Can AI Predict Football and Poker? What the Data Says

Every couple of months somebody claims an AI has cracked football predictions, or poker. The real research looks a lot less exciting. Machines are useful in both games now, but they fail in different places, and knowing where keeps you from trusting a number just because it looks confident.

Football: plenty of data, plenty of luck

Football is very hard to predict because goals are rare and draws are common. One deflection or one red card can flip a result the numbers had right. That’s why even the best models never get close to certainty.

The clearest public test is the 2017 Soccer Prediction Challenge. The organizers gave teams over 216,000 historical games across 52 leagues from 35 countries and asked them to predict 206 future games. The winning model, gradient-boosted trees built on team ratings, called the right result 52.4% of the time. A 2024 review of the field adds something more telling. On the challenge’s own probability score, forecasts taken straight from bookmaker odds beat every top entry, even though the winning model called slightly more results right (52.4% vs. 51.9%). The market takes account of team news, injuries and sentiment in a way that a model only trained on past scores won’t. (Later models built on the same dataset have since beaten the bookmakers’ score, though not by much.)

That doesn’t mean AI is useless in football. In 2024 Google DeepMind and Liverpool FC released TacticAI based on a dataset of 7,176 Premier League corners. Liverpool experts compared set pieces suggested by the AI with real ones (without knowing which were which), and the AI set pieces were preferred 90% of the time. But the task was to help design a set piece, not to predict the winner.

Chatbots as tipsters

People are also asking general chatbots for predictions and the results are grim here. ForecastBench, a benchmark from the Forecasting Research Institute, puts the same questions about future events to models and to people. In the original 2024 results, expert forecasters scored 0.093 on the Brier scale (lower is better). The best model scored 0.111, and the general public, at 0.107, did about as well as it did. Models have been closing that gap fast since then.

There’s a practical problem too: a chatbot only knows what it was trained on or what it manages to look up. And it’ll be equally confident in describing a striker’s form whether the data is from last week or from last season.

Poker: fixed rules, hidden cards

Poker looks like the easier case for a machine. The rules don’t change. There’s no weather, there’s no ref. But you can’t see the other players’ cards so you have to use probabilities and your bluffing rate to decide the best move to make.

Purpose-built programs cracked that years ago. Carnegie Mellon’s Libratus beat top professionals at heads-up no-limit hold’em in 2017. Pluribus, built with Facebook AI, did the same at a six-handed table in 2019. Both taught themselves by playing against themselves, and neither could read any text at all.

Language models are another story. A 2025 UC Berkeley / Georgia Tech benchmark, PokerBench, found that the best off-the-shelf model, GPT-4, got the best move and bet size right only a little more than half the time across 11,000 poker situations. And in a public play-money cash game with nine chatbots in fall 2025, Meta’s Llama 4 lost its whole bankroll. Plus, one analysis of the hand logs showed some models misstating their own hands.

Where both go wrong: sample size

The most common mistake in football and in poker isn’t in the model. It’s judging the model too early.

In a 2026 bot comparison on Zenodo commercial poker bots played against each other without human players. One machine-learning bot led the main sample of 17,295 hands until the final week. Then the other one overtook it and finished at +47.1 big blinds per 100 hands (bb/100) against +13.2. Looks like a clear winner but in reality the gap was smaller than the margin of error. And 62% of the winner’s profit came from just ten hands. The paper concludes that a difference of 30 bb/100 between such bots over 15,000 hands cannot be distinguished from luck. Halving the margin of error takes four times as many hands. As a group, though, the two machine-learning bots clearly beat the older rule-based ones.

In football the problem is even worse: it’s not thousands of hands, but 38 league games a season for a Premier League team. If a tipster is correct 30 times in 50, they’re either good at it or they just got lucky. PokerBotAI’s guide to variance and sample size walks through this with poker examples, and the logic carries straight over: short runs tell you very little.

What this means for you

  • Treat a model’s output as a probability, not a pick. A model that gets 52% of football results right is a good one.
  • Compare it with the market. A model that can’t beat bookmaker prices adds little.
  • Don’t take a chatbot’s facts on trust. Check team news yourself.
  • Judge any system, human or machine, over hundreds of predictions, not one good weekend.

AI is getting better at both. Right now, though, the best results still come from narrow tools designed to do one thing, used by people who know how much randomness is still in the numbers.