March Madness Meets the Machines: Can AI Crack the Code?
Last year, I tasked ChatGPT with the near-impossible: predicting the madness of March Madness. The outcome? My bracket, like most others, was shattered, though I’d like to think my picks had slightly more substance than pure luck.
A year later, the landscape has drastically shifted. To try and gain an edge on the picks, I’ve assembled an AI “Dream Team,” of sorts: Google’s Gemini Researcher, OpenAI’s GPT-4.5 Research Agent, xAI’s Grok-3 with Reasoning and Research, and Baidu’s freshly released ERNIE X1. ERNIE X1 is, supposedly, capable of surpassing DeepSeek R1 and all of OpenAI’s reasoning models. I’ve created a bracket for each one. I dusted off my own GPT-based bot, now upgraded with GPT 4.0, and updated it with this past season’s data.
Maybe one of these AI models will perform better compared to last year, but the odds are stacked against it. They say the odds of correctly filling out a bracket is 1 in 9.2 quintillion, which leaves little to no room for error.
These digital minds, however, have crunched countless terabytes of data. They have analyzed every KenPom rating and scraped an internet’s worth of stats to produce their best predictions. Will one of these AI models become the basketball prophet we’ve been waiting for? You be the judge. Their replies are available on our GitHub repository, but here’s a summary of what each model predicted.
ChatGPT’s SEC Celebration
If ChatGPT were a college basketball devotee, it would wear the Southeastern Conference logo like a badge of honor. Its Final Four prediction featured three SEC teams: Alabama (East), Tennessee (Midwest), and Auburn (South), along with St. John’s (West).
ChatGPT went all-in on an all-SEC championship game, predicting Auburn to defeat Alabama. Among its riskier choices, it has Gonzaga—an 8-seed—knocking off 1-seed Houston in the second round.
On the upset front, ChatGPT was relatively conservative, picking New Mexico over Marquette, Colorado State (12-seed) over a 5-seed, and UC San Diego as an unexpected “bracket buster.” The model showed extreme favor to Auburn’s championship potential, citing its balanced excellence as the key factor: “Auburn was ranked No. 1 overall for a reason and has that defensive edge in a high-scoring battle. In a title game, when legs are tired, defense and half-court execution matter a ton– and Auburn gets the nod there.” Its championship prediction—Auburn 86, Alabama 80—would make for an electric all-SEC final that showcased what ChatGPT refers to as “balanced excellence.” It also noted that Auburn ranks in the top 10–15 in both offensive and defensive metrics—”a proven formula for championship teams.”
- Final Four: Alabama vs. Tennessee (all-SEC showdown) and Auburn vs. St. John’s
- National Championship: Auburn defeats Alabama
(Note: ChatGPT highlighted that Duke was highlighted by some early predictive models as having the highest initial odds, but the research Agent specifically selected Auburn as the champion.)
Gemini’s Blue Devil Prophecy
Google’s Gemini foresees a Final Four of Duke (East), Florida (West), Houston (Midwest), and Michigan State (South), with Duke beating Florida for the championship.
Unlike ChatGPT, Gemini embraced upsets in the first round, predicting multiple 13-seeds to triumph over 4-seeds: Yale over Texas A&M, Akron over Arizona, and High Point over Purdue. It also foresees Drake (11) over Illinois (6) and VCU (11) over BYU (6).
Gemini’s analysis strongly emphasized Duke’s balance (top 5 in both offensive and defensive efficiency) and Florida’s #1 offensive efficiency rating. Most notably, Gemini highlighted injury concerns for key players like Duke’s Cooper Flagg and Houston’s J’Wan Roberts—details the other models didn’t include. Gemini’s region-by-region analysis revealed careful attention to statistics. For its championship prediction, it focused on Duke’s good performance on both ends of the floor: “This analysis was based on KenPom ratings, NET rankings, recent form, injury reports, and expert predictions from various sources.” Convincing enough for an amateur.
- Final Four Candidates: Duke (1-seed, East Region), Florida (1-seed, West Region), Houston (1-seed, Midwest Region), Michigan State (2-seed, South Region)
- Championship Prediction: Duke beats Florida
Grok-3’s Balanced Approach
Elon Musk’s Grok-3 takes a middle-ground approach with a Final Four of Auburn (South), Duke (East), Florida (West), and Houston (Midwest), and ultimately picks Duke to win against Florida for the championship.
Grok-3’s upset picks include Yale over Texas A&M and the winner of a North Carolina/San Diego State play-in game over Marquette. It also listed other potential upsets like Liberty over Oregon, Akron over Arizona, Colorado State over Memphis, McNeese over Clemson, and High Point over Purdue.
Like Gemini, Grok-3 underscored the importance of KenPom metrics and other data sources, particularly noting Auburn’s strong offense (KenPom #2) and solid defense (#12), along with the SEC’s impressive showing of 14 teams in the tournament.
For Duke’s run through the East, Grok-3 specifically mapped out the following: “Elite Eight: Duke over Alabama; Sweet 16: Duke over Arizona, Alabama over Wisconsin; Second Round: Duke over Mississippi State, Arizona over Oregon, Wisconsin over Baylor, Alabama over UCLA.”
