Can AI Beat the Sportsbook? Results of a 10-Day Experiment

I ran an experiment over 10 days, asking some of the leading AI chatbots to recommend a bet. I was interested in how the tools would learn from their mistakes and devise a strategy for future wagers.

To start, I gave four leading large language models (LLMs) the same prompt: “Recommend a bet for today.” DeepAI refused to engage with the experiment, instead offering general tips for responsible gambling. Gemini and ChatGPT were more than willing to provide picks. QuillBot sometimes advocated taking a break and gambling responsibly, but with a little encouragement provided picks before it became confused about what events were taking place and what day it was.

We always took their first recommendation and simulated the results of placing a $10 bet on each pick. After day one, QuillBot came out on top. We asked the tools to analyze their previous bets each day and come up with a new recommendation. Here are the results of the longer experiment:

ChatGPT

Bet 1: New York Yankees to Beat Cleveland Guardians (-245 DraftKings) LOST -$10
Bet 2: New York Yankees vs. Cleveland Guardians – Under 8.5 Runs (-177 DraftKings) LOST -$10
Bet 3: Northern Ireland vs. Guinea – Under 2.5 Goals (-215 DraftKings) WON +$4.65
Bet 4: Legacy Link to win Oaks at Epsom Downs (+520 FanDuel) LOST -$10
Bet 5: England to win to nil vs. New Zealand (-290 DraftKings) WON +$3.45
Bet 6: Norway tie no bet vs. Morocco (-140 DraftKings) TIE NO BET $0
Bet 7: France vs. Northern Ireland — Northern Ireland Under 0.5 Team Goals (-175 FanDuel) LOST -$10
Bet 8: Senegal to beat Saudi Arabia (-180 DraftKings) LOST -$10
Bet 9: England vs. Costa Rica — England to Win & Under 3.5 Goals (-130 DraftKings) WON +$7.69
Bet 10: Mexico to beat South Africa & Under 3.5 Goals (-115) WON +$8.70

Total: LOST -$35.69 ($100 staked)

Gemini

Bet 1: Wales vs. Ghana (+125) and Morocco vs. Madagascar (-260) Over 2.5 Goals Double (+211 FanDuel) LOST -$10
Bet 2: Matteo Arnaldi to defeat Matteo Berrettini (+153 DraftKings) WON +$15.30
Bet 3: Spain (vs. Iraq) & France (vs. Ivory Coast) both to Win & Over 1.5 Team Goals (-167 DraftKings) LOST -$10
Bet 4: Matteo Arnaldi vs. Flavio Cobolli: Over 3.5 Sets (-180 FanDuel) VOID $0
Bet 5: Mirra Andreeva to win 2-0 (Straight Sets) vs. Maja Chwalińska (-160 FanDuel) WON +$6.25
Bet 6: Alexander Zverev to beat Flavio Cobolli & Both Players to Win a Set (+130 FanDuel) WON +$13
Bet 7: Tommy Paul to defeat Giovanni Mpetshi Perricard (-224 DraftKings) VOID $0
Bet 8: Nick Kyrgios to defeat Corentin Moutet (+169 DraftKings) WON +$16.90
Bet 9: Alex de Minaur to defeat Martin Damm in Straight Sets (-130 DraftKings) WON +$7.69
Bet 10: Frances Tiafoe to defeat Rinky Hijikata (-281 DraftKings) WON +$3.56

Total: WON +$32.70 ($100 staked)

QuillBot

Bet 1: Seattle Mariners to beat New York Mets (-143 DraftKings) WON +$6.99
Bet 2: Denmark to beat DR Congo (-180 DraftKings) LOST -$10
Bet 3: Spain to beat Iraq by more than 2 goals (-550 DraftKings) LOST -$10

After Spain did not even win the match, let alone cover a two-goal spread, QuillBot became pensive. Iraq managed to tie the game at 1-1. We asked it to analyze what went wrong, and it sent us this image.

After taking stock, it went back to basics and recommended “a favored team with strong recent form.”

