Chinese AI Dominates World Cup: 66% Accurate in Fan Prediction Contest

Artificial intelligence systems created by 12 leading Chinese tech companies have demonstrated higher accuracy than human fans in predicting match outcomes during the 2026 FIFA World Cup, achieving an overall accuracy rate of 66 percent so far.

The prediction competition, jointly organized by a major Chinese computer and technology company and a popular streaming service, pit these 12 AI models against approximately 35 million human spectators. After the first 100 matches, the humans achieved a correct prediction accuracy of 59 percent, though they led early on, they have since fallen behind the AI models.

The AI group includes DeepSeek, Qwen, Jiutian from China Mobile, Baidu’s Ernie Bot, Tencent’s Hunyuan, Kimi, Zhipu AI, MiniMax, JieYue Star, iFlytek Spark, SenseTime Xiaohuan, and Lenovo’s Tianxi AI. They correctly predicted the outcomes of 788 out of 1,200 matches after the quarter-finals’ Argentina versus Switzerland game concluded.

Hu Yanping, a distinguished professor at Shanghai University of Finance and Economics, commented that the AI’s overall predictive accuracy falls within his expected range of 60 to 80 percent. “This World Cup prediction contest serves as a real-world testing ground for assessing the reasoning ability and limitations of AI models,” he said. “Understanding both their strengths and shortcomings provides valuable insights for future improvements.”

As the tournament advanced, more data became available regarding team form, squad rotations, group standings, and tactical approaches, which enhanced AI’s ability to process and analyze large amounts of information. Meanwhile, human predictions tend to be influenced by factors such as team popularity, personal preferences, and emotional biases, the report highlighted.

However, the AI models struggled with unexpected draw outcomes and upsets during the initial phase, making incorrect predictions for 11 draws and four matches won by lower-ranked teams, which are often called upsets. These blind spots largely resulted from the models’ difficulty in accurately predicting matches that end in a draw or surprise results.

Hu pointed out that AI’s challenge in predicting draws mainly stems from its strong capacity to evaluate relative team strength but a limited ability to determine if that strength can translate into goals within the match duration—a task that warrants further research. The models’ inability to foresee unpredictable in-game events is within expected performance bounds, he added.

Additionally, AI predictions tend to be biased towards traditional powerhouse nations, leading to uniform forecasts that often miss unexpected defeats of top teams. All 36 predictions made by the AI for knockout stage matches involving Germany, the Netherlands, and Brazil proved inaccurate.

Predicting exact scores was particularly difficult for the models. Out of 1,200 score predictions, only 145 were correct, resulting in a 12 percent accuracy rate. The best-performing model correctly predicted the exact score in just 17 percent of cases.

All 12 models favored Brazil to win the tournament, but the team was eliminated in the Round of 16. Conversely, none of the models correctly predicted England’s progression to the semi-finals.