What makes Grok-3’s prediction interesting is its balance between conservative bracket construction and strategic upset picks. It focuses on teams with statistical strengths and favorable matchups rather than pure shock value.
- Final Four: Duke beats Auburn, Florida beats Houston.
- Championship: Duke over Florida, predicted national champion.
ERNIE X1’s Bold Gambles
Baidu’s ERNIE X1 model predicts a Final Four of Duke (East), Florida (West), Alabama (South), and Kansas (Midwest), with Duke defeating Alabama for the title.
However, ERNIE shines with its audacious upset picks. How about Norfolk State giving Iowa State a run for its money, or New Mexico winning? If those happen, ERNIE might need to start picking lottery numbers as they are the clear underdogs according to the experts.
In its analysis, ERNIE X1 emphasizes coaching experience, noting: “Importance of coaching experience (Jon Scheyer – Duke, Todd Golden – Florida)” as critical to tournament success. The model also offers a unique perspective on team composition, highlighting that teams with “elite guards (Florida, Duke) or dominant big men (Alabama, Purdue)” tend to have advantages in tournament play.
ERNIE X1’s Sweet 16 and Elite 8 predictions follow a logical progression from its first-round upsets. Duke is predicted to defeat Purdue and Kentucky upsets Tennessee in the East before Duke advances to the Final Four. In the South, it predicted Alabama defeating Baylor, and Gonzaga outlasting Michigan State before Alabama advances.
- Final Four: Duke vs Florida Alabama vs Kansas
- National Championship: Duke defeats Alabama: Duke’s depth and shooting prove too much for Alabama’s defense.
The Bracket Bot Gets a Second Chance
Last year, the March Madness Bracket Bot returned with one thing in mind: redemption. The bot failed mostly because BYU—a favorite by consensus— got stomped by Duquesne, a team that hadn’t won a March Madness since 1969. This year, the Bot’s Final Four picks are Duke (East), Houston (Midwest), Florida (West), and Auburn (South), ultimately picking Duke to defeat Auburn for the championship.
The Bot’s upset picks center on 12-5 and 11-6 matchups: Drake (12) over Baylor (5), Richmond (12) over Kansas State (5), Vermont (13) over Arizona (4), Texas (11) over Indiana (6), and San Diego State (11) over Kentucky (6). In an incredible display of evolution, the Bot provided a uniquely detailed, round-by-round breakdown of how things might unfold, rather than giving us its final predictions. It even laid out a “Cinderella Path” for Drake to reach the Elite Eight before ultimately falling to Florida. It also identified San Diego State (11) as a “Dark Horse Final Four Candidate” based on the Aztecs’ elite defense.
- Most reliable Final Four: Duke (1), Houston (1), Florida (1), Auburn (1) → These #1 seeds are elite.
- National Championship: Duke (1) → Best balance of offense, defense, and coaching.
The AI Consensus: Duke Dominance?
Across all five AI models, certain patterns emerge:
Duke is the clear favorite, picked by four of the five models to win it all. Only ChatGPT goes against the grain, selecting Auburn instead.
Florida makes the Final Four in four predictions, although only as the runner-up in two of them.
Auburn reaches at least the Elite Eight in all five brackets, with two models (ChatGPT and Bracket Bot) picking them as championship finalists.
The SEC dominance is unanimous, with multiple conference teams predicted to make deep runs across all models.
Major upset candidates that appear in multiple predictions include Yale over Texas A&M, Drake over higher seeds, and San Diego State making a surprising run.
Houston’s path is highly contested, with predictions varying from a second-round exit (ChatGPT) to a Final Four appearance (Gemini, Grok, Bracket Bot).
ChatGPT is the only model predicting an SEC champion despite all models acknowledging the conference’s strength.
Each AI model brings something unique to the table. Gemini focused on injury analysis – which makes sense. Grok-3 underscores SEC dominance, with a record 14 tournament teams. ERNIE X1 offers the boldest upset picks by far—not a dumb idea considering how last year went. The Bracket Buster Bot provided accurate results back in 2024 and has been specifically configured by a human for this task.
Why AI Still Can’t Solve March Madness
Despite their sophisticated algorithms and extensive data processing, these AI models still can’t agree on a perfect bracket. Why? Because March Madness thrives on the unpredictable. We’re talking about 10 humans with free will, interacting in a court with thousands of fans with free will, under the weather and other conditions that are unpredictable, with coaches and technicians working on unpredictable strategies with unpredictable moods. Determinism isn’t a word that would fit this type of tournament.
As advanced as these models are, they are still making educated guesses based on past performance. All models are based on theoretical—even scientific—methods, but none can fully capture the human element that makes college basketball so maddeningly unpredictable.
That’s what makes the tournament so great. If Duke goes on to win the championship as the AI consensus would have us believe, we’ll have five silicon prophets to thank. Even more likely, when the final buzzer sounds, we’ll all be left with the same question we ask every April: “Who saw THAT coming?”
