FOXBOROUGH, United States. All 12 Chinese AI systems participating in a Lenovo and Migu Video project to predict matches at the 2026 World Cup picked Germany before its meeting with Paraguay. The Round of 32 match finished 1-1 after extra time, before Paraguay won the penalty shootout 4-3.
Sina Finance reported that none of the 12 systems backed Paraguay. The Associated Press described the result as the biggest upset of the 2026 World Cup.
Which systems took part
The “World Cup Prediction: Human versus AI” challenge includes Lenovo Tianxi AI, DeepSeek, Kimi, Baidu’s ERNIE Bot, Qwen, Zhipu AI, MiniMax, Tencent Hunyuan, iFlytek Spark, SenseTime’s Xiao Huanxiong, StepFun and China Mobile’s Jiutian. The project covers all 104 matches and compares the systems’ predictions with user selections.
Before Germany played Paraguay, Kimi predicted a 3-1 German win, DeepSeek and Qwen forecast 3-0, while ERNIE Bot and StepFun chose 2-0. The other participants also favoured Germany, according to Sina Finance.
How Paraguay eliminated Germany
Julio Enciso put Paraguay ahead in the 42nd minute and Kai Havertz equalised in the 52nd. Jonathan Tah appeared to score in extra time, but the goal was disallowed after video review because of a foul on goalkeeper Orlando Gill.
Gill made two saves in the shootout and Tah sent Germany’s final attempt over the crossbar. José Canale converted the decisive kick. AP reported that Paraguay became the first team to defeat Germany in a World Cup penalty shootout. Germany entered the match ranked 10th by FIFA and Paraguay 41st.
One match does not erase the earlier results
The unanimous miss does not mean every prediction in the project failed. Organiser figures cited by Sina Finance showed that after 66 matches, the AI systems had selected the correct result direction 61.9% of the time, compared with 54.6% for participating users.
During the group stage, Tencent Hunyuan correctly identified 29 of the 32 teams that advanced. MiniMax and iFlytek Spark named 28 each, while DeepSeek, Zhipu AI and Lenovo Tianxi AI selected 27.
What the unanimous prediction shows
The result demonstrates the limitation of a point prediction, but it does not measure the overall capabilities of a language model. The full prompts, data sources, prediction timestamps and degree of independence between systems were not disclosed. It is therefore unclear whether the consensus reflected similar methods, shared information or Germany’s obvious status as favourite.
The episode also illustrates the difference between the most likely outcome and a guaranteed result. A reasonable selection of the favourite can still be wrong in a single match, particularly in a knockout game decided by penalties. The same caution applies when assessing Chinese models in specialised benchmarks or new projects involving language models and AI agents.
Sources: Sina Finance on the 12 predictions, Lenovo and Migu Video challenge data, and the Associated Press match report.
The image was generated with artificial intelligence for Cifrum.kz and is conceptual. It is not a match photograph and does not depict real football players.
Comments on this article