AI in sport is strongest where a machine has to see and measure. That covers optical tracking of players and the ball, statistics built on that tracking, officiating aids, and workload models that flag injury risk. AI is weakest when it is asked to forecast what has not happened yet. A peer-reviewed review finds that result-prediction accuracy varies by sport and levelled off in a large soccer dataset. In one club study, an injury forecaster beat older methods, yet about half its alerts were false. Officiating systems also fail, as when Wimbledon’s was turned off in error.
The rest of this article covers what the major leagues actually run, what the research says about reliability, and where the limits sit.
How is AI used in sports today, from performance to fan experience?
This article groups AI in sport into four layers: perception, derived statistics, decision support and fan-facing features. Each layer depends on the one beneath it, so the quality of the tracking data limits everything built on top.
Perception here is camera-based tracking. In the NBA, Sony Hawk-Eye uses 3D pose detection to track 29 points on every player’s body, plus the ball, 60 times per second. FIFA’s semi-automated offside system uses 12 tracking cameras under the stadium roof to follow the ball and up to 29 data points per player, 50 times per second. Computer vision also works on game video. The NFL’s Digital Athlete was trained to recognise helmets from every angle, then to spot helmet impacts and match them against tracking data to identify the players involved.
Derived statistics turn tracking data into readable numbers such as expected field goal percentage, defensive attention and completion probability. Decision support turns data into advice on tactics, training load or a refereeing call. Fan features sit at the top. Wimbledon’s AI Commentary, built by IBM on watsonx foundation models, narrates highlight clips with toggleable captions. MLB plans to use Google’s Gemini platform to personalise content and recommendations for fans. Getting those features onto a live broadcast reliably is a production problem of its own, covered in governing real-time AI in live sports production.
How is the NBA using AI?
The NBA uses AI to turn optical tracking into new statistics. Its “Inside the Game” platform with AWS processes tracking data on 29 points per player and uses machine learning to generate real-time insights for the NBA App, NBA.com and live broadcasts. Sportico reported that the platform combines Sony Hawk-Eye’s body tracking with AWS’s AI services and rolls out during the 2025-26 season.
The individual metrics show what the models do:
- Defensive Box Score: AI works out in real time which defender is responsible for each offensive player, so every box-score stat can be credited against a specific defender.
- Shot Difficulty: Expected Field Goal % takes into account the shooter’s orientation and setup, defensive pressure and the positions of the players around them.
- Gravity: custom neural networks process optical tracking 60 times per second to measure how much defensive attention a player draws.
- Play Finder: built on Amazon Bedrock and Amazon SageMaker, it searches thousands of games for plays similar to a given one.
- Handle Score: rates ball-handling on a single scale.
Instead of writing rules for what a good handle looks like, the team borrowed human preference alignment from large language models. Experts compared clips in pairs and picked the better one, producing over 10,000 A/B comparisons across more than 1,000 plays. That is a sensible way to model something experts can recognise but cannot define: the label becomes a judgement between two clips rather than a rule. Sportico gave a measured verdict on the stats overall: at their best, they “add numbers to debates already being had”.
NFL and MLB: tracking at league scale
The NFL and MLB also build statistics on tracking data at league scale. Since 2017, AWS has stored the NFL’s tracking data, nearly 300 million data points per season, and NFL engineers build and deploy their Next Gen Stats models in Amazon SageMaker. The completion probability model takes 10 on-field measurements, such as pass distance and the distance between the receiver and the nearest defenders.
Google reports a 300 ms improvement in stats delivery, which enables real-time analysis of ball and strike calls. Bat tracking adds five high frame-rate Hawk-Eye cameras that measure swing speed and the path of the bat relative to the incoming ball. What those swing metrics mean analytically is covered in our guide to AI in sports analytics.
How is machine learning applied to athlete performance and training?
Here machine learning reaches athlete performance in two ways: tactical advice from tracking data and training-load management from workload and sensor data.
