In the 2014 World Cup quarterfinal, Brazil’s best player took a knee into his lower back and fractured a vertebra. Neymar was out for the rest of the tournament. Two days later his teammates walked onto the field for the semifinal having just watched it happen, and lost 7-1 to Germany, one of the most lopsided results in World Cup history.
Nobody needs a model to explain that losing your best player hurts a team. The harder question is whether anyone could have seen it coming, and whether it would have changed anything if they had.
This is the fourth post in our series on AI and machine learning at the 2026 World Cup. The first post laid out the two ways this technology is used in the tournament, on the pitch and off it. The second asked whether a model can predict who wins. The third looked at how teams use data to find eligible players nobody else was tracking. This post stays off the pitch but turns to a different question: once a team has found the players it wants, how much data can help keep them healthy enough to use.
Fitness talk isn’t the first conversation, but it becomes one
I asked Finn McCallum, the player and coach I’ve been talking with throughout this series, whether data informs how players, coaches, and trainers talk about fitness and fatigue, or whether that’s still mostly instinct and feel. He said the first time a player mentions his leg feels off, that conversation is rarely data-driven. But once the same complaint keeps coming up, that’s usually what prompts a coach or trainer to start pulling workload numbers and tracking data instead of just going by feel.
Instinct still comes first. Data enters once a nagging issue turns into a pattern, or once a team has to make a call that affects a whole tournament instead of one match.
A retractable pitch and a pattern worth testing
Real Madrid’s stadium has a retractable pitch, real grass grown in sections that slide underground into a greenhouse when there’s no game, so the venue can double as a concert space. It’s the kind of thing you’d never guess had anything to do with injuries. According to Finn, the club has recorded more ACL tears in the two or three years since installing it than in the decade before. He was careful to say it’s not proven, just a strong hunch inside the sport, but the mechanism makes sense. A pitch built to retract can’t have much depth to it, just a few inches of turf over a slab, and there’s a real difference between planting a cleat into that and into a few feet of natural soil.
This is the kind of pattern computer vision is suited to test, not because a camera can see turf depth, but because it can see what supposedly causes the tear: how a player moves in the moment before their body fails. A friend of Finn’s recently sent over a paper on using body pose estimation, tracking joints and limb angles frame by frame, to flag risky movement patterns before an injury happens. Soccer already has one of the highest ACL injury rates of any sport, the byproduct of constant cutting, pivoting, and decelerating from full speed, so there’s real value in catching the warning signs early. An ACL tear, by the numbers, usually isn’t a collision. It’s a plant and a turn, the motion pose estimation is built to catch, less useful for predicting that a player will get hurt than for noticing the mechanical signature that tends to precede it, across thousands of hours of footage a human eye would get tired of watching by minute forty.
Rotation is a data decision wearing a coaching decision’s clothes
The clearest real example from this World Cup didn’t involve a camera at all. Norway reached the knockout round for just the second time in the country’s history, their first since 1998, and once they’d already secured it with a game to spare, manager Ståle Solbakken sat nine of his regular starters, Erling Haaland among them, for the final group match against France. Haaland told reporters beforehand that he “couldn’t care too much” about the result, since France would probably win the group and the tournament anyway. Norway lost 4-1. They then beat Ivory Coast and stunned Brazil in the knockout rounds before losing to England in the quarterfinal, the best World Cup run in the country’s history.
Norway had already gotten what it needed and protected its most valuable asset from a game with nothing left to gain, then went on a run that outlasted every earlier expectation. Compare that to Portugal, whose center-midfield trio, Bruno Fernandes among them, had played almost every minute of every match since the previous August between club and country commitments. By the time the World Cup arrived, none of them looked like themselves, and Portugal went out well short of expectations. No one’s suggesting an algorithm would have kept Bruno Fernandes on the bench. But workload tracking, the same kind of data that flags an ACL risk, is the tool that turns “he seems tired” into a specific number: he’s played 58 competitive matches since August, and here’s what that has historically done to a player’s output.
The workload numbers don’t always point the same direction, either. Michael Olise and Harry Kane spent the season as starters for a Bayern Munich team that went deep in every competition it entered, then turned around and played seven or eight of their countries’ eight possible World Cup matches for France and England, still performing at a high level. Workload data can flag who’s running on empty. It can’t tell a coach in advance which players will hold up under the same load anyway.
The decision an algorithm should never be allowed to make
Ahead of the 2018 World Cup, France center-back Samuel Umtiti was carrying a knee injury from the club season. He delayed surgery, managed the pain with an injection, and played through it anyway. He scored the winning goal in the semifinal against Belgium with a header off a corner. France won the whole tournament. Umtiti was never quite the same player again, and by his early thirties he wasn’t really playing much at all.
Ask any fan what a coach should do if a data model flags a star player as a serious injury risk going into a semifinal, and you’ll get some version of the same answer: it depends how badly you want to win. That’s a call a human has to make, not a model. A system can flag that a player’s movement resembles the last twenty players who got hurt before it happened. It can’t say whether this particular player, in this particular tournament, should play through it anyway, and no dataset was ever built to answer that.
This same divide runs across AI and machine learning generally, in soccer and elsewhere. The tool can compress a mountain of footage into a signal a human couldn’t have spotted alone. What happens with that signal, whether a coach benches a starter, whether a player takes the injection anyway, still belongs entirely to the person making the call. The technology doesn’t inherit that responsibility just because it did the noticing.
A 48-team tournament raises the stakes
This World Cup runs longer than any before it, up to eight matches instead of seven if a team goes all the way, but the recovery windows haven’t grown with it: roughly a week between group games once travel back to base camp is counted, and about four days between knockout matches, close to what Argentina had during their run to the title in 2022. More games, same body. That’s not a reason to fear this kind of technology creeping into a sport that’s always run on instinct. If anything, it’s the reason rotation, load management, and injury-risk modeling carry more weight now than they ever have, not to take the decision away from the people who’ve always made it, but to make sure they’re making it with better information than “he says he feels fine.”
A coach who’s spent twenty years reading this sport doesn’t need replacing. He just gets a better set of eyes for the eighty-seven minutes a game he can’t watch closely enough on his own.
