Where AI Actually Works in Sports: Ball-Tracking Cameras, Injury-Prediction Wearables, and the Line-Calling Controversy

In July 2025, Wimbledon retired its human line judges after 147 years and handed the calls entirely to Hawk-Eye's Electronic Line Calling (ELC) system. Within days, the tournament had to publicly apologize: during a match on Centre Court, the system failed to call a ball out because an operator had accidentally switched it off mid-point, and nobody caught it in time for the point to be replayed correctly. The underlying ball-tracking technology — the same core system that's been calling lines in cricket and tennis for two decades — wasn't the problem. The problem was that a fully automated pipeline still has a human-operated kill switch, and removing 300 years of judgment from a stadium doesn't remove every point of failure, it just relocates them.

That's a useful frame for AI in sports generally. The computer-vision and sensor-fusion problems — where is the ball, where is the player, did the foot cross the line — are largely solved and have been in production for years. The harder, less-solved problems are downstream: predicting things that don't have a clean ground truth, like injury risk, or replacing judgment calls that fans and leagues still want a human to own.

Optical tracking and automated officiating

Hawk-Eye (owned by Sony since 2011) is the incumbent here, and its reach goes well beyond tennis. The same core technology — six to ten synchronized high-speed cameras triangulating a ball's position to within a few millimeters — powers goal-line technology in the Premier League, DRS ball-tracking in cricket, and the semi-automated offside system FIFA deployed at the 2022 and 2026 World Cups, which adds a chip-embedded ball and 29 limb-tracking data points per player, sampled at 50 times per second, to flag offside positions in seconds rather than the minutes a VAR review used to take.

The reason this application area is mature isn't just camera quality — it's that the problem has a verifiable ground truth. A ball either crossed a line or it didn't; a physical position in 3D space is measurable and auditable, and the system's calls can be checked frame-by-frame against physics. That's a fundamentally easier problem than "was that a foul" or "will this player get injured," which is why officiating AI has stayed narrowly scoped to spatial fact-finding (ball position, offside lines, foot faults) rather than judgment calls (fouls, intent, dangerous play). The NBA has experimented with computer-vision foul detection internally but has not deployed it as an in-game officiating tool — fouls involve intent and contact severity that don't reduce cleanly to coordinates.

The Wimbledon incident is also a preview of a recurring pattern: when a vendor's tracking system is the sole source of truth and there's no fallback, an operational failure (a misconfigured camera, a dropped frame, a disabled feed) becomes a competitive-integrity incident instead of a minor bug. Leagues that have kept a human review layer on top of automated tracking — the NFL's use of Sony's Hawk-Eye-powered system for measuring first downs, introduced in 2025 to replace the chain gang, still has referees confirm the automated spot — have avoided this failure mode so far.

Player tracking and real-time analytics

This is the highest-volume commercial application of AI in sports, and it's arguably where the vertical is most mature. The NBA has used Second Spectrum's optical tracking system (six cameras per arena, tracking all ten players plus the ball at 25 frames per second) since 2017, feeding models that generate real-time shot-quality metrics, defensive matchup data, and the "quantified" stats (like expected field goal percentage) now baked into broadcasts. Stats Perform, which acquired much of Opta's tracking business, provides similar infrastructure across soccer leagues including the Premier League and MLS.

The NFL's Next Gen Stats, run jointly with AWS, takes a different sensor approach: RFID chips embedded in shoulder pads (not cameras) transmit player location ten times per second, which AWS SageMaker models turn into win-probability curves, expected rush yards, and pass-rush pressure metrics that show up on broadcasts within seconds of a play ending. The technical bottleneck here isn't the tracking — RFID and optical tracking are both well-understood problems now — it's inference latency under a hard real-time constraint. A win-probability model that updates 30 seconds after a play is graphics-department trivia; one that updates before the broadcast cuts back from replay is a product.

