12 Chinese AI Models Beat 35 Million Human Fans in World Cup Prediction Contest With 66% Accuracy

A World Cup prediction contest co-hosted by Lenovo Group and Migu Video has delivered striking results: 12 Chinese AI models posted 66% accuracy on win-draw-lose predictions across the first 100 matches, outperforming approximately 35 million human fans who achieved 59% accuracy. The human group led after the first seven days of the tournament but has remained in second place ever since.

The AI lineup includes DeepSeek, Qwen, China Mobile's Jiutian, Baidu's Ernie Bot, Tencent's Hunyuan, Kimi, Zhipu AI, MiniMax, JieYue Star, iFlytek Spark, SenseTime Xiaohuan, and Lenovo's Tianxi AI. Together they got 788 of their 1,200 predictions correct after the 100th fixture.

Hu Yanping, distinguished professor at Shanghai University of Finance and Economics, told Yicai that the 66% accuracy aligns with her prior projection of 60% to 80%. She described the contest as a practical testbed for evaluating AI reasoning capacity and limitations, noting that both strengths and shortcomings provide insights for optimization.

As the tournament progressed, richer data on team form, squad rotations, group standings, and tactical tendencies gradually emerged, amplifying AI's strength in aggregating and processing large volumes of information. By contrast, human judgments were more easily skewed by team popularity, personal fandom, and emotional bias.

However, the AI models showed clear blind spots. All 12 gave wrong predictions for 11 fixtures that ended in draws and for four upsets won by lower-ranked teams. The models' general deficiency in predicting draws stems from their ability to gauge relative team strength paired with an inability to judge whether that strength translates into goals within match duration. All 36 forecasts for Germany, the Netherlands, and Brazil in elimination rounds turned out wrong due to the models' historical bias toward traditional powerhouse nations.

Score prediction proved to be the models' biggest weakness. Out of 1,200 exact score predictions, only 145 were correct — a 12% accuracy rate. Even the top-performing model only managed 17%. All 12 models predicted Brazil would win the title, yet Brazil was eliminated in the Round of 16, and none picked England to reach the semifinals.