(NOTE: This article is part of an ongoing series documenting an experiment with using AI to fill the NCAA brackets and see how it fares against years of human experience. The original article is as follows.)
After the first two rounds, I am at the top of both my friends’ pool with sixty entries and my family pool with six. For a broader sample size, my brackets stand in the 98th percentile among the 27 million brackets entered into the ESPN challenge.
Were ChatGPT’s picks perfect? No. It didn’t see Florida getting dunked—and that will hurt me later on. But we got 13 out of the Sweet 16 picks correct. Which is almost certainly better than I could have done on my own.
Interestingly, all of my pools are still very competitive. There are no runaways. As I mentioned, I asked Chat GPT to construct a bracket that was optimized to win my pool and another that was designed to more correctly predict the outcome of the tournament. The difference was that the first took a few more chances to separate me from other folks picking favorites. Through the first two rounds, the bracket that was designed to win the pool is doing better—but only slightly.
I will be back next weekend to report on how we did in the Sweet Sixteen and Elite 8.