years.
On first use, it appears to be nothing more than a glorified search bar or a means to write overly formal emails to complain to utility companies. But spend enough time with the technology and refine your prompts, and you can unlock its full potential.
This extends to the world of fitness, where the rapid integration of AI appears to be gunning for the jobs of coaches and personal trainers.
Apps such as Strava’s recently acquired Runna use AI to analyse and monitor if you’re on track with a training plan, suggesting adjustments if it thinks you’re off the pace; users of HumanGo AI can call upon their own AI chatbot, Hugo, to create and modify training plans, while answering all the questions you'd ask a real-life coach.
But how good is generative AI at creating training plans? And will it actually be able to replace a coach? I fired up ChatGPT to find out.
- Read more: How is AI impacting cycling? Five ways Artificial Intelligence is changing the way we ride
OK computer
As a seasoned marathon runner with a 2hr 43min PB and years of race-specific blocks under my belt for running, cycling, swimming and triathlon, I know my way around a training plan.
From free guides found on the internet to a bespoke schedule drawn up and updated weekly by a coach for when I tackled the Paris-Roubaix Challenge and Manchester Marathon in the space of eight days in 2024, I've also experienced the sizeable difference between a generic, one-size-fits-all plan where you're left to your own devices and a tailored and structured system where a qualified professional is on hand to guide you through each session and a WhatsApp message away to answer any questions.
To put ChatGPT through its paces, I thought I’d start with a fairly generic prompt that would be the goal of a lot of beginner-to-intermediate runners: "Could you create me a sub-4 16-week marathon training plan? I currently run 3-4 times per week."
Its response was a lot longer than I expected, including plan goals, a breakdown of the sort of things I’d be doing each week, the 16-week plan itself, explainers of key workouts (such as pacing targets and interval instructions) and general tips on warming up and cooling down, listening to my body, hydrating and fuelling properly and considering a half marathon tune-up in weeks 8-12.
Each week featured four runs – three easy and one hard, with a long run on a Sunday that built up to 18 miles. There was also a recovery week every fourth week. So far, so good. Delve a bit deeper, though, and things started to look a bit basic.
While the plan followed a periodised, polarised approach, the hard day itself only tasked me with running at two different intensities – tempo and 5K pace – throughout the entire 16 weeks.
Anaerobic work, such as sprints and high-intensity interval training, was completely absent. There was also a lack of strength and conditioning work, despite the known scientific benefits that combining running-specific strength training and endurance training can have on everything from body fat percentage and lean mass to running economy and VO2 max.
“It's very generic, it's very repetitive, and when I saw this programme, my first thought was ‘that is so boring’,” says Faye Johnson, running coach and level 4 PT at Maximum Mileage Coaching.
To keep things interesting, Johnson says she would have employed a selection of different sessions depending on the particular training block, while also gradually introducing different types of speed work – “whether that be a progressive run, maximal intervals or threshold intervals, or, just below that, so tempo” – adding that it’s something that beginners might not have done a lot of.
“There's also not a single hill session in here. I obviously wouldn't overdo it, but if someone's doing an undulating marathon, or even a relatively flat one, hills are really good for strength and it translates into better speed when you go back to a flat route.”
Coach in the machine
Johnson concedes that the plan itself wasn't bad and would probably achieve the desired goal, explaining that the structure and weekly breakdown were nice additions.
ChatGPT is able to pump this out in seconds because the core of its output is based on any publicly available training plan on the internet.
Trained on a near-infinite amount of text and data, it’s no surprise that the plan has structure, science and coaching principles at its heart.
When asked, it responded that it had based the plan on principles from “several reputable coaching sources and methodologies”, and included “input from RRCA and USATF-certified coaches’ practices”.
“If you know what you're doing, and you’re pretty comfortable with what your pace is – you know what your easy pace is, you know where you are when you're pushing yourself to threshold pace or beyond that – it's a pretty straightforward and easy program to follow,” says Johnson.
She adds that this sort of coaching would probably work for those who are happy working solely off a set plan.
“There are certain people who like to have structure and accountability; they have something to follow and it helps motivate them, keep going and a reason to think, ‘that's on the plan today, I've got to go out and do that’.”
But this isn’t the sole reason people turn to a coach, which is where the cracks in ChatGPT’s coaching abilities start to appear.
Cannot compute
“Some people value that someone has put together a programme for them, and if they miss that session, what are the implications?” says Johnson, who coaches her clients remotely using TrainingPeaks.
“People think, ‘she's going to see exactly what I have and haven't done’. You can't hide anything.” This lack of accountability could leave some with a lack of motivation, while there’s no financial outlay to act as a carrot, like there is when employing a coach.
But there is another benefit to a coach having an overview of each training session that ChatGPT can’t compete with – a human will be able to see if you’re following the plan correctly, even if you can’t see it yourself.
“One of the biggest things for me is that ‘easy run’ is quite vague,” she says. “It translates into so many different things.
"I get people who I know damn well have been puffing during their easy run, so I give them that feedback whereas, if you were following a programme based solely on pace, it doesn’t account for variables like heat, tiredness, etcetera.”
At the extreme end, this can lead to overtraining and injury. “If you put this programme into a particular person's hands, they're going to wear themselves out and not know when to rest or pull back and subsequently pick up an injury.
"There are a couple of people that I'm coaching who don't know when not to train. I'm almost there to tell them to back off because they're notoriously always doing too much.”
To see what would happen if I did get a sore calf three weeks in, I went back to ChatGPT with my theoretical injury.
Its response was to replace the speed work with an easy run or cross-training, advise me not to run through the pain, provide a treatment and recovery plan, tips on checking my shoes weren’t worn out, and flag any symptoms that meant it was time to visit a physio.
While all sound and sensible advice, the lack of analysis means that, unlike a coach, ChatGPT would always struggle to find the root cause of my sore calf; even if the tightness did ease, there’s no reason it wouldn’t occur again weeks down the line.
Trust issues
And that is the major flaw with using ChatGPT as a coach – it relies on a lot of heavy lifting from the user: whether that’s coming up with a tailored, specific prompt to get the most accurate training plan for your goal event; following the schedule with an understanding of what it’s asking you to do and how to complete each workout to achieve the desired aims; and knowing when you might need said plan adapting because it’s too easy, too hard or you’ve picked up a niggle.
Even if you’re one of those people who can do all this, there is an underlying need to trust the generative AI’s output – something that can be difficult when the plans it creates are free and the technology itself has been around for only three years.
“As an individual, I don't know that I would default to AI – maybe in part because I don't know how to use it to its best, but there is a trust element to it,” concludes Johnson.
“How can an AI platform account for all the little nuances that you have in regards to an activity plan of any sort?”
For me, until it’s able to replicate the insight and feedback of a real-life coach and back it up with actual credentials and testimonials to boot, I’ll be restricting ChatGPT’s use to life-admin tasks.
