The contest between artificial intelligence and humans has moved from the chessboard to the football pitch. In a World Cup prediction challenge, a panel of 12 AI models collectively achieved a 65.7% accuracy rate in forecasting match outcomes over 100 games, surpassing the 58.9% average of the public. This marks a significant evolution from the famous AlphaGo victory a decade ago.
The earlier triumph of AI occurred within the closed, rule-bound system of the board game Go. In 2016, AlphaGo's victory over champion Lee Sedol demonstrated that machines could reach or exceed top human levels in complex but fixed environments. The current challenge, however, unfolds in the dynamic and unpredictable arena of international football.
During the 2026 FIFA World Cup, LENOVO GROUP and Migu co-launched a "World Cup Prediction Man vs. Machine" event. The LENOVO GROUP Tianxi AI served as the convener, leading a team of 12 models including DeepSeek, Qwen, China Mobile Jiutian, Baidu ERNIE, Tencent Hunyuan, Kimi, Zhipu, MiniMax, Jieyue, iFlytek Spark, and SenseTime's Xiao Huan Xiong. This AI coalition competed against sports pundits, professionals, and the general public in predicting the 32 qualifying teams, and later the outcomes of all 104 matches, including exact scores and the eventual champion.
This time, the machines confront a continuously evolving open system. While historical team data and tactics provide a foundation for judgment, a football match is not a game that can be reliably predicted by data alone. Player form, in-game decisions, tactical shifts, and even random events can dramatically alter the course of a game. The 104 matches represent 104 independent events with constantly changing conditions and incomplete information, not a single problem to be solved repeatedly.
The goal was not to find an infallible oracle, but to observe the capabilities, differences, and common limitations of multiple AI models in a real-world scenario through a hundred consecutive data points. LENOVO GROUP recently released a report titled "Hundred-Game Observations on the World Cup Prediction Man vs. Machine – AI's Advantages, Differences, and Boundaries," providing the industry's first quantifiable, verifiable, and continuous sample based on data from the first 100 matches.
AI's Collective Victory Over Humans
After 100 matches, the AI panel's accuracy in predicting match results was higher than that of human participants. This advantage was not established from the outset. Initially, on June 13, AI accuracy was only 43.8%, trailing the public's 54.4% by 10.6 percentage points. By June 18, AI accuracy rose to 48.6%, surpassing the public's 45.0% for the first time. From that point, AI maintained an overall lead, widening the gap to 6.8 percentage points by the 100th match.
This outcome can be described as an "elevation of the average." AI excels at raising the long-term average in high-frequency, continuous, and informationally complex scenarios, rather than providing a certain answer for every single match. AI's strength lies in gradually reducing judgment errors across a large volume of games, whereas humans might be very accurate in a single instance but struggle to maintain consistency over a long tournament.
A closer look reveals that the "AI victory" was the result of 12 different models taking turns to perform, with no single model dominating throughout. The report notes that no AI was universally dominant; different models led during different stages. For instance, Tencent Hunyuan correctly predicted 29 of the 32 qualifying teams. After 72 group stage matches, China Mobile Jiutian and Tencent Hunyuan were tied for first with 49 correct calls. In the round of 32, Qwen led with 14 correct predictions, while in the round of 16, Tencent Hunyuan, Kimi, and iFlytek Spark each got 7 correct.
Through the first 100 matches, China Mobile Jiutian led the cumulative tally with 71 correct predictions, followed by Tencent Hunyuan and Qwen with 70 each, and LENOVO GROUP Tianxi AI and SenseTime's Xiao Huan Xiong with 69 each. The leader was only one game ahead of second place, and the top six were separated by just three games. The model with the longest consecutive correct prediction streak (Qwen with 14) did not top the overall 100-game leaderboard, highlighting that short-term accuracy and long-term error reduction are distinct capabilities.
The World Cup produces one champion team, but this AI prediction experiment is difficult to summarize with a single "champion" model. Different answers emerge depending on whether one looks at the initial 32-team prediction, a specific knockout round, the longest streak, or the cumulative total. Shorter samples are more influenced by single-game results, while longer samples emphasize the ability to consistently minimize errors.
Collectively, this data shows that AI's ability to overtake and maintain an advantage over humans in World Cup prediction was not due to one "strongest model," but to 12 models each contributing during different phases. The rising AI prediction curve is essentially an average built from 12 different answer sheets. Different models exhibit clear strengths in specific scenarios and stages, suggesting AI capability should be measured in continuous, layered, and verifiable data.
The Draw: AI's True Adversary
The data further reveals that AI's predictive advantage was not evenly distributed. For the 76 matches that ended in a win or loss, the 12 AIs achieved an 81.0% accuracy in predicting the winning side, significantly higher than the public's 71.8%. However, for the 24 matches that ended in a draw, AI and public accuracy plummeted to 17.0% and 16.0% respectively. The clear advantage AI held in predicting wins/losses almost completely vanished when it came to draws.
Cape Verde's performances were a classic example. Before holding Spain to a 0-0 draw and Uruguay to a 2-2 draw, not a single AI predicted a draw in either match. The AI models recognized the disparity in team strength but failed to adequately estimate the possibility of a weaker team securing a draw through defense, tempo control, and limited counter-attacks.
