March Madness brackets are notoriously difficult to predict, but one reporter decided to see if artificial intelligence could improve his chances this year. He created two brackets in parallel using OpenAI’s ChatGPT and Google’s Gemini to test the models’ ability to analyze the men’s tournament. The experiment, now in its third year, aimed to see if the AI tools could handle the complex data required for a high‑stakes basketball prediction.
Comparing the Models
ChatGPT identified key factors for success, noting that 21 of the last 23 champions ranked in the top 21 for points per possession on offense. The model provided a list of themes with evidence to back them up.
Gemini offered a more structured approach. Using a feature called NotebookLM, the platform analyzed sources like KenPom ratings, NCAA NET rankings, and ESPN data. It created a five‑point championship model that included specific metrics such as a minimum of 33% team three‑point shooting and AdjEM, which measures adjusted efficiency margin.
Related: Chelsea signs Circle Internet Group as shirt sponsor
Both systems highlighted the advantage of “chalk,” or picking higher‑seeded teams. They suggested that NIL money and the transfer portal help powerhouses restock talent more easily. Gemini added a unique data point called “Neutral Court Translation Analysis,” or NTS, which measures how well teams perform away from home.
Another metric, “The Gauntlet,” noted that 88% of recent champions maintained a NET ranking in the top 25 and played a moderate pace of 62 to 69 possessions per game.
Gemini occasionally misplaced teams. It put the South Region’s No. 3 seed Illinois in the Midwest bracket.
Related: Jake Paul pushed Joshua harder than Prenga says MVP aide
ChatGPT also had trouble keeping track of winners, sometimes pitting an 11‑seed against the No. 5 seed instead of the No. 3 seed when an upset occurred. The reporter intervened manually, prompting the AI game by game to correct these errors.
Diverging in the Final Four
Both systems eventually agreed on the same Final Four teams: Duke vs. Houston and Arizona vs. Michigan. However, they split on the semifinals. ChatGPT projected Michigan to defeat Houston, while Gemini favored Arizona over Duke.