Ninety-four percent. That’s the rate of agreement between neural networks and international experts when evaluating an athlete’s technique. This figure, drawn from a systematic review of 73 studies published in 2025 in Bioengineering (MDPI), doesn’t suggest that AI replaces the eye of coaches. It says something more troubling: that athletic performance is now also the product of a pipeline of data, algorithms, and infrastructure investments that the richest federations have been quietly accumulating over the past decade.
Before Los Angeles 2028, the question is no longer whether technology is changing elite sport. It already has. The question is who benefits from it, how it is governed, and what it does to the value of records.
The Essentials
- A systematic review of 73 studies published in 2025 shows that neural networks achieve 94% agreement with international experts in evaluating sports technique.
- Predictive AI models reduce repeat injuries by 23% and reach 85% accuracy in anticipating hamstring injuries.
- Major national federations in football, cycling, and athletics have integrated these tools into their elite programs, widening the gap with federations unable to keep up.
- The absence of common standards makes it difficult to compare records between generations and raises a governance question that the IOC has yet to settle.
94% Agreement: What That Means in Practice
To understand this figure, you first need to know what is being measured. Technical evaluation in elite sport—whether the takeoff phase of a high jump, the angle of attack of a rower, or the posture of a javelin thrower—has historically rested on the experienced eye of coaches trained over years. These judgments are not arbitrary. They are the result of tacit expertise that is difficult to formalize.
What the systematic review published on NCBI reveals is that neural networks trained on annotated video sequences manage to replicate this judgment in nine cases out of ten. Not by guessing, but by capturing biomechanical micro-patterns that the human eye does not always account for simultaneously: the position of the center of gravity at time T, the synchronization of muscle chains, left-right asymmetry over multiple movement cycles. These algorithms do not replace the coach. They provide him with a layer of information that the coach alone cannot produce in real time.
What changed between 2020 and today is the maturity of the technical pipeline. Wearable inertial sensors, which measure acceleration and rotation at several hundred hertz, now cost a fraction of what they were worth ten years ago. Computer vision embedded in standard cameras enables three-dimensional analysis without marker suits. And deep learning models, trained on increasingly large databases, gain in generalization. The result is a biomechanical analysis system that twenty years ago was reserved for university laboratories and now integrates into the training routines of several dozen national federations.
Predicting Injuries Before They Occur
The most advanced application of these tools is not performance optimization. It’s injury prevention. And that’s where the data is strongest.
The same systematic review reports that machine learning models achieve 85% accuracy in predicting hamstring injuries, the muscle group most affected in sprinting and change-of-direction sports. Even more significant: protocols incorporating these predictions reduce repeat injuries by 23%. This last figure deserves attention. A repeat injury is not simply poor performance. It is often a shortened career, chronic pain, a premature exit from competition. Reducing repeat injuries by a quarter concretely changes the life trajectory of athletes.
How do these models work? They aggregate heterogeneous signals: training load data over several weeks, biomechanical asymmetries captured by inertial sensors, history of previous injuries, neuromuscular fatigue indicators from jump tests. None of these signals, taken in isolation, predicts much. It is their combination, processed by ensemble algorithms, that reveals precursor patterns.
Several European football clubs have used these systems since 2022. FC Barcelona, through its sports science department, has published internal results since 2023 showing a reduction in days absent for muscle injury. A growing number of national athletics federations have deployed similar protocols, notably in partnership with their national sports institutes. These examples are not isolated cases: they signal a structural trend in federations with the means to invest.
Major Federations Have Gained a Head Start
Football, cycling, tennis, and athletics were the first to industrialize these tools, for straightforward reasons: high television revenues, sponsor interest in measurable performance, and an already-embedded culture of advanced statistics. In these disciplines, the technical staff of national teams and professional clubs now integrate data science specialists alongside strength and conditioning coaches and sports medicine physicians.
Cycling may be the most developed example. The INEOS Grenadiers team, which dominated the Tour de France for several seasons, is known for its intensive use of aerodynamic modeling and power optimization algorithms. Its sports director publicly describes their approach as “real-time management of individual physiological constraints.” In practice, this means that each stage is planned from models that account for personalized fatigue curves for each rider, updated nightly.
Athletics follows the same logic, but with an additional constraint: record comparability. When Mondo Duplantis improves his own pole vault world record, is the performance strictly comparable to that of Sergei Bubka thirty years ago, who competed without personalized biomechanical modeling? The question is not rhetorical. It is at the heart of current discussions at World Athletics, even if the organization has not yet produced a formal framework to address it.
AI in Sport Follows the Same Logic of Exclusion as the Economy
Productivity gains linked to AI are not distributed uniformly. This is true in industry and services, as data on the transformation of work by AI agents shows. It is also true in sport.
