A defender lunges, a striker runs on, and 60,000 people freeze. Twenty seconds later a graphic appears on the big screen: a 3D reconstruction of a shoulder, a knee, a line. Offside. No flag went up in real time — a computer drew that line. This is AI referee technology, and it now sits inside football, tennis, cricket, baseball and rugby at the highest level.
It is one of the few places where ordinary people watch an AI system make a decision that matters, live, in front of a crowd. That makes it a useful case study in what these systems actually do well — and where they still fall over.
What is AI referee technology?
AI referee technology is a combination of high-speed cameras, sensors and machine learning models that track the ball and players many times per second, then calculate whether a rule was broken. It does not “understand” the sport. It measures positions in three-dimensional space and compares them against a rule written as geometry.
That distinction explains almost everything about how these systems behave. They are excellent at questions with a measurable answer — was the ball in or out, was this shoulder ahead of that one — and useless at questions requiring judgement, like whether a challenge was reckless or a player was feigning injury.
The clearest way to see this is in three sports that have gone furthest: football, tennis and baseball.
How does AI decide an offside in football?
FIFA’s semi-automated offside technology uses 12 dedicated tracking cameras mounted under the stadium roof. They track up to 29 data points on every player — every limb and extremity relevant to an offside call — 50 times per second. A sensor inside the match ball reports its own motion to the video operation room 500 times per second, pinpointing the exact moment of the kick.
Those two data streams answer the two halves of the question. The ball sensor establishes when the pass was played; the camera array establishes where every relevant body part was at that instant. Combining an independent motion source with an independent vision source is what makes the call both fast and hard to dispute.
Crucially, the system does not award the decision. According to FIFA, video match officials must manually verify the proposed kick point and offside line before the on-field referee is told anything. The word “semi-automated” is doing real work: the machine proposes, a human confirms.
Can AI umpires call balls and strikes better than humans?
In baseball, yes — measurably, on the narrow question of whether a pitch crossed the strike zone. Major League Baseball launched its Automated Ball-Strike (ABS) Challenge System for the 2026 season. Pitch tracking is continuous, but the machine only speaks when challenged: each team gets two challenges per game and keeps them if the challenge succeeds.
The design detail worth noticing is who may challenge. Only the pitcher, catcher or batter can call for a review, and they must do it immediately after the pitch — no dugout coaching, no video room. That keeps the game moving and prevents the technology from turning every at-bat into a tribunal.
The spring training numbers show the human umpires were not being humiliated. Across 1,844 challenges, 53% were successful — meaning nearly half the time, players challenged a call the umpire had got right. Batters were notably worse at judging their own strike zone than the defence: 45% success for batter-initiated challenges versus 60% for the pitching side, per ESPN. Teams averaged 4.32 challenges a game.
The takeaway: AI officiating does not reveal that human umpires are bad — it reveals that players’ certainty about close calls is roughly a coin flip.
Tennis and cricket: the millimetre problem
Tennis went furthest fastest. In 2025, Wimbledon replaced human line judges with automated electronic line calling for the first time in 148 years. The underlying Hawk-Eye system triangulates the ball’s position from a ring of about ten high-speed cameras and reconstructs its flight path.
It is precise, but not infinitely precise. Hawk-Eye Innovations puts the system’s average error at a couple of millimetres, and independent testing against a high-speed surface camera has reported a mean error of roughly 3.6 mm. In tennis, that margin is smaller than the fuzz on the ball, so nobody argues.
Cricket has been more honest about the uncertainty, and built it into the rules. In an LBW review, Hawk-Eye measures three things: where the ball pitched, where it struck the batter, and the predicted path to the stumps. That third one is an extrapolation, not a measurement — the system is forecasting where the ball would have gone. Because prediction carries more error than observation, the ICC’s “Umpire’s Call” rule holds that if less than 50% of the ball is projected to hit the stumps, the on-field decision stands and the team does not lose its review.
Umpire’s Call is widely disliked by fans, but it is one of the most intellectually honest features in sport: an explicit admission that the AI’s confidence interval is wide enough that a human’s original judgement deserves the benefit of the doubt.
Where AI referee technology still gets it wrong
The failures are rarely the algorithm. They are the plumbing around it.
Wimbledon 2025 is the case study. During a Centre Court match between Anastasia Pavlyuchenkova and Sonay Kartal, the electronic line-calling system had been deactivated in error across part of the court. A ball that clearly landed out went uncalled, the point was replayed, and Pavlyuchenkova said the game had been stolen from her. The All England Club apologised, and afterwards removed operators’ ability to manually deactivate ball tracking at all.
That is the recurring pattern across every industry that deploys this kind of system, not just sport. The model was fine. A human turned part of it off, and no process caught it. The same failure mode shows up in AI-assisted air traffic control, where the hard problem is not the prediction quality but the handover between machine output and human action.
The other limitation is scope. No system in professional sport today judges intent. Whether a tackle deserves a red card, whether contact was simulated, whether a bowler’s action is legal in spirit — these remain entirely human, because they are not geometry problems.
Why every major sport keeps a human in the loop
Notice the design choice repeated across all four sports: the AI is a witness, not a judge. FIFA requires a video official to validate the offside line. MLB triggers the machine only when a player challenges. Cricket defers to the on-field umpire inside the margin of error. Wimbledon is the exception — fully automatic line calls — and it is the one that generated a public apology in its first season.
This is not sentimentality about tradition. It is risk management. Automated calls are fast, consistent and free of crowd pressure, but when they fail they fail silently and at scale, and a sport’s legitimacy rests on spectators believing the result. Keeping a human accountable for the final decision is what makes the technology acceptable to the people watching.
The same logic governs how AI is used in medicine, aviation and finance. Sport is simply the version broadcast to millions of people at once — which is why it is becoming the public’s real education in what AI decision-making looks like. If you want a sense of where else AI is quietly reshaping sport, our piece on how AI predicts sports injuries covers the side fans never see.
Frequently Asked Questions
Is AI referee technology fully automatic?
In most sports, no. FIFA’s semi-automated offside technology requires video match officials to verify the machine’s proposed offside line before the referee is informed. Tennis electronic line calling is the main fully automatic exception.
How accurate is Hawk-Eye?
Hawk-Eye Innovations puts average error at a few millimetres, and independent testing has reported a mean error of around 3.6 mm for measured ball position. Accuracy is lower for predicted ball paths, such as the projected trajectory used in cricket LBW reviews.
Why does cricket still have Umpire’s Call?
Because the projected path of the ball after impact is a forecast, not a measurement, and carries more uncertainty. If less than 50% of the ball is predicted to hit the stumps, the on-field umpire’s original decision stands and the reviewing team keeps its review.
Can AI decide fouls, red cards or player intent?
No. Current systems measure positions and trajectories, which makes them suited to offside, line calls and strike zones. Judgements about recklessness, simulation or intent remain entirely with human officials.
Does AI officiating make sport fairer?
On measurable calls, it is more consistent than human eyesight and immune to crowd pressure. It does not remove controversy — it moves the argument from whether the official saw it correctly to whether the rule and its margin of error are set correctly.
