AI can research team news, analyse statistics and produce football probabilities in seconds. But can it actually find betting value that the market has missed?
The 2026 World Cup provided one of the cleanest tests yet. Researchers tracked leading AI models across all 104 matches, making predictions before kick-off and comparing them with real betting-market prices.
The answer was sobering. The best AI systems were good at identifying likely winners, but the betting market was at least as good. In one study, none of four frontier AI models beat the market on probability accuracy, while simply backing the market favourite made more virtual profit than any of them.
Four Leading AI Models Took on the Betting Market
A study submitted on July 20 tested Claude Opus 4.8, ChatGPT GPT-5.5 Thinking, Gemini 3.1 Pro, and Grok in Expert Mode across every match at the 2026 World Cup.
Each model researched the fixture using web access, produced probabilities for the home win, draw and away win, then decided whether to place a virtual bet and how much to stake.
Researchers also collected pre-match 1X2 odds, allowing the betting market to act as a fifth forecaster.
The four AI systems agreed on the same most likely result in 92% of matches.
That sounds impressive until you look at what they were agreeing with.
Their estimated probabilities tracked the betting market extremely closely, with correlations between 0.97 and 0.99. The researchers concluded that the models were largely recovering information that the market had already priced in rather than finding something the bookmakers had missed.
| Forecaster | Result Accuracy | Brier Score | Betting ROI |
|---|---|---|---|
| Betting market | 68.3% | 0.4688 | N/A |
| Grok | 68.3% | 0.4706 | +10.3% |
| ChatGPT | 68.3% | 0.4729 | +8.0% |
| Claude | 66.4% | 0.4705 | -18.1% |
| Gemini | 65.4% | 0.4828 | +3.7% |
Lower Brier scores indicate better probabilistic forecasts. AI betting returns came from each model’s own staking decisions rather than an identical flat-staking strategy.
Some AI Bets Still Made Money
This is where the result gets more interesting.
Grok returned +10.3% on its virtual bets, ChatGPT made +8.0%, and Gemini returned +3.7%. Claude lost 18.1%.
So saying “AI couldn’t make money betting on the World Cup” would be wrong.
The problem is proving that those profits came from an actual predictive edge.
A simple benchmark that placed a flat $100 virtual bet on the market favourite in every match made $1,041, more absolute profit than any of the AI agents. The tournament was also unusually kind to favourites, which means a 104-match sample can make a particular staking approach look better than it really is.
The study’s authors explicitly warn against treating any individual model’s positive return as evidence of a profitable betting strategy.
That matters. A profitable run and a proven edge are not the same thing.
AI Had the Most Trouble Where Betting Gets Interesting
A separate study published on August 4 followed six frontier AI models throughout the same tournament.
This test was even broader. The models produced predictions across seven markets for all 104 matches, alongside group-winner and tournament forecasts. Researchers ended up with 4,494 scored predictions.
Their average match-result accuracy was 63.9%, roughly level with simply selecting the bookmaker’s favourite.
The models also tended to agree with each other far more often than they were correct. Combining several AI opinions therefore offered little extra help.
More tellingly, accuracy dropped in the closest matches. The systems performed better when one team was clearly stronger, which is precisely where the betting market also has the easiest job assigning a favourite.
Another 2026 World Cup benchmark tested 13 AI systems across the 104 matches. Its best system achieved only small improvements over betting-market and human-fan baselines on match-result and exact-score accuracy, although it performed better on a broader scoreline measure.
Three different tests point towards much the same problem.
Why Is the Betting Market So Hard to Beat?
Modern AI has access to much of the same public information as everyone else: injuries, expected line-ups, rankings, form, statistics, tactical analysis and betting prices.
But bookmakers aren’t setting odds in an information vacuum either.
Prices are shaped by trading models, professional bettors and the collective weight of money entering the market. By the time an AI discovers that one team is stronger, that information will often already be reflected in the odds.
There is another problem. If multiple AI models search the same web, consume similar data and arrive at similar conclusions, asking more models does not necessarily create independent opinions.
In the four-model World Cup test, ChatGPT and Grok selected the same top outcome in every match.
More computing power does not automatically mean more betting value.
Does This Mean AI Is Useless for Betting?
No. It means the bar should be much higher than asking an AI which team will win.
AI can process large datasets, research changing information, test assumptions and help build probability models far faster than doing everything manually. Business Insider also reported during the World Cup on a separate experiment in which several AI models researched matches and traded with virtual bankrolls using live prediction-market prices.
The harder step is converting that information into a probability estimate better than the price already available.
That’s the part bettors should care about.
Picking Spain to beat an outsider at short odds doesn’t prove much. Finding repeated cases where a team’s true chance appears meaningfully higher than the market price, then demonstrating that advantage over a large sample, would be far more interesting.
The 2026 World Cup tests haven’t shown that yet.
For now, AI looks much better at understanding the betting market than beating it.
