What happened

Ahead of the 2018 World Cup in Russia, a range of AI systems, from machine learning betting models to academic prediction engines and corporate AI demos, attempted to forecast the tournament’s outcome. 

Many models chose favourites such as Germany, Spain, Brazil, or France for entirely different probabilistic reasons, yet failed either to predict knockout-stage dynamics or, in many cases, the eventual champion (France). 

Goldman Sachs initially predicted a Brazil vs. Germany final with a Brazil victory. Dortmund/Ghent/Munich academic team's AI simulated the tournament 100,000 times, predicting Spain as the most likely winner, followed closely by Germany.

Some predictions were published widely through the media and technology company communications, creating an impression of precision and reliability. 

The incident revealed the inability of these systems to account for rare events, tactical surprises, and high-variance outcomes inherent in international football, leading to misplaced trust and, in certain cases, financial or reputational losses for those who relied on the forecasts. 

Why it happened

The failures appear to have stemmed from several factors:

• Model limitations: Football outcomes depend heavily on non-quantifiable factors (injuries, referee decisions, team morale), which most models failed to capture.

• Small datasets: International tournaments provide limited historical samples, reducing statistical robustness.

• Overfitting and spurious correlations: Some models leaned heavily on past performance or ELO-type metrics that did not generalise to an unpredictable tournament.

• Opaque methodology: Many commercial prediction systems lacked transparency about their inputs, assumptions, and uncertainty ranges, leading users to overestimate their accuracy.

• Corporate incentives: Tech companies promoted their models as showcases of AI capability, sometimes downplaying uncertainty or methodological weaknesses.