AI betting tips: can artificial intelligence beat the bookmaker?

What AI betting tips are

The term covers three different things:

1. You ask a chatbot. You type "who will win Arsenal v Tottenham?" and get a tip with reasoning that reads like an expert's analysis.
2. Apps, websites and channels with "AI tips". They show a confidence percentage, bets "picked by the algorithm" and winning streaks. Some are free and live on affiliate links, some sell paid tips or a subscription.
3. Real machine learning models. These are built by bookmakers, odds suppliers and professional betting syndicates. Nobody sells their tips in a messaging app, because they make their money from them in other ways.

This article is about the first two groups. The third matters for a different reason: it sets the odds that AI tips lose to.

What the 2026 tests showed

The 2026 World Cup: four models against the market

Jiacheng Ding, Cong Guo and Jason Xu (WC2026-Agents, July 2026) asked four leading models for a forecast before each of the 104 matches of the 2026 World Cup: Claude Opus 4.8, ChatGPT (GPT-5.5), Gemini 3.1 Pro and Grok. The benchmark was the bookmakers' odds. Accuracy was measured with the Brier score: the lower it is, the better the forecast.

What follows from this:

• No model predicted better than the market. Two came within two thousandths of it, but "almost as good as the odds" is not enough in betting, as we show below.
• In 92% of matches the four models picked the same favourite. Their hit rate (66–68%) barely differed from that of the odds (68.3%). In practice an AI tip is the bookmaker's price told in other words.
• A blind stake on the favourite from the odds made more than any model. That is not a strategy, just the effect of a tournament where favourites often won. But it shows how much "intelligence" the models added to the price itself: none.
• Betting against the market lost money for all four. Where a model disagreed with the odds, it was usually the model that was wrong.
• Three of the four models ended in profit, but 104 matches prove nothing. +10% on about a hundred bets is well within ordinary luck, and backing the favourite blind did better still.

Premier League: a whole season, and bankruptcies

In its KellyBench study (April 2026), the company General Reasoning gave the models more time and more tools. Each got a virtual £100,000, historical statistics, line-ups and closing odds with a margin of about 5.3%, and then went match by match through the 2023/24 Premier League season. It could build its own statistical models, write code and decide its own stakes. Each model played five times.

• Every model lost money on average. The best, GPT-5.4, lost 7.9% of its bankroll, Claude Opus 4.6 11.2%, GLM-5 51.6%, Gemini 3.1 Pro 66.0% and Kimi K2.5 89.6%.
• 3 runs out of 25 ended in profit. GLM-5, Gemini and Kimi had runs that ended in bankruptcy.
• The Dixon-Coles model from 1997 beat three of the five models. It is a simple goals model almost 30 years old, which also lost in this test (−15.4% on 249 bets), but less than most of the chatbots.
• An experienced quantitative analyst made 5.1% on 39 bets. The sample is too small to speak of an edge, but the style of play was different: few bets, and only where he saw a reason.

According to a report on the study, every model could recite the Kelly formula for stake size, but none could apply it consistently. The best scored 26.5% of the points in the assessment of strategy quality, against 73.1% for the human.

