Gauge against the machine: Could AI replace coaching expertise?

Will you live to see a Tour de France decided by artificial intelligence? What if I told you that you might have already? Everyone is talking about AI and whether it is threatening to take over the world, but what exactly is it? 

In the most basic terms, it’s computer software that is able to learn patterns from information and then perform tasks that previously required human intelligence. Already it is outperforming humans in a multitude of tasks. But what about its potential in cycling? Can we trust AI to set our training? In other words, is it able to learn the patterns of your body better than you or your coach?

Though still in its infancy, AI is already infiltrating the cycling world. It is used for analysing performance data, predicting future trends, and has even helped develop a yellow-jersey-winning nutrition strategy. Thanks to its ability to rifle through so much data, AI has the capacity to detect trends in training – for example, the correlations between sessions and fitness gains. 

Two early movers, the apps Spoked and Athletica, have been using AI for almost a decade. Using AI to analyse your training is, first and foremost, much cheaper than hiring a human coach.

Coach vs App

Spoked – AKA “the AI cycling coach” – was founded by former JLT-Condor rider Richard Lang in 2016. His ambitious goal was to build an AI coach that would be as good as having a flesh-and-blood coach and physiologist at your disposal at any time of day. Once the user has entered their athletic history and achievement goals, the app ‘talks’ to other platforms, such as Strava and TrainingPeaks, and within minutes produces a weekly training plan.

Lang explains that it is not only data but personal feedback that make Spoked a credible coach-replacement. “We created a training framework that ensures a rider is hitting the demands that align to the goal and training phase. The individual parts of the sessions scale up or scale down depending on other factors such as how much time the rider has available.” The app prescribes workload based on objective physical readiness, claims Lang. “It is all based on a central engine that combines five variables: time and zone percentage, sleep, perceived difficulty, sleep, mental and physical freshness. We understand that there can be errors within data, which is why we bring in that personal feedback too.”

Drawing on personal feedback is the greatest challenge for artificial intelligence coaching. Whereas a good human coach not only sets workouts but also becomes a friend, motivator, therapist and psychologist all in one, an AI app is solely a data-cruncher.

Lang pushes back on this, insisting that Spoked adapts and responds to individual differences. “We’re not just a workout aggregator that slots in workouts,” he says. “We have a hybrid approach, and we still rely on a rider’s feedback through a short health questionnaire after each ride. We’re scanning both a rider’s health and training load to decide which workouts to prescribe.”

What do Spoked-coached riders think of their machine overlord? Tim Rice, a 43-year-old programme manager from Preston in Lancashire, started using Spoked at the beginning of 2023 to prepare for the Mallorca 312 sportive. 

“It’s great that there is the flexibility to change the day, duration and structure of a session as things crop up,” he says, “knowing that the AI will replan, on the fly, for the remaining weeks’ workouts to accommodate for the change. I’ve had a [human] coach before, which was significantly more expensive for a plan that was wholly dependent on them. Spoked uses AI at a low cost to set and tailor the plan.” 

Rice appreciates the comparatively low cost of the app – £9.99 per month for the ‘pro plan’ – and having access to advice from Lang. “I’ve seen improvements in PRs, was able to complete Mallorca 312, and just this past week felt confident enough to enter my first crit,” he adds.

The rival app

Athletica.ai is a Canadian rival to Spoked, which in a similar way prompts users to connect to Strava. Based on a rider’s previous six weeks of training, it then captures what co-founder Paul Laursen calls a “training fingerprint”. Athletica uses Google’s ‘sentiment analysis’ tool to interpret post-ride comments. 

“After your session, we prompt you for your RPE [rated perceived exertion] and comments,” says Laursen, who was the lead performance physiologist at Sport New Zealand for the London and Rio Olympics. “These comments usually capture an athlete’s mood, something a human coach would consider. Over time, our users find that it is as though AI is starting to understand them.”

Artificial intelligence has the power to bring data-analysis to the masses, providing meaningfully individualised sessions based on past performances. Still, this is no guarantee that it can surpass or even match the intuitive and experience- led decisions made by a living, breathing coach. Exactly how well apps like Spoked and Athletica stack up against traditional, tried-and-tested methods will take time and research to uncover.

The usefulness of AI may depend on the rider’s own ability to determine what type of training best suits them based on their interpretation of the signals from their own body. I’ve raced at an international level for the past five years and I like to think I know my own body better than any tracking device or fitness gadget. I often sense instinctively when my power meter needs calibration – my ‘in-built power meter’ lets me know.

The Chat Bot

ChatGPT is an artificial intelligence chatbot founded by OpenAI which accrued over 100 million users in its first two months. It has a breadth of capabilities from answering questions to writing essays by ‘web scraping’ information – in essence, it is a very enhanced form of predictive text. With this in mind, I asked ChatGPT to build me a plan for the final week of training in the lead up to the National 10-mile TT Championships, a big target for the second half of my season.

It was a bizarre experience watching ChatGPT spew out a training plan that appeared, on the face of it, quite carefully and judiciously designed.

Decisions that a coach might labour over for hours were being made by the chatbot instantaneously. That said, some of the sessions set by ChatGPT were questionable. The first red flag was ‘High Intensity Hill Reps’ just three days before the target race.

