‍#广场预测世界杯赢40000U

AI Predicts the “World Cup Top 32” with Accuracy Surpassing Humans

With the group-stage dust of the 2026 World Cup in the United States, Canada, and Mexico finally settled, the complete list of the 32 teams has been officially released. While traditional powerhouses and dark-horse teams fight fiercely on the pitch, off the field a phased “settlement” has also arrived for an AI-led “man-versus-machine prediction showdown.”

In the “World Cup Prediction Man vs Machine Battle” launched by Lenovo Group together with Migu Video, 12 mainstream domestic AI large models and human experts each predicted the outcomes of 104 matches. The final results show that the overall prediction accuracy of the AI side surpassed that of the human experts. Moreover, some AI models even nailed underdog teams such as Cape Verde—teams that have no historical World Cup data—producing certain “anti-consensus” outcomes.

Image source: Provided by Lenovo official

From May 28 until before the World Cup kicked off, Lenovo Tianxi AI, serving as the “chief convener,” sent an answer sheet titled “2026 World Cup Top 32 Team Prediction Unified Examination” to 12 mainstream domestic AI large models and human experts, requiring them to submit their own “answers” before the matches began.

On June 28 Beijing time, with the Group J closing match ending in a 3:3 draw between Algeria and Austria, the rankings of each AI prediction were also revealed. Among them, Tencent Hunyuan topped the chart with 29 teams correctly predicted to advance, followed closely by MiniMax and iFlytek Spark, each with 28 correctly predicted teams. The AI camp’s overall win rate reached 61.9%, leading the human experts by 7.3%.

Even more noteworthy is that, in the pre-tournament predictions, four of the 12 AI models boldly backed the “newcomer” Cape Verde. This “anti-consensus” prediction was verified one by one by Cape Verde’s performances: the team repeatedly held traditional powerhouses such as Spain and Uruguay to draws and advanced with an unbeaten record.

Marlin, Technical Director of SenseTime, explained to a reporter from Times Finance that the reason AI large models were able to pick Cape Verde—the “biggest dark horse”—was that the models could see through the apparent veneer of on-paper strength and dig into deeper data. Although Cape Verde is a World Cup “newcomer,” many of its players developed within Europe’s league system, and in recent qualifiers the team has performed strongly. By capturing deep data variables such as defensive discipline, counterattack efficiency, and player structure, the AI models reached conclusions that were more rational than human experience.

However, while AI demonstrated its ability to deliver “anti-consensus” picks, it also exposed limitations under extreme uncertainty. Take the match between Cape Verde and Saudi Arabia as an example: the 12 AI models’ predictions diverged into three outcomes—DeepSeek, Kimi, Jiejue, and iFlytek Spark predicted a win for Saudi Arabia; Tongyi Qianwen, China Mobile Jiutian, Tianxi AI, Tencent Hunyuan, and SenseTime Xiao Huanxiong predicted a draw; Baidu ERNIE, Zhipu, and MiniMax backed Cape Verde to win. Although the final result was a draw, none of the models precisely hit the final score of 0:0.

This phenomenon reveals a common blind spot in current AI predictions: an “overestimation of offensive firepower.” Even if five models predicted a draw, the scorelines they gave each included at least one goal. According to Lenovo’s official data, among the nine draws in the group stage, the AI’s prediction hit rate was less than 3%. Judging from this, AI large models are better at handling structured data and deterministic trends. But for a team sport like football—where real-time mentality, sudden injuries, and many other contingencies all play roles—there are still parts that AI large models cannot estimate.

This World Cup produced a stream of dark horses, making AI predictions turn out varied and wildly different. This actually exposes a weakness of large models: “convergence of underlying logic.” Because large models, in essence, are “probability compressors” that infer outcomes from historical data. In matches where team strength is clearly defined and favorites and underdogs are familiar, everyone uses the same data and naturally reaches the same conclusions. But once they face an unfamiliar team, or a match features innovative tactics or extreme styles, AI can “collectively malfunction” due to a lack of historical references. That is because they do not truly understand the game; when faced with the unknown, they only make guesses based on probabilities.

In fact, the “man vs machine battle” over World Cup outcome predictions is not a one-off performance by Lenovo alone—multiple top large model companies have already joined in. Tongyi Qianwen launched a dedicated football prediction AI assistant, which not only covers all 104 matches for users to compete alongside AI in predictions, but also simultaneously initiated the “Pitch Project”—when users accumulate a certain number of prediction points, Tongyi Qianwen will donate to build football pitches for rural schools.

Moonshot AI (Kimi) also built 300 dedicated Agents, each responsible for specific areas such as tactical analysis, player condition tracking, schedule calculation, and odds monitoring, ultimately generating a 224-page in-depth prediction report that showcases its capability to coordinate complex tasks with multi-agent collaboration. In addition, Anthropic’s Claude Fable 5 model also made predictions based on macro variables such as the tournament structure (48 teams participating, the champion needing to play 8 matches), North American summer heat, and the teams’ roster age curves.

These diversified modes of participation not only transformed AI prediction from a simple “win-loss guessing game” into a comprehensive technology showcase spanning data analysis, multi-Agent collaboration, public-welfare interaction, and macro-level inference. They also turned this World Cup “man vs machine battle” into an excellent testing ground for major companies to verify their ability to deploy large models. But at an even higher dimensional level, in complex systems such as business decision-making, macroeconomic assessment, and even social governance, AI also faces a game between “data completeness” and “real-world chaos.”