AI has wormed its way into every crevice of the 2026 World Cup. It’s dreaming up sloppified soccer jerseys, collating thousands of on-field data points, and even guarding venues in the form of robot surveillance dogs that presumably can’t be bribed with sausage.
Yet for all its busywork, AI remains blissfully ignorant on the one metric that matters: who wins and who loses.
During what must have been a slow news day Monday, the sports writers at USA Today quizzed Microsoft’s Copilot on the day’s World Cup matches. There were four contests altogether: Spain-Cape Verde, which Copilot predicted would be 3-0; Belgium-Egypt, predicted to be 2-1; Uruguay-Saudi Arabia, predicted to be 2-1; and Iran-New Zealand, predicted to be 1-0.
Perhaps as expected, the predictions seriously missed the mark. In reality, each match ended in a draw, an outcome Copilot failed to even consider as a possibility. Belgium-Egypt and Uruguya-Saudi Arabia both ended 1-1, while Iran and New Zealand traded blows in a 2-2 tie. Arguably the most devastating rebuke was Cabo Verde, whose now-viral goaltender Josimar “Vozinha” Dias stood on his head to hold a top-tier Spanish team to 0-0.
Copilot’s predictive analysis was telling. As USA Today writes, the AI model reasoned that Spain’s attackers would pepper Cabo Verde’s inadequate defenses with so many shots that they would eventually crumble, exposing an obviously disproportionate match-up. As Spain learned the hard way, that forecast was probably more indicative of the kind of buzzy media hype Copilot was ingesting than any well-crafted analysis.
That said, Microsoft’s AI isn’t the only one taking a red card. Earlier this month, journalists asked ChatGPT to predict the results of the NBA finals between the New York Knicks and the San Antonio Spurs. Though the Knicks won on game five in spectacular fashion over the weekend, ChatGPT originally pegged the Spurs as the 2026 NBA champs, declaring that San Antonio superstar Victor Wembanyama would help drag the series into game seven.
The failed predictions come after a bombshell pre-publication study showed large language models like ChatGPT and Copilot are horribly equipped to predict the outcome of sports, or even analyze important plays and games after they’ve happened.
During one test of top AI models’ ability to predict the outcome of various three- to 15-minute game segments, even the best performing model only got it right 43 percent of the time. That indicates a major performance gap in LLM’s ability to forecast real-world outcomes, even under the tightly-controlled conditions of a soccer match. As the researchers wrote: “humans reach 58.9 percent overall and remain well-calibrated, in contrast to [AI] models.”
Put together, it’s clear LLMs remain way behind on their ball knowledge. While that’s bad news for anybody hoping to score on a few World Cup prop bets, it’s even worse for a tech industry which has burned hundreds of billions of dollars trying to turn LLMs into complex reasoning machines.
**More on sports: **US Soccer Scanning Videos of Millions of Youth Players to Identify New Stars