How accurate has AI been predicting World Cup game outcomes?
Before this tournament kicked off, a previous MediaFeed article on AI World Cup predictions noted that every major model placed Spain, Argentina, and France in their top tier, and that four of seven frontier AI models picked Spain as champion. Several weeks into the 2026 World Cup, most of that consensus is holding up. Some of it isn’t.
The following are the predictions, what actually happened, and where the gap lies.
What every model agreed on
Seven frontier AI models were handed the same 2026 World Cup draw. Four picked Spain, three picked Argentina. Every model placed Spain, Argentina and France in its top tier. That consensus broadly survived the actual tournament. France, Spain and Argentina all reached the semifinals. England, which the models slightly underrated, also made it. Brazil and Portugal, which several models had going deep, did not. Right about the shape. Wrong about several of the names.
The one model that went its own way
Nvidia’s Nemotron model arrived with formations, pressing intensity, and expected-goal rates for each team, rating Ecuador’s defense above Germany’s, and predicting Turkey would win Group D, with the US finishing last. Neither happened. But Nemotron’s overall semifinal bracket (Spain, France, Argentina, and Brazil) was three-quarters right, which is better than most human analysts.
The daily prediction tools and their track records
Several platforms publicly tracked individual match predictions throughout the tournament. France beat Morocco 2-0 was a perfect call. Spain beat Belgium 2-1 and was rated three out of four. DeepSeek achieved 81% accuracy on individual match predictions, significantly higher than the 67% achieved by Al Jazeera’s Kashef model at the 2022 World Cup and the 71% reached by similar models in 2014 and 2018. The improvement reflects four years of additional training data, not a fundamental shift in how predictable football is.
What the models consistently missed
Opta described Cape Verde’s round of 32 run as expected to end emphatically against Argentina. It wasn’t emphatic. As our previous MediaFeed article on World Cup host city stories noted, Cape Verde goalkeeper Vozinha’s mother had her visa fast-tracked so she could watch him play; she was in the stands for Cape Verde’s 2-2 draw with Uruguay, a result no model predicted. The models got the eventual outcome right. They got the texture completely wrong, which is not a small distinction in a sport where texture determines whether a match ends 1-0 or 2-1.
The mathematician with three in a row
German mathematician Joachim Klement correctly predicted three consecutive World Cup winners using a model incorporating GDP per capita, population size, and football infrastructure. His 2026 pick was the Netherlands. He describes his methodology as completely irrational and like playing the lottery. Netherlands did not reach the semifinals. He has been right three times and wrong once.
The bottom line
The AI models predicted the tournament’s shape more accurately than its stories. Spain, France, Argentina and England in the semifinals was broadly what the consensus foresaw. What it didn’t foresee was how those results arrived. The numbers are getting better at predicting which team advances. They are not getting better at predicting what it means.
Ask us! What questions do you have about content, strategy, pop culture, lifestyle, wellness, history or more? We may use your question in an upcoming article!
Related:
Like MediaFeed’s content? Be sure to follow us.
This article was syndicated by MediaFeed.co.