TacticAI is an AI football tactics assistant. Developed with Liverpool FC, it predicts who will receive a corner kick and whether a shot follows, and it generates alternative player setups. It was trained on 7,176 valid corners from Premier League tracking data captured at 25 frames per second. Expert raters favoured its suggestions over the existing tactics 90% of the time, and they could not reliably tell its generated setups from real ones (average F1 of 0.60 on that task). The authors chose corners on purpose: they happen about ten times a match and start from a fixed position, so the problem is unusually well structured.
On training load, the NFL’s Digital Athlete combines Next Gen Stats with video, equipment sensors, weather and play type. Its Risk Mitigation Modeling analyses training data to estimate each player’s ideal training volume while keeping injury risk down. Each game week it processes about 6.8 million video frames and logs around 100 million player positions.
A 2026 Frontiers in Digital Health review describes systems that use phone cameras and wearables to assess joint angles and movement quality in real time, then adjust sets, intensity or rest. The same review argues that these systems are a supplementary tool and not a replacement, because empathy and leadership remain central to coaching.
Can AI really predict sports injuries?
AI has flagged injury risk better than older single-dimension methods, but without certainty. The measured evidence here comes from a study of an Italian professional football club. A multi-dimensional forecaster built on GPS workload data reached about 50% precision, while existing single-dimension approaches managed under 5%. In the study’s main test, the model caught around 80% of injuries.
The study’s limits matter as much as its result. It followed 26 players for 23 weeks and recorded 23 non-contact injuries. The model performed poorly in the first weeks and only stabilised after about 14 weeks of data collection. In the end, only 3 of 55 candidate features carried the prediction. A precision of 50% also means that half of the alerts are false alarms, and the authors note that unnecessary rest is costly when a squad is short of players.
The NFL describes injury prediction as a future goal for its own system. The league says the Digital Athlete “may be able to help predict and prevent injuries” in time, not that it already does. Its models for measuring the forces behind concussions are still in development.
Is there an AI that can predict sports results, and how reliable is it?
Many models predict match results, with accuracy that varies widely by sport. A review by Bunker and Susnjak covered 31 papers across nine sports. Several of its results show the limits:
- In NBA forecasting, the best model reached 67.8%, below the best expert predictions on teamrankings.com at 69.6%.
- In baseball, an SVM’s predictions were not significantly more accurate than those derived from betting odds.
- In football, draws were especially hard to predict, with accuracy of around 46% on three-way prediction against roughly 70% when draws were excluded.
- In the 2017 Open International Soccer Database Competition, model accuracy stayed modest even with more than 216,000 matches available.
Three further findings explain the ceiling. Feature engineering mattered more to accuracy than the number of matches. Some studies used standard cross-validation, which shuffles matches so that future matches predict past ones, a pitfall the review identified in a number of studies. And a model built for one sport generally cannot be applied to another without a new experimental process. Comparisons between studies are difficult in the first place, because they differ in sport, dataset, features and seasons.
Betting markets are a hard benchmark to beat. One betting study reports that strategies to beat bookmakers have produced inconsistent returns, and its own profitable strategy came from exploiting mispriced odds rather than from out-forecasting the market.
Result prediction asks a model to beat experts and betting markets. Tracking a ball at 50 Hz is a repeatable task where machines help officials decide faster and more accurately. TechnoLynx scopes AI work around outcomes that are plausible at current capability, and treats “smarter than the experts” as a far riskier proposition than “do this repeatable task for us”.
Where AI in sport is dependable and where it is fragile
| Use | What the evidence shows | Reliability today |
|---|---|---|
| Player and ball tracking | 29 body points at 60 Hz in the NBA; 29 points at 50 Hz for FIFA offside | In live league and tournament use |
| Derived stats (NBA, NFL, MLB) | Real-time stats from tracking; NFL completion model uses 10 inputs | Live; adds numbers to existing debates |
| Set-piece tactics | TacticAI suggestions favoured 90% of the time by experts | Focused on corner kicks |
| Injury risk | 50% precision vs under 5% for older approaches, in one club study | Useful as a flag; half of alerts false; needed 14 weeks of data |
| Match results | NBA models below expert picks; baseball no better than betting odds | Modest; some studies used flawed validation |
| Officiating | ABS challenges near 50% success in 2025 Spring Training; Wimbledon system turned off in error | In live use; Wimbledon blamed human error |
What are the disadvantages or limitations of using AI in sports?
