AI models from OpenAI, Anthropic, and Google competed in NCAA March Madness bracket predictions as part of a Wall Street Journal office pool, outperforming most human participants after the first round while keeping strong chances of winning.
Despite an initially shaky start, where the models failed to understand the bracket structure, they quickly corrected course and began making calculated picks, including strategic upsets and crowd-diverging choices that gave them a clear edge in the rankings.
Details
The three models entered the 124-person pool under pseudonyms after being provided with scoring rules, statistical data, and public pick distributions, with a clear objective: not to predict the most likely outcomes, but to maximize winning probability.
- Claude spent 12 minutes analyzing the bracket and chose Illinois as champion, diverging from popular picks
- ChatGPT and Gemini selected Michigan, favoring a safer approach
- Initially, all models misread the bracket format and produced impossible matchups before correcting themselves
As games began, the models applied risk-management-like strategies:
- They targeted calculated upsets and succeeded in several cases
- They leaned heavily on favorites to minimize downside risk
- They incorporated external factors such as injuries and team developments
After the first weekend, all three AI models maintained perfect Final Four predictions, avoiding major pitfalls that eliminated many human entries, especially after a top-seeded team was knocked out.
Current projections suggest Claude has the strongest chance of winning—particularly if Illinois takes the title—while ChatGPT and Gemini remain in contention with narrower paths.
What’s Next?
Attention now shifts to the later rounds to see whether AI models can secure an outright win, especially if a potential Michigan–Illinois final matchup materializes.
