AI models struggle to analyze sports games, basic recognition accuracy only 74%

A new study from the University of North Carolina at Chapel Hill and Northeastern University finds that major AI models perform poorly at analyzing professional sports. The team created a test called SVI-bench with 35,000 hours of basketball, soccer, and hockey footage to check AI's ability to see, reason, simulate, and act on its own.

On basic tasks like spotting what a player is doing, models like ChatGPT and Google Gemini got about 74% accuracy. But when asked why a play happened, success dropped to around 40%. For example, AI called a weird three-point shot "the first one of the game" instead of noting it hit the backboard top first.

Things got worse on predicting moves—AI was almost guessing randomly. And when asked to think like a sports commentator using game data, accuracy was just 5%. Researcher Lorenzo Torresani says AI can describe what it sees but can't explain why things happen or what comes next, and this gap shows up in other jobs too.