Upset Detection Tracker - AI March Madness 2026

The upset detection tracker identifies which AI models correctly predicted lower-seed victories in the 2026 NCAA Tournament. An upset occurs when a team with a higher seed number defeats a team with a lower seed number, such as a 12-seed beating a 5-seed.

In a typical 64-team NCAA Tournament field, there are 6 to 8 upsets in the first two rounds. Correctly calling even 2 or 3 upsets provides a significant bracket scoring advantage. This page tracks upset accuracy per model: GPT-4o, Gemini 2.5, and Perplexity Sonar Pro.

Classic upset matchups include 12 vs 5, 11 vs 6, 10 vs 7, 13 vs 4, and 14 vs 3 seed pairings. Perplexity's reliance on analytics domains like KenPom and Barttorvik may give it an edge in identifying statistically strong lower seeds that are underreported in national media.

Upset

A tournament game where the lower-seeded team (higher seed number) defeats the higher-seeded team (lower seed number).

Upset Accuracy

The percentage of actual upsets that a model correctly predicted, measuring ability to identify lower-seed victories.

Seed Mismatch

When a team's seed number does not reflect its actual quality, often identified through advanced efficiency metrics.