Anthropic's AI assistant Claude has run 50,000 simulations of the World Cup, drawing on match data going back to 1872. The exercise was part of a test of AI-assisted forecasting in a prediction contest.
The scope of the simulations
Each simulation modeled the entire tournament from group stage to final, using historical results to inform probabilities. The dataset covers more than 150 years of international football, including matches from the first official international game in 1872 between Scotland and England. Claude processed that data to generate thousands of possible outcomes, weighting factors like team form, head-to-head records, and tournament history.
Fifty thousand runs is a large sample for a single tournament. By comparison, many forecasting models run a few thousand simulations. The scale suggests Anthropic wanted to test not just accuracy but consistency — how often the same team wins across different random seeds.
Why the historical data matters
The World Cup itself began in 1930, but the dataset reaches back decades earlier. That includes early British Home Championship matches, Olympic football tournaments from the early 1900s, and other international friendlies. For teams with shorter histories — like newer national sides — the model likely relied on more recent data or regional patterns.
Using data from 1872 means the simulations account for eras with very different styles of play, rules, and competitive balance. A match from 1900 is not directly comparable to a 2022 World Cup game. The model would have to adjust for that drift, though Anthropic has not detailed how it handled the changing nature of the sport.
The prediction contest
The simulations were part of a broader test of AI-assisted forecasting. The contest pitted Claude's predictions against other methods — possibly human experts, statistical models, or both. Prediction contests for major tournaments are common, with participants ranging from sports journalists to data scientists. This one specifically tested whether an AI assistant could improve forecasting accuracy.
Anthropic has not released the results of the simulations or said how they compared to other entries. The company also hasn't disclosed whether the same approach will be used for future tournaments, like the next Women's World Cup or continental championships.
What is clear is that the test pushed Claude beyond its usual text-generation tasks. Running 50,000 simulations requires significant compute and a structured reasoning process — not just predicting the next word but calculating probabilities across a complex bracket.
Whether Claude's simulations outperformed human forecasters remains an open question. The contest results have not been made public.
