Update: AI Won My March Madness Bracket

UPDATE: Amazingly, my AI picks won the official Micro Center office March Madness pool, with UCONN's championship win securing my victory. For this feat of futuristic sports betting, I used a combination of ChatGPT, Google's Gemini, and NVIDIA's ChatRTX. All were useful in some way, but most of the actual picks came from asking Gemini for the winner of individual hypothetical matchups -- although ChatRTX gave a solid list of first round picks to get me started. 

 

But this year, I had another idea. With access to an ever-growing array of AI tools, shouldn't technology be able to help me make smarter picks? 

For this experiment, I asked OpenAI's ChatGPT, Google's Gemini, and NVIDIA's ChatRTX for advice. The first two chatbots are cloud-based, and able to do independent research online, while ChatRTX is a local AI, downloaded to your computer and accessing only local documents you allow. 

ChatGPT

For ChatGPT, I started by asking for the full monty: "Can you fill out my March Madness bracket with the most likely winning answers?"

ChatGPT replied: "To fill out your March Madness bracket with the most likely winning predictions, you can consider some insights gathered from various expert predictions. For instance, Purdue, Houston, and UConn have been highlighted as strong contenders for the top seeds this year. Purdue has an impressive record and is expected to be a significant player in the tournament, especially with notable wins throughout the season. Houston has shown robust performance in the Big 12, and UConn, being the reigning national champions, also presents a formidable lineup."

 

Screenshot: Dan Ackerman 

Good general advice, but the system seemed averse to providing specific picks. I decided to try a smaller, more manageable question: "For the first round, who should I pick in each March Madness matchup?"

The reply read, in part: "For your March Madness first round picks, it's essential to consider expert analyses and predictions across various matchups...Remember, while expert predictions can guide your decisions, the unpredictable nature of March Madness is what makes it exciting, so there's always room for personal intuition and gut feelings in your picks."

That made it feel like ChatGPT was being overly cautious and trying to make sure I wasn't betting the house on its advice. Which I appreciate, but it wasn't getting me any closer to a filled-in bracket. 

I drilled down even more, asking about a specific game that didn't seem like a blowout. "Who will likely win their first round game, BYU or Duquesne?"

Things got a little more specific here, but felt like ChatGPT was really trying to avoid being nailed down to an answer. The chatbot said: "BYU has had notable performances throughout their season, which could factor into their strength going into the tournament. Duquesne, on the other hand, showcased resilience and skill, especially highlighted by their win against Dayton, as mentioned in their season overview."

After referring me to DraftKings and other sports sites for more information, it closed with: "For the most informed decision, consider combining these data points with your analysis, factoring in team dynamics, key player performances, and overall season success. Always remember that upsets are a hallmark of March Madness, making it both unpredictable and thrilling."

Maybe asking for gambling advice is a sensitive topic, and ChatGPT knows to deflect specific questions like this. So I moved on to the next contender, Gemini. 

Google Gemini 

For Google's recently renamed AI, I asked a similar opening question: "I'm participating in an office March Madness pool for 2024. Give me a complete bracket suggestion list based on the most likely outcomes."

Gemini started with a nicely formatted, if general, reply: "While I can't guarantee the most likely outcomes (upsets are a big part of March Madness!), I can provide a suggested bracket based on rankings, recent performance, and expert predictions. Here's a breakdown. 

• First & Second Rounds: Generally, pick the higher seed to advance; Look for potential upsets based on recent performance and matchups that favor a lower seed. Resources like CBS Sports and Sporting News can help identify these.
• Sweet Sixteen & Elite Eight: Here, the competition gets tougher. Consider past performance in the tournament and recent wins against ranked opponents.
• Final Four & Championship: Select the top remaining seeds with a strong chance of winning. Balance this with potential Cinderella stories that have momentum."

Then Gemini gave me some actual specific advice on the four regions, although not the full list of picks I was asking for. That was some actionable info, at last. But still not granular enough for me. So even though it would take longer, I decided to go all-in and ask about each individual matchup. I picked a non-obvious contest and asked, "Who will win the first-round game between Mississippi State and Michigan State?"

This was apparently the magic formula, and I started getting highly specific answers. Gemini gave me a decent breakdown and concluded: "This is a close matchup, but Michigan State's edge in defense and Tom Izzo's tournament experience might give them a slight advantage."

I tried again with another matchup -- Clemson vs. New Mexico -- to make sure it wasn't a fluke, and got another solid answer: "Most analysts favor New Mexico with a slight edge, with some predicting a close game. While upsets are a hallmark of March Madness, New Mexico seems to have the upper hand based on recent form and overall strength." 

Now that I had a decent formula for at least getting some specific advice, I moved onto a local AI bot, NVIDIA's ChatRTX. 

ChatRTX

This is an AI system (formerly known as Chat with RTX) that you install locally on your computer, and it requires at least a 30-series or 40-series NVIDIA GPU to run. I'm running the Mistral LLM on it, but it also supports Meta's Llama 2. 

The key to the hype around local AI, including ChatRTX, is that you can personalize it by giving the AI access to your own content -- basically by putting all your data in a specific folder and telling ChatRTX to use that data. It can read text and DOC/DOCX files, PDFs, and XML. 

Screenshot: Dan Ackerman

From that, I got a pretty solid, if not surprising, list of first-round picks, although there was some disagreement between Gemini and ChatRTX, with Gemini picking New Mexico over Clemson and ChatRTX picking Clemson. The ChatRTX answers were straightforward lists, with no editorializing or warnings about Cinderella stories. 

Presenting my AI March Madness bracket

In the end, I decided to fill in my bracket relying most heavily on Gemini, which included online predictions and the general internet zeitgeist in addition to pure stats. From that initial list of winners, I presented Gemini with the hypothetical second and third-round games, and so on, and ended up with a championship game between UConn and Purdue. Gemini was reluctant to pick a potential winner, but when pressed, said: "If I had to pick one, I would lean slightly towards UConn. But, it would likely be a close game."

For the tie-breaker total score of that final game, Gemini said: "Simulating the game using a statistical model, one possible outcome could be a total of around 166 points. This considers the average points scored by teams in the NCAA and adds some variation."

Both Gemini and ChatGPT picked UConn as the most likely overall champ, but I couldn't get ChatRTX to commit to a winner. When asked to restart from scratch and pick the two teams most likely to meet in the finals, ChatGPT picked the same UConn vs Purdue final matchup, but Gemini suggested Houston vs. Tennessee as a strong possibility. Depending on the exact phrasing, previous conversation, and even when it was asked, it was possible to get different answers from these AI bots. 

You can see my entire bracket below, or at ESPN. As AI answers are largely based around research and stats, it pretty much follows a lot of the conventional wisdom, with only a handful of even minor upsets. That might mean that these AI picks are playing it overly safe, or that they're just playing the odds. We'll return to these predictions as the tournament progresses to see how our AI picks held up.