Both AI and humans tend to look for a winner and systematically underestimate the possibility of a stalemate. A draw represents a situation where the on-field process undermines paper strength. A strong team may dominate possession and shots but fail to score, while a weaker team may slow the game through dense defense. For relatively weaker teams, defensive counter-attacking is the most realistic and effective tactic in a World Cup, minimizing potential heavy losses while creating chances against an attacking opponent.
Of the 15 matches where all 12 AIs were collectively wrong in their predictions, 11 were draws, and the other 4 were won by the pre-match underdog. Draws accounted for nearly three-quarters of these collective failures, constituting the most stable and clear challenge for AI prediction.
Beyond draws, AI also showed collective misjudgments in other scenarios. In the elimination matches of Germany, the Netherlands, and Brazil, the 12 AIs made 36 collective directional predictions and got none correct. The models could recognize the historical strength advantage of traditional powerhouses but failed to capture the risk of that advantage failing to materialize in a single knockout match.
When historical records, team reputation, and squad strength create a clear hierarchy, models tend to converge on the same prediction, potentially underestimating reversals caused by specific tactical matchups, on-the-day form, and psychological factors. In the knockout stages, the historical advantage of traditional strong teams leads to highly consistent pre-match judgments, which in turn exposes a common bias when that advantage fails in a single game.
The report notes that in 40 of the first 100 matches, all 12 AIs unanimously predicted the same outcome. In 10 of those 40 matches, this unanimous consensus was wrong. "The most dangerous moment for AI may not be when opinions are divided, but precisely when all models believe they cannot be wrong." High consensus indicates multiple models may be extracting similar conclusions from similar signals, but when key variables are not incorporated, it can instead amplify a shared blind spot.
AI Wins on Average, Humans Create Peaks
If following the logic of the AlphaGo era, the World Cup prediction contest might also seek a definitive answer: who won, who lost. However, the two tasks are fundamentally different. AlphaGo aimed to continuously approach an optimal solution within fixed rules, while World Cup prediction requires judging the most likely outcome before the answer is known. This distinction is precisely where the value of this experiment lies.
A fan from Chongqing provided a contrasting sample. Before the tournament, he correctly predicted 31 of the 32 qualifying teams. According to the report's simplified probability models, this corresponds to reference odds of approximately 1 in 181 million under a pure random selection assumption and 1 in 212,000 assuming group knowledge. This made him the most outstanding individual among over 35 million participating users, while the top AI result was Tencent Hunyuan's 29/32. His judgment stemmed from long-term experience watching football—in the human world, it holds true that individuals with more experience and information access can, at certain moments, appear "smarter."
The Argentina vs. Switzerland match (3-1) provided another example. All 12 AIs predicted an Argentina win, but none predicted the exact 3-1 scoreline. At a related event, two ordinary audience members did predict that exact score. This does not negate AI's average advantage over 100 matches in directional prediction, but it again shows that in the more demanding task of predicting exact scores, an individual human can still, in a single moment, surpass all models.
This does not mean humans as a whole are more accurate than AI. AI's lead on the long-term average line across 100 matches demonstrates that models are developing the ability to process large amounts of structured information and consistently control errors. "AI elevates the overall average level, while humans leave behind the most unforgettable peak moments of this summer," the report states. The two are not mutually exclusive but together constitute different forms of judgment capability in the real world.
Subjecting AI to Public Scrutiny
The LENOVO GROUP report presents AI's average advantage, common blind spots, model differences, and human peaks. AI can raise the baseline of judgment but cannot provide all the correct answers.
This forms the greater significance of this "Man vs. Machine 2.0" challenge: subjecting AI to open testing in the real world. If the value of AI is no longer to provide a perpetually correct answer, but to improve judgment quality in uncertain environments, then how to verify this capability becomes the next key question. The report, in some ways, addresses this.
The primary value of this material lies in its methodology. For some time, the validation of large model capabilities in the domestic market has long relied on laboratory leaderboards and standardized tests. While useful for横向比较 technical metrics, these often remain distant from real application scenarios. The World Cup schedule provided a rare observational perspective.
On a deeper level, this experiment also reveals the potential for平移 of intelligent analytical capabilities. Previously, systematic information integration and trend judgment were primarily concentrated within专业机构, with high access costs for普通用户. The continuous hundred-match data indicates that AI is正在 pushing analytical capabilities to a broader audience with lower barriers—users without deep专业背景 can also form relatively rational decision judgments. This trend is not limited to sports but is同样正happening in scenarios like financial investment, consumer choice, and market analysis.
As the convener of the event, the performance of LENOVO GROUP Tianxi AI also addresses a business-relevant proposition. It did not achieve first place in any single category, yet it ranked within the top four across four key milestones: the pre-tournament 32-team prediction, the 72 group stage matches, the round of 32, and the cumulative 100-game leaderboard. It was the only model among the 12 that did not fall behind at any stage. For enterprises integrating intelligent systems into supply chains, customer service, and decision support systems, this kind of long-term reliability often holds more practical significance than occasional moments of brilliance.
AI does not need to be an infallible "prophet." Its more important role is to help people acquire information faster, understand possibilities more accurately, and make better choices in complex environments. In reality, whether in business operations, market analysis, or an ordinary person's消费选择, many problems are closer to football matches than to Go. The choices people can make under重大突发 moments are often只有一次. The world is fascinating precisely because it cannot be fully predicted. If all variables could be exhausted by algorithms,竞技 would lose its charm, and our商业 would lose its value.
Comments