What's genuinely new since 2023 is teams using these tracking feeds for in-game and film-study decisions, not just broadcast graphics — Second Spectrum's data now feeds directly into coaching software used for opponent scouting, and several NFL teams have built internal models on top of Next Gen Stats' raw player-location data for defensive personnel decisions. The maturity gap between "broadcast enhancement" and "decision support" has mostly closed for tracking data specifically.

Injury prediction and load-management wearables

This is where the field gets genuinely uncertain, and where the gap between vendor marketing and peer-reviewed evidence is widest. Companies like Catapult Sports and STATSports sell GPS-and-inertial-measurement wearables — now standard equipment in the NFL, most European soccer academies, and Olympic training programs — that log distance covered, acceleration/deceleration counts, and "player load," a proprietary composite metric meant to proxy cumulative physical stress. The pitch is straightforward: feed weeks of load data into a model, flag players whose workload pattern resembles those who got hurt before, and either rest them or adjust training.

The catch is that soft-tissue injury (hamstring strains, ACL tears — the injuries teams most want to predict) is multi-causal in ways that make it a poor fit for supervised learning on load data alone. A 2020 systematic review in the British Journal of Sports Medicine found that most published injury-prediction models had modest-at-best external validity — they performed well on the team/dataset they were trained on and degraded sharply when applied to a different roster, because individual injury history, genetics, sleep, and off-field stress aren't captured by a GPS vest. Teams that have publicly discussed their internal models (several NBA teams have described "load management" decisions, like resting healthy stars on back-to-backs, as informed by wearable data) still treat the output as one input to a medical staff's judgment, not a standalone prediction to act on.

The adoption pattern reflects that uncertainty: wearables for load monitoring are near-universal across major pro leagues, but wearables for load prediction driving autonomous rest/train decisions are not — the output almost always routes through a human athletic trainer, which is the same pattern seen in AI-assisted medical diagnosis more broadly: the model flags risk, a licensed professional decides.

Fan-facing and broadcast AI

The lowest-stakes, fastest-moving application area is broadcast enhancement — AI-generated highlight reels (IBM's Watson-powered highlight selection at Wimbledon and the US Open, which scores points on crowd noise, player reaction, and shot difficulty to auto-generate reel candidates), automated camera direction that follows the ball without a human operator (used in lower-tier and youth sports broadcasts where a full camera crew isn't economical), and AI-generated play-by-play for markets that wouldn't otherwise get commentary. None of this has the integrity stakes of officiating or the safety stakes of injury prediction, so adoption has moved faster and with less scrutiny — but it also hasn't fully replaced premium broadcast talent for marquee events, where audiences still expect a named human voice.

Sports-betting odds models are the other fast-growing corner here, and the one with the least public technical detail — sportsbooks like DraftKings and FanDuel run proprietary real-time pricing models that adjust in-game odds based on live win-probability shifts, but the models themselves aren't published or peer-reviewed, for obvious competitive reasons, which makes this the hardest application area to evaluate from the outside.

Maturity by application area

The takeaway

The pattern across every application area is the same one that shows up in AI-in-healthcare and AI-in-finance coverage: AI works best in sports where the target variable has a verifiable ground truth (ball position, player coordinates) and works worst where the target is a noisy, multi-causal outcome with no clean label (will this hamstring tear, was that worth a foul). Officiating and tracking systems have quietly become infrastructure — nobody's debating whether Hawk-Eye's ball physics are accurate anymore, the debate is about operational reliability and who's accountable when the pipeline fails mid-match. Injury prediction is still in the phase where the marketing outruns the peer-reviewed evidence, and the leagues spending the most on wearables are, correctly, still routing every prediction through a human trainer before it changes a player's workload.

If you're evaluating a vendor's AI claim in this space, the fastest diagnostic is the same one that works in any domain: ask what the ground truth label was during training, and whether the model's performance has been validated on data it didn't see during development. Spatial tracking problems tend to pass that test. Injury and outcome prediction, for now, mostly don't.