Artificial intelligence (AI) in sports nutrition

Over the past decade, sports nutrition has quietly become one of the most technologically driven areas of performance support. Dietitians and athletes now interact with artificial intelligence (AI) every day, often without realising it: readiness scores pushed to their phones upon waking, automated messages interpreting training data after a ride or run, and wearable-generated summaries telling athletes whether they recovered “well” or “poorly.” This steady infiltration of AI into daily practice has been so seamless that many now take it for granted. At the same time, rapid advances in generative AI have led to bold claims that nutrition planning, intake analysis, and even behaviour modification can be automated entirely.

The reality is more nuanced. AI has already transformed several aspects of sports nutrition, but it remains vulnerable in others. The dietitians and performance professionals who will gain the most value are not those who resist technology, nor those who embrace it uncritically, but those who understand when it is reliable and when it is not. This article examines where AI currently excels, where it falls short, and how it can be integrated responsibly into an evidence-based nutrition workflow.

AI in the everyday life of an athlete

The most influential form of AI in sport is not futuristic or experimental. It is the AI embedded in routine digital tools. A typical morning might involve a wearable watch summarising sleep architecture and readiness; GPS and route prediction software managing traffic and commute; and, after a training session, automated notifications from Strava, TrainingPeaks or Garmin describing power, heart rate and recovery trends. Nutrition software and consumer wellness apps build upon these same data streams.

These examples illustrate a key point: AI is already shaping athlete behaviour. The question for practitioners is whether they can ensure that nutritional decisions informed by AI are grounded in physiology rather than marketing.

AI is already shaping athlete behaviour. The question for practitioners is whether they can ensure nutritional decisions informed by AI are grounded in physiology rather than marketing.

Where AI works really well

AI performs best when a problem has a clearly defined structure, when outcomes can be assessed objectively, and when the input data is accurate. In these circumstances, AI is not a threat to professional expertise but a powerful amplifier of it. One of the most useful applications is in nutrition planning for predictable endurance environments. In modern professional cycling, for example, power-meter files, course metadata and environmental conditions can be combined to estimate stage demands and energy requirements with impressive precision. Ten years ago such calculations required a spreadsheets, constant manual adjustment and hours of labour; today, AI-powered software can automatically convert predictions into individualised meal plans for an entire squad with a single click.

A similar logic applies to continuous glucose monitoring (CGM). The sheer volume of CGM data makes manual interpretation challenging and interpretation in realtime near impossible. Machine-learning models excel at detecting trends, repeated responses to certain foods, and patterns that elude human observers. Here, AI does not replace practitioner interpretation; rather, it relieves practitioners of the burden of extracting basic patterns from vast datasets so that they can focus on decision-making and athlete guidance. However, if we leave it completely up to AI to interpret the data this can be problematic (more about this later).

AI performs best when a problem has a clearly defined structure, when outcomes can be assessed objectively, and when the input data is accurate.

AI is also becoming increasingly relevant in knowledge management. The speed of scientific output in sports nutrition means that no practitioner can realistically read every new paper. Generative AI makes it possible to summarise research rapidly, identify recurring themes and extract mechanistic foundations or key findings. When a practitioner then evaluates the quality of those summaries through a critical lens, AI becomes an enabler of evidence-based practice rather than a risk to it.

In summary, AI works best in scenarios where:

1. The problem is well defined.

2. Accurate inputs are defined.

3. The answer is relatively black-and-white.

Where AI does not work so well

The limitations of AI become clear the moment a question requires context, interpretation or tolerance for uncertainty. Some of the most prominent examples come from studies evaluating responses of large language models to questions in domains known to contain misinformation. While AI performed exceptionally well on questions with clear, uncontroversial answers, performance deteriorated sharply when the prompts were complex, ambiguous or scientifically contested. In these contexts, responses were often superficially persuasive but physiologically unsound, and citations (if included) were frequently fabricated.