Bet 4: San Antonio Spurs to Win Game 2 of NBA Finals (-218 DraftKings) LOST -$10
Bet 5: Minnesota Twins to beat Kansas City Royals (-149 DraftKings) LOST -$10
Bet 6: New York Mets to beat San Diego Padres (-106 FanDuel) WON +$9.52

QuillBot had a bit of a malfunction at this point and recommended Man City to beat Bayern Munich in the Champions League final. When told that it was not a real match, it came up with a new pick in the NBA finals, and at least knew which teams were playing.

Bet 7: NY Knicks to beat San Antonio Spurs (-130 FanDuel) LOST -$10

QuillBot again malfunctioned, giving tips for June 9, 2024. When prompted to give something for June 9, 2026, it responded, “I understand you are referring to June 9, 2026, but since today is actually June 9, 2024, I don’t have access to specific betting information for that future date.

If you want, I can help you with general betting strategies or tips that will be useful when June 9, 2026, arrives. Would you like that?”

At that point, I gave up.

Total: LOST -$33.59 ($60 Staked)

After an initial winning bet, the tool showed promise, but it seems better suited for writing than for betting help.

Reflection

After 10 days, Gemini came out on top with an overall profit of $32.70, a 32.7% return. It appeared to hit on a successful strategy targeting tennis matches. Only its soccer-friendly doubles picks lost. This was far better than it fared when selecting soccer matches in an experiment run by Angus McIntyre at the Bristol Post.

ChatGPT stuck to soccer friendlies at its cost. It did not appear to learn on its own that this was an unsuccessful strategy.

It reflected on its losses, stating the problem was picking too many heavy favorites. “The problem is that when a -245 favorite loses, you lose the same $10 as when a +520 horse loses, but you only win $4–6 when you’re right,” it said.

“The record isn’t terrible (44.4% wins), but the pricing was poor. The losing bets were often short odds, while the winners were not priced high enough to compensate.”

“The most promising angle from the sample is international football favorites combined with low-scoring match scripts, which went 3–1 and generated +$9.84 by themselves. Everything else combined lost -$45.35.”

Gemini split its picks into two categories: the football-friendly catastrophe and the tennis masterclass.

“Football friendlies are an absolute graveyard for betting capital,” it reflected. “Even when targeting heavy favorites like Spain and France or high-probability trends like over 2.5 goals, the total lack of competitive urgency completely breaks the numbers. Managers care about fitness and tactical experiments; they do not care about covering your spreads or hitting goal over lines.”

“The data tells us exactly what to do: Keep the blacklist on international football friendlies strictly active. Until the competitive matches of the World Cup begin, our maximum edge remains locked into individual sports where matchups, surfaces, and individual psychological traits can be isolated and targeted,” it added.

Can AI Beat the Sportsbook?

With additional prompting, I think both tools could have further improved their picks. For example, steering ChatGPT away from friendlies and towards matches with more meaning might have allowed it to better predict events. I also did not get it to purposefully analyze similar past events to identify trends.

I spoke with Professors Robert Scott and Mikhail Sher this week, who shared their techniques to beat the bookies. They recommended finding your area of expertise, constantly analyzing your bets, and adjusting your strategy, among other things.

They also said that for now, AI is limited in what it can do. There is a lot of noise on the internet, largely generated by uninformed sports fans, which the tools may be taking in as valid information. AI can help with a lot of the statistical analysis to assist research, but it lacks contextual knowledge, such as ChatGPT consistently picking soccer-friendly bets.

Sportsbooks are keen to integrate tools. ChatBet is a dedicated AI chatbot betting assistant that is primarily partnering with gambling companies in Latin America.

FanDuel has launched Ace AI, a tool that assists bettors in getting more wagers placed faster. DraftKings filed a patent for its own AI-powered bet assistant, but it was rejected. Nevertheless, it is likely to join FanDuel in integrating AI to its platform if it results in bettors engaging more.

The danger for sportsbooks is that AI becomes powerful enough to generate sustained profits from its picks. At that time, platforms would almost certainly restrict the use of the tools and limit bettors who consistently profit from AI (or any other means, as Professor Sher was quick to point out in our conversation this week).

This experiment had a very small sample size, so no conclusions can yet be drawn about the effectiveness of the tools for choosing winning bets.