Artificial intelligence (AI) in sport

Artificial intelligence (AI) has rapidly become one of the most frequently referenced concepts in high-performance sport. It is discussed in recruitment, in training planning, in tactical decision-making, and increasingly in sports nutrition. Yet although the term is used widely, the understanding of what AI is, how it works and what the underlying mechanisms are, is often limited.

To practitioners sitting in the applied performance space (coaches, sport scientists, nutritionists, physiologists) AI can appear simultaneously exciting and intimidating. It promises unprecedented insights, but it is also described in language that feels far removed from the day-to-day reality of supporting athletes.

This gap in understanding matters. If AI is to contribute meaningfully to performance, it must be more than a buzzword or a marketing feature. It must be understood as a set of tools built on clear principles, integrated thoughtfully into applied practice, and interpreted by people who understand both sport and context. The purpose of this article is not to oversell AI or reduce it to hype, but to explain how it works, why it is developing so fast, and where it currently has the most value inside elite sport.

AI is not new, but something has changed

Although it is often portrayed as a recent invention, AI has been advancing for decades. Alan Turing proposed the Turing Test in the 1950s to evaluate whether a machine could convincingly imitate human intelligence (1). In 1957, the first Perceptron, a precursor of the neural networks behind modern deep learning, was developed (2,3). Most of the theory behind today’s AI has existed for more than half a century.

What has changed is not the fundamental concept, but the environment around it. Two developments unlocked AI’s current power:

1. The explosion of data collection, especially through wearables, connected devices, power meters and GPS.

2. Cloud computing, which provides the processing capacity to train models on massive datasets.

   
   

Cycling offers a clear example. The shift from inconsistent historical training logs to standardised power-meter files and formats means that riders can now arrive at a team with ten or more years of structured performance data already available. When this volume of data meets the computational power of the cloud, AI becomes operational rather than theoretical.

AI didn’t suddenly become intelligent, it finally received the resources it needed: massive data and the computing power to learn from it. In sport, that combination is starting to turn information into intelligence, and theory into practice.

AI didn't suddenly become intelligent, it finally received the resources it needed: massive data and the computing power to learn from it.

What AI actually is

AI is not a single technology but a family of approaches that has evolved over several decades (see figure above). In the 1950s, AI simply referred to the idea that machines could imitate elements of human intelligence. Over time, new branches emerged as both data and computing power expanded.

Machine Learning (1980s-present)

This is where AI systems began learning directly from data rather than relying on hard-coded rules. In sport, this includes predictive models that learn relationships between workload, recovery, and performance, helping coaches anticipate training responses or race outcomes. Machine Learning is particularly suited to the time-series data common in sport science: power, heart rate, GPS, temperature, and more.

Deep Learning (2010s-present)

Deep learning is a subset of machine learning that uses neural networks to model complex, nonlinear patterns, the kind found in images or continuous signals. It’s used in motion analysis, injury-risk modelling, or automatic detection of events in race footage.

Generative AI (2020s-present)

The newest family builds on deep learning to create new content, text, images, even simulated data. Generative AI models such as large language models (LLMs) can summarise or draft material rapidly, making them useful for knowledge management, literature review, or communication. However, their reliability depends entirely on how well they are grounded in verified information.

Generative AI is the most visible form of AI today, but for applied sport science, Machine Learning and Deep Learning remain the extremely useful but still largely underused. They work quietly in the background, turning years of performance data into practical predictions. In short, AI in sport isn’t one thing. It is a continuum, from early rule-based systems to predictive machine learning and now generative models, each stage adding new capabilities built on the same foundations. More details can be found on history of AI in “The Master Algorithm by Pedro Domingo” (4).

How an AI model is built in practise

From the outside, it can appear that AI simply produces answers. In reality, building a model that practitioners can trust is slow, laborious and highly technical.

The workflow of a typical sports-performance AI project looks like this:

1. Define the problem precisely.

2. Collect large volumes of relevant data.

3. Prepare and clean the data, removing errors and filling gaps.

4. Engineer meaningful features (performance metrics, workload summaries, etc.).

5. Split the dataset into a training set and a test set.

6. Train multiple models and compare performance.

7. Tune the best model.

8. Test the model on data it has never seen.

9. Deploy the model in a format that is accessible to practitioners.

It is during this process that the single most important principle of AI emerges: The model is only as good as the data used to create it.

While the algorithms behind AI are often highly sophisticated, the work of a data scientist rarely involves inventing new ones. In most applied contexts, including sport, the task is to select, adapt, and fine-tune existing models using domain-specific data.

For that reason, data quality becomes the decisive factor. The more complete, consistent, and contextually accurate the data, the more meaningful the model’s predictions will be. Across industries, research consistently shows that 60–70% of the total effort in AI projects is spent on data collection, cleaning, and preparation, not on algorithm development (5).

The model is only as good as the data used to create it.

AI cannot repair fragmented, inconsistent, or low-quality data; it magnifies it. In sport, where performance data are often noisy and heterogeneous, the greatest value comes not from building new algorithms but from building better data foundations to feed the ones we already have.

AI in professional sport

Perhaps the clearest example of AI’s value in elite sport is talent identification in cycling. Historically, scouting was limited by geography, staffing and opportunity: a sports director would attend a few domestic races and recommend promising riders. The number of athletes evaluated was small, and international riders were frequently overlooked.

When athletes voluntarily shared their performance files through platforms like TrainingPeaks, however, entirely new possibilities emerged. A cloud-based system was built to automatically:

• Standardise data from all athletes.

• Compute performance scores relevant to different racing roles.

• Update progression with each new training file.

• Rank hundreds of athletes in real time.

The result was not that an algorithm replaced scouting. Rather, the algorithm surfaced athletes whose performances might otherwise have gone unnoticed. The lesson is not that AI is magical. It is that AI works when it combines high-quality data with human interpretation. Coaches and performance staff still had a critical look at the data and made the final decisions; the model simply widened their field of view.

Why AI in expanding in sport now

Beyond raw computing power, sport has become fertile ground for AI because it involves:

• High-frequency data collection.

• Repetitive decision cycles.

• Clear performance consequences.

• A competitive advantage for early adopters.

Sport rewards small improvements and punishes inefficiency. If one organisation can screen 500 riders instead of 50, generate nutrition plans instantly, rather than manually, or identify “readiness” changes before symptoms appear, the advantage compounds season after season. This explains both the momentum and the anxiety around AI. The question is no longer whether AI can influence performance, but how practitioners can ensure that influence is reliable and ethical.

Interested in learning more? Watch the series of talks on 'AI in sport science and nutrition' with Kevin Yven and Professor Asker Jeukendrup on mysportscience academy.

References

1. Turing, Alan. (October 1950) “ Computing machinery and intelligence” M ind LIX 236 433-460.

2. Rosenblatt, Frank (January 1957). "The Perceptron—a perceiving and recognizing automaton". Cornell Aeronautical Laboratory. doi:10.1093/mind/LIX.236.433.

3. Rosenblatt, F. (1958). "The perceptron: A probabilistic model for information storage and organization in the brain". Psychological Review. 65 (6): 386–408. doi:10.1037/h0042519. ISSN 1939-1471. PMID 13602029.

4. Domingos, Pedro.The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World. (September 2015).

5. D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, Dan Dennison. Hidden Technical Debt in Machine Learning Systems Advances in Neural Information Processing Systems 28 (NIPS 2015).