What would an established coach make of the plan? “In principle, it isn’t bad,” says Matt Bottrill, a multiple national champion and consultant to the Lotto- Dstny squad, “but I would be worried that the midweek intervals might be too much a few days prior to the target race. In a taper week, I prescribe half the number of intervals compared to a normal week.” 

Bottrill and I agree that, although ChatGPT’s plan seems to be broadly on the right lines, it’s probably less than optimal – hardly surprising, given that its recommendations are derived from generic data available on the web rather than my personal training history.

Tech that talks: we sat down woth ChatGPT

Who better to interview about the potential impact of AI than ChatGPT itself? We sat down with the chatbot to grill it on the pros and cons of AI-led training

CW: What are your thoughts about using AI as a cycling training tool?

ChatGPT: As an AI language model, I don’t have personal thoughts or opinions, but I can provide you with an analysis of the potential for artificial intelligence in cycling, training for both amateur and professional riders.

The potential for artificial intelligence in cycling training is truly remarkable, and it holds great promise for both amateur and professional riders. For amateur riders, AI can offer affordable coaching solutions. Many cycling enthusiasts may not have access to professional coaches or extensive training resources.

AI-based systems can bridge this gap by providing training plans based on individual capabilities, goals, and even constraints such as time availability and fitness levels.

Professional riders can also benefit immensely from AI in training. AI algorithms can analyse vast amounts of data from various sources. By processing this data, AI can identify patterns, uncover hidden insights, and deliver real-time feedback to athletes and their coaches.

CW: What is the biggest risk of using AI for cycling training?

ChatGPT: The biggest risk of using AI training for both amateurs and professional cycling lies in the potential for overreliance on technology and the disregard for individualised training needs and human expertise. While AI can offer valuable insights and data analysis, it should be seen as a tool to complement traditional coaching methods rather than a replacement for them.

CW: Overall, is using AI in cycling training a good thing?

ChatGPT: It’s important to note that while AI can provide valuable guidance, human expertise and judgement remain invaluable. AI should be seen as a tool to augment human capabilities rather than replace them entirely. Ethical considerations regarding data privacy, algorithm transparency, and fair competition also need to be addressed as AI becomes more integrated into the cycling world.

AI nutrition 

While AI coaching may be in its early stages, AI nutrition has already entered the world of pro cycling. As reported in the New Scientist in April, Jumbo-Visma used AI technology at last year’s Tour de France – where, let’s not forget, they won the yellow, polka dot and green jerseys as well as six individual stages. By inputting a rider’s height, weight, age and role, along with external factors such as the weather and stage type, an estimated daily calorie amount is given.

This AI model was pitted against the Jumbo coaches’ previous model, and once post-stage data was analysed, scored on a 0-100% scale. The machine-learning model scored 82%, while the previous model scored 52%.

Remember when Tadej Pogačar was by his own admission “a bit short on fuel” on Stage 11 of last year’s Tour de France and Jonas Vingegaard crossed the line 2.51 ahead, effectively winning the yellow jersey? You have to wonder whether AI nutrition-planning made the difference. 

Contacted for more details about their use of AI in nutrition strategies, Jumbo-Visma did not reply.

AI strength work: ‘Our algorithm selects from 1,000 exercises’

Artificial intelligence isn’t only being used for on-bike training. Programme.app uses AI to design resistance training sessions that can be completed at home with limited equipment. We asked the app’s co-founder Sean Klein how it works:

“Programme works by taking the client’s goal, training experience and available equipment and time to generate a resistance training programme specifically for their situation. The goal of Programme is to provide training sessions that correspond precisely to the individual, with progressive overload over time, keeping the client moving towards their goals.

“The algorithm draws on a movement library of over 1,000 exercises labelled with difficulty, movement category, weight multiplier, repetition multipliers and session position. Once the user has entered their equipment, ability and time available, the algorithm selects the session structure.

“Our biggest limitation is that we cannot teach technique. We can give cues and tips and provide videos, but we are not there in person, working on the minutiae of the exercises. We cannot adapt sessions based on life events either.

"When you coach someone in person, it can be clear if they are having a bad day, unfortunately an algorithm will struggle to take this sort of information into account.”

Pros and cons

So, is AI more intelligent than humans when it comes to cycling training? As yet, the jury is out. AI is excellent at sifting through data and identifying patterns, while cycling’s obsession with tracking data via apps and wearables makes it a prime candidate for feeling the full force of the AI revolution. It is conceivable that AI could be used to analyse multiple data streams at once: sleep, heart rate (including heart rate variability) and body temperature, alongside workload and intensities. It may then be able to extrapolate the physiological effects of specific sessions, guide recovery and subsequent training, while growing more accurate with the more data it is fed.

As a full-time cyclist, the rise of AI interests me but doesn’t necessarily excite me. In theory, training sessions could be precisely dictated down to the last watt of effort, the last second of interval duration.

Right now, we exist in a state of fine balance. There is enough data to allow targeted training, but not so much that human decision-making is redundant. I don’t want the next generation coming through to be even more obsessed with data; I don’t want them to be forbidden by machines from having a good, old-fashioned, unstructured smash-up at the end of group rides.

Our growing reliance on data is not going to be held back, nor is the desire for athletic improvement. Where AI takes us next on that journey is difficult to predict.

If we time-travelled back to the 1970s and tried to explain the effect power meters would have, contemporaries of the Merckx-Hinault era would laugh at us. It remains to be seen whether AI is like the power meter or the flying car – revolutionary or overhyped?