And the results were pretty surprising.\u00a0ChatGPT<\/a> hit on upsets and the collapse of favorites, and I kept climbing the leaderboard in my pool. By the time the Final Four rolled around, I was suddenly in striking distance of winning the whole thing.<\/p>\n Ultimately, I didn’t win. But I came close enough that it didn’t feel like the original joke experiment I intended, and it instead felt like something worth testing again.<\/p>\n So yeah, this year, I ran it back.<\/p>\n When the 2026 NCAA tournament bracket dropped on Selection Sunday, I once again turned to ChatGPT to build my March Madness bracket from scratch. I wasn’t suddenly going to become a college basketball sicko overnight, and that’s kinda the point of the experiment.\u00a0<\/p>\n Could AI help someone who mostly watches the NBA put together a bracket that felt informed, stayed alive through the early rounds and maybe even held up deep into the tournament?<\/p>\n Before we get to how it all turned out, here’s how ChatGPT made its picks.<\/p>\n What AI gets right (and wrong) about March Madness<\/p>\n One thing AI quickly understands about March Madness is that predicting it perfectly is basically impossible. Even if you knew the exact probability of every game outcome, the chances of picking a perfect bracket are astronomically small, 1 in 120.2 billion<\/a>, and that’s if you know ball. You’re far more likely to be attacked by a shark.<\/p>\n That randomness of this tournament is exactly why it’s so fun and why office pools exist in the first place.<\/p>\n What ChatGPT tends to get right is structure. It understands the historical patterns of the tournament. Higher seeds usually advance, certain midseeds are common upset candidates, and picking too many underdogs early is usually a mistake.<\/p>\n Where AI struggles is the same place most humans do: the unpredictable chaos that defines March Madness. Hot shooting nights, injuries, coaching adjustments and momentum swings all play a part.<\/p>\n But that’s also why the experiment is interesting to me. If AI can build a bracket that survives the early rounds and avoids obvious mistakes, it might do better than the average bracket filled with random guesses in most pools.<\/p>\n An OpenAI spokesperson reached out to me about my story and offered this statement: “ChatGPT can be a fun and useful tool for anyone filling out their brackets this year. Whether you follow the sport closely or just want a little help making your picks, it’s free to use and can break down team stats, compare matchups, and help you think through different approaches depending on how you want to play it.”<\/p>\n How I asked ChatGPT to pick my bracket<\/p>\n My experiment lat year worked best when I didn’t simply ask ChatGPT: “Who will win March Madness?” Instead, I gave the AI a structured prompt that tried to mimic how analysts evaluate similar tournament matchups. The goal isn’t perfection but instead building a bracket that balances favorites with a few realistic upsets. The kind of bracket that might actually win a pool.<\/p>\n The prompt I decided on for this year:<\/p>\n “You are helping me fill out a March Madness bracket. Use historical tournament trends, team seeding and general basketball analytics to suggest winners for each matchup. Avoid unrealistic brackets with too many early upsets, but include a few plausible ones, including possible dark-horse and Cinderella teams. The goal is to build a bracket that could realistically win a pool. Use this as a guide: https:\/\/www.ncaa.com\/march-madness-live\/bracket<\/a>“<\/p>\n Note:\u00a0I used ChatGPT 5.4 Thinking.<\/p>\n You can adjust the prompt depending on how risky you want the bracket to be. If you’re in a large pool with thousands of people, it makes sense to tell ChatGPT to be more aggressive with upset picks in order to stand out. In a smaller pool, a more conservative strategy, leaning toward higher seeds and avoiding early-round chaos, often gives you a better chance of staying competitive.<\/p>\n <\/a> <\/p>\n