Data or Instinct? AI on the Pitch

At the 2022 World Cup in Qatar, twelve dedicated tracking cameras were mounted under the stadium roof, capturing up to 29 points on each player's body, 50 times every second, while the ball itself carried a sensor sending data 500 times a second. When an attacker received the ball from a teammate in an offside position, the system alerted the video officials within seconds. Even so, the decision remained with humans: the referee on the pitch confirmed it, and in subtle cases, such as interference with an opponent, he judged for himself. The scene sums up what is happening: a machine that sees more precisely, and a person who holds the decision.

What do the numbers measure?

Statistics are nothing new to sport. Perhaps its most famous milestone is Michael Lewis's book Moneyball (2003), about the Oakland Athletics baseball team, who competed in 2002 on a payroll of around $44 million while the New York Yankees' exceeded $125 million, because they sought players the market undervalued using measures such as on-base percentage. That was advanced statistical analysis, not AI in the strict sense, but it opened the door to thinking in numbers.

Today models are trained on tracking of positions and ball paths to produce measures such as "expected goals" (xG): the probability that a shot becomes a goal given its location, angle and the way it was set up, and in some models defensive pressure and the goalkeeper's position. It separates a striker who scored from a poor chance through luck from one who consistently creates high-value chances.

But the measure has declared limits: it reflects how an "average" player would do in the same situation, not the skill of the particular player, and its value for the same shot differs from one model to another. It is a tool of description, not a final verdict.

Injury prediction: promising but immature

It looks like an application whose usefulness is beyond dispute: wearable devices tracking training load and movement, and a model warning of strain that may precede a tear. But the scientific evidence is more modest than the publicity suggests. A systematic review (2021) found, among 249 studies, only 11 meeting its criteria; their samples ranged from 25 to 363 participants, and model accuracy varied from a near-random result (52 per cent accuracy) to a good one (AUC = 0.87). The authors concluded that the models can identify the players at highest risk and the most important risk factors, that the quality of the studies needs improving, and that the interpretation of models deserves more attention.

Here the question of privacy becomes clear: who owns the data of a player's body? And to what extent may a club know about him what he does not know himself? A legitimate question arises, not a documented incident: what if a predicted injury-risk score were used to exclude a player from a transfer or contract before he has even fallen ill? It follows that the party that owns the data and the permitted purpose should be settled in advance, and that the player should give informed consent.

The model's limit: it learns from the past

A model learns from records of what has happened, so it leans towards what has worked before. This is a logical inference, not a statistical result: if all clubs adopt similar sources and models, their decisions may become similar, and room for tactical diversity narrows. And the creativity that breaks expectation is, by definition, what the model does not find in its archive. The wise course is to leave human instinct its authority over ambiguity, and the machine its authority over pattern.

What this means for us

On 11 December 2024, FIFA's Extraordinary Congress in Riyadh confirmed the Kingdom as host of the 2034 World Cup. This event will require more than stadiums: local analytical minds that know how to ask. The tools are available to whoever buys them; the real difference lies in the quality of the questions and the soundness of interpreting the answers, and in building a team that understands the tool's logic and limits before it buys it.

Sound rules

• The algorithm is an aid to decision, not a substitute for it.
• Diversify data sources and modelling methods.
• Protect the player's dignity and biometric data; they are not sold or used without his consent.
• Invest in understanding the tool alongside buying it.

When the stands erupt at a goal in the last moments, it is not because its probability was high, but because people rejoiced at what people like them had done. The best thing technology can do is reveal the beauty of the game without claiming to have grasped it whole.