AI brackets outperformed humans
AI brackets are actually scoring — models like Gemini and Claude hit roughly 96% accuracy in a test (130 points with 3 misses), outpacing ChatGPT and Grok in the same exercise. The experiment is turning heads about AI's predictive strengths in tournament modeling and bracket construction ( ).
Several independent reporters and outlets ran head‑to‑head AI bracket tests this March, including Axios’ Ina Fried and a WBUR segment that retested chatbots after problems the prior year. (axios.com) Sporting News fed Google’s Gemini the full 2026 men’s tournament bracket and published round‑by‑round picks and upset calls for every region. (sportingnews.com) Sportsbook Review said it “trained” Anthropic’s Claude to produce game‑by‑game picks plus score projections and published a detailed writeup and upset recommendations on March 20, 2026. (sportsbookreview.com) Covers published a 10,000‑simulation tournament model that used KenPom and advanced efficiency metrics to project Final Four probabilities and pathway odds. (covers.com) Bracket scoring is not standardized: common systems increase point values by round rather than awarding one point per correct game, which changes how raw totals (and leaderboard placement) are calculated. (printyourbrackets.com) Testing methods varied across reports — some teams prompted models directly, others incorporated external data or custom training — so published comparisons reflect different inputs and not a single uniform benchmark. (sportingnews.com) News coverage emphasized that last year’s failure modes (models misreading matchups or missing later‑round pairings) were corrected this season through clearer prompts and data‑first workflows, a shift noted by multiple outlets. (axios.com)