Nutrition, of course, contains many such grey areas. Questions around supplements, relative energy deficiency, gastrointestinal resilience during endurance events, or the nutritional management of extreme environmental stress are rarely reducible to a binary “right” or “wrong.” These are the situations where generative AI appears most confident and most unreliable at the same time.

While AI performed exceptionally well on questions with clear, uncontroversial answers, performance deteriorated sharply when the prompts were complex, ambiguous or scientifically contested.

Food-image recognition (perhaps the most heavily advertised AI application in nutrition) faces similar challenges. The promise is appealing: athletes photograph their meals and technology calculates energy and nutrient composition automatically. Yet in practice, the difficulty of distinguishing visually identical foods, detecting oil or sauces added after cooking, separating layered items on a plate, and estimating portion size means that accuracy is often insufficient for high-performance sport. Even in tightly controlled research environments, performance varies widely. These systems will undoubtedly improve, but the claim that they can fully replace human oversight in elite settings remains premature.

A third limitation lies in the misconception that AI can compensate for poor input quality. If athletes under-report snacks, omit supplements or record meals inconsistently, AI does not correct the problem, it amplifies it. Rather than producing a cautious estimate, it generates a confident prediction based on flawed data. In sports nutrition, where health and performance are at stake, that confidence can be dangerous.

The future of AI in sports nutrition

There is growing interest in the development of digital twins, refering to computational models that represent the metabolic identity of an individual athlete. In theory, digital twins could predict glycaemic or gastrointestinal responses to specific meals, forecast fuelling requirements under different physiological conditions, or simulate adaptations to dietary strategies without exposing the athlete to unnecessary risk.

The concept is scientifically fascinating and has strong clinical parallels in personalised medicine. However, the barrier is data quality. Creating a digital twin of an elite athlete would require exceptionally detailed and standardised datasets across nutrition, energy expenditure, metabolic biomarkers, training load, sleep and environmental factors. Most performance environments simply do not generate such comprehensive datasets consistently enough to build robust models. The idea has tremendous potential, but the infrastructure required to deploy it remains aspirational rather than operational in high-performance sport.

AI 'hallucinations'

Occasionally, AI does not merely provide incomplete or imprecise information, it invents it. These so called 'hallucinations' are the product of models that are designed to generate plausible responses rather than to verify accuracy. The consequence is that AI will sometimes give an answer even when an answer does not exist. The same mechanism that enables AI to generate elegant, fluent text also enables it to generate mistakes persuasively.

The consequence [of AI hallucinations] is that AI will sometimes give an answer even when an answer does not exist.

This failure mode exposes the most important lesson for practitioners: AI does not understand uncertainty, ethics or responsibility. It cannot say “this depends,” “we do not know,” or “this is unsafe.” High-performance nutrition is full of situations where these statements are necessary. AI can support expertise, but it cannot replace wisdom.

Towards an AI-augmented model of sports nutrition practise

Across the entire field, a consistent theme emerges. AI reshapes sports nutrition not by replacing professionals, but by redirecting their attention. Dietitians who once spent hours calculating carbohydrate needs can now focus on behaviour change and fuelling strategy adherence. Instead of manually scanning the literature, they can evaluate new evidence critically. Rather than performing administrative tasks, they can devote time to communication, culture and athlete engagement. AI eliminates bottlenecks; humans provide judgement.

It is unlikely that AI will replace skilled practitioners in the foreseeable future. However, it may replace practitioners who resist the tools that others use effectively.

Summary

AI is not inherently beneficial or harmful for sports nutrition. Its value depends entirely on how it is used. When the problem is well defined, the data are structured and the consequences of error are low, AI can be extraordinarily efficient. When the situation requires nuance, contextualisation or ethical reasoning, human oversight is indispensable. The future of sports nutrition will not be defined by a competition between humans and technology, but by collaboration, with AI handling speed and scale, and practitioners guiding interpretation, personalisation and decision-making.