Ask any of the leading artificial intelligence systems to fill out a March Madness bracket, and they’ll all give you the easy answer: Duke.
That’s what experiments run by The Daily Wire and CBS Sports predicted this week. The Daily Wire fed tournament data to four AI models — ChatGPT, Anthropic’s Claude, Google’s Gemini, and Perplexity. CBS Sports ran its own test with ChatGPT, Gemini, and Microsoft Copilot.
Every model arrived at the same pick: The Blue Devils, who enter the tournament as the No. 1 overall seed in the NET rankings, the AP Poll, and KenPom efficiency ratings.
“ChatGPT, Gemini, and Perplexity all had Duke beating Arizona in the final,” The Daily Wire reported. “The models read the same numbers, weighted the same KenPom rankings, and arrived at the same safe, predictable conclusion: the No. 1 overall seed with the nation’s best defense and best player wins it all.”
But when The Daily Wire pushed their four models harder — feeding them historical upset rates, injury data, regional geography advantages, and instructing them to act as “upset analysts” — the brackets got interesting in a hurry.
All four models abandoned Duke as their champion. Three of the four landed on Houston as a Final Four team. ChatGPT and Gemini both picked the Cougars to win it all, and predicted Kingston Flemings as the tournament’s Most Outstanding Player
All four models also agreed on the same first-round upsets. They each picked 12-seed Akron to knock off 5-seed Texas Tech — citing Texas Tech’s devastating loss of star J.T. Toppin to a torn ACL and Akron guard Tavari Johnson’s 20-plus points per game average. All four also tabbed 11-seed VCU to upset North Carolina, a team missing projected top-5 NBA pick Caleb Wilson due to a broken thumb. And all four liked 9-seed Iowa over 8-seed Clemson, noting the Hawkeyes are ranked 11 spots higher in KenPom despite the lower seed.
CBS Sports experiment with ChatGPT, Gemini, and Copilot found similar patterns. Copilot flagged Akron over Texas Tech and VCU over North Carolina as high-confidence upsets, and added 13-seed Hofstra over 4-seed Alabama to its list of first-round shockers. The CBS version of ChatGPT — apparently using a less aggressive upset prompt — went with a more conservative Final Four of Duke, Arizona, Houston, and 6-seed Tennessee.
Gemini, in the CBS test, looked further down the bracket for potential Cinderella runs, flagging 10-seed Santa Clara, 11-seed South Florida, 12-seed High Point, and 12-seed Northern Iowa as teams capable of deep tournament runs.
Not everyone is buying what the algorithms are selling. CBS Sports noted that its SportsLine computer model — which is not AI-generated and runs 10,000 simulations per game — projected a double-digit loss for Akron in the first round. It also cooled on Tennessee, projecting the Volunteers to be knocked out in the first round by the play-in winner between SMU and Miami (OH).
Presented with clean data and no further instruction, every A.I. model produces the same boring guesses.
Ultimately, though, it will come down to the very human players on the courts.
And not even A.I. can guess what March Madness will actually deliver.