Federations from low- and middle-income countries cannot absorb the cost of a complete biomechanical pipeline. A professional-quality markerless motion capture system costs between €15,000 and €50,000 to install. A data scientist specialized in sports science costs between €60,000 and €120,000 per year depending on markets. A national injury prevention program integrating AI requires medical and analytical infrastructure that most Olympic committees in Africa, Central Asia, or Latin America cannot finance.
The result is a widening bifurcation. On one side, athletes trained in technologically dense environments, with near-continuous biomechanical monitoring and personalized predictive models. On the other, athletes equally talented who train with methods from twenty years ago, exposed to the same unanticipated injury risks, without access to technical optimization tools.
This gap is not new in elite sport. Unequal funding of Olympic delegations is as old as the modern Games. But the technological dimension gives it new depth. Previously, the comparative advantage of a well-funded federation lay mainly in physical facilities, doctors, qualified coaches. Today, it also lies in a structural informational advantage that an individual athlete cannot make up for through training alone.
The IOC has so far approached this question from the angle of material equipment, with strict rules on running shoes or swimming suits. It has not yet equipped itself with an equivalent framework for performance analysis tools. This is a governance gap, not an accidental omission: lobbies of wealthy federations have no interest in seeing their technological advantages regulated.
What the Pioneers Are Actually Doing
Describing the gap without naming those attempting to close it would be incomplete. Several initiatives show that wider diffusion of these tools is possible.
The World Anti-Doping Agency (WADA) has financed since 2023 a program for sharing biomechanical data between national federations, with the stated objective of creating common reference databases accessible to less well-funded Olympic committees. The program is still modest: about twenty participating federations, three pilot sports (athletics, weightlifting, swimming). But the architecture is in place.
On the academic side, the biomechanics research group at the University of Queensland has developed analysis protocols based on consumer smartphones, capable of producing kinematic estimates with error less than 5% compared to laboratory systems. These methods, published in open access, are already being used by federations in Kenya and Ethiopia for technical analysis of their distance runners.
In the United States, the TeamUSA Data Initiative, launched by the U.S. Olympic Committee in 2024, aggregates data from 28 sports and provides national coaches with a common biomechanical analysis platform. The centralized approach enables significant economies of scale: instead of each federation building its own pipeline, they share common infrastructure. This model could inspire similar initiatives in other Olympic committees, provided political will and funding follow.
The problem remains entire for federations lacking even the framework conditions allowing these less expensive alternatives. Training a sports data analyst assumes a university capable of hosting them, a national sports program capable of employing them, and an institutional ecosystem capable of valuing this work. These conditions are missing in a majority of countries represented at the Games.
Los Angeles 2028: A Deadline Without Governance
In the two years separating today from Los Angeles, technological gaps between federations will not close spontaneously. Ongoing investments by the best-resourced actors continue to accelerate, driven by competitive pressure and sponsor appetite for narratives of measurable performance.
The real question that the 2028 Games will pose is not technical. It is political. Will the IOC and international federations acknowledge that technological infrastructure is now part of sports preparation as much as physical facilities, and that this justifies an equalization mechanism? Will they define common standards for biomechanical analysis, so as to render athletic performances comparable from one generation of athletes to another?
These questions do not yet have clearly identified institutional advocates. World Athletics launched in 2025 a working group on technology and performance, but its mandate is limited to material equipment. The IOC has constituted a commission on AI in sport, with first recommendations expected for 2027, that is, after the start of the Los Angeles qualification cycle.
The problem with the absence of governance in a rapidly changing domain is that implicit rules form anyway. They form to the benefit of those with the means to shape them. If the IOC and international federations do not act before 2028, the conversation about technological equity in sport will probably be held after the Games, with already deeper inequalities to correct.
The medal has long depended on financing. It now also depends on the laboratory. The question is whether this dependence will be organized collectively, or left to market forces.
Sources
- Scoping review on AI and sports biomechanics – Souaifi et al. 2025 (PMC/NCBI)
- World Athletics, working group on technology and performance, 2025 (worldathletics.org)
- INSEP, injury prevention program through predictive modeling, 2024 (insep.fr)
- University of Queensland, smartphone-based kinematic analysis protocols, open access (uq.edu.au)
- U.S. Olympic Committee, TeamUSA Data Initiative, 2024 (teamusa.org)
- ANR Project FULFILL – INSEP / FFA – sprint biomechanics
- INEOS Grenadiers – performance restructuring and aerodynamics (2024)
- Google / LA28 – AI partnership for 2028 Olympics
- PMC – Hamstrings: muscle group most affected in sprinting sports
- Bioengineering (MDPI) – Original publication of the scoping review