The main limitations are data scarcity, poor transfer between contexts, and operational failure in live systems.
Data is scarce in sport. TacticAI’s authors note that gold-standard football data is scarce, because the Premier League plays only a few hundred games a season. The JAIR review found that sports datasets tend to be small, and that the push for ever more features raises the risk of over-fitting.
In officiating, the recorded failure was operational. At Wimbledon in 2025, the first year electronic line calling replaced line judges, the system was turned off in error on part of Centre Court. Three calls went unpicked by the system, and the chair umpire, not told it was off, called two of them. The club’s fix was procedural: it removed the operators’ ability to deactivate ball tracking by hand. Wimbledon put the source down to human error.
FIFA’s offside workflow keeps officials in the loop. Its offside system raises an alert, but video officials check the kick point and offside line by hand before the on-field referee is told. Complex incidents can still need a full review, and judgements such as whether a player was interfering stay with the referee. In MLB’s 2025 Spring Training test of the ABS Challenge System, teams challenged just over four pitches per game with a success rate near 50%, and catchers succeeded most often at 56%.
How far human judgement should stay in the loop is a wider question than officiating, and it is the focus of our piece on what AI in sports analytics actually does and where it stops.
What are the ethical risks of AI in sport?
The 2026 Frontiers in Digital Health review names three critical ethical risks: privacy invasion, data bias and unclear responsibility when something goes wrong.
Privacy risk follows from how much data these systems need: identity details, health metrics, location, sleep and eating patterns. If that data leaks, rivals could study an athlete’s weaknesses, and location data could expose athletes to tracking or harassment. Bias arises when training data over-represents some groups. The review’s example is a coach trained mostly on European and American athletes giving advice that raises injury risk for athletes with different physical profiles. On responsibility, the review argues that a human coach should approve AI training plans and that team doctors should confirm return-to-play decisions, so that accountability is clear when an athlete is hurt. Its recommended safeguards are data minimisation, diverse training data, bias audits and monitoring across the model’s lifecycle.
A bias audit or a sign-off step is a component of the system and has to be designed in from the start.
Who can build the engineering behind a sports AI system?
TechnoLynx builds the engineering layer sports AI depends on, under AI Engineering for Sports Analytics: computer vision models that turn camera feeds into positions and poses, the pipelines that move tracking data, and GPU inference fast enough for live use. TechnoLynx puts feasibility and return first: is the target outcome plausible at current capability, and does it pay off? That keeps a project away from outcomes that would need a model to outperform experts.
Some teams do not need this. A club that only needs the league’s standard tracking feed and existing stats can use what its league and that feed’s vendor already supply. An in-house team makes sense once the club holds proprietary data and wants to control the models built on it. When choosing a partner, ask how they would validate a model on time-ordered data, given the cross-validation pitfall the JAIR review describes.
Where AI in sport goes next
No source here forecasts the future of AI in sport, so this section is our judgement. The near-term direction we see is richer tracking feeding more derived stats, with generative interfaces on top. The NBA describes Play Finder as groundwork for future generative AI built on player tracking data, and the NFL says the Digital Athlete will support real-time injury-risk analysis during games. Large events such as the Games raise the same questions on a bigger stage, as discussed in how artificial intelligence could transform the Olympics.
We expect the gap between sensing and forecasting to persist. We expect better cameras and models to keep improving how precisely sport is measured. Match outcomes stay hard to predict, and the JAIR authors call that unpredictability one of the main reasons people enjoy sport. The open question for any club or league is which decisions it wants a model to make, and which it only wants a model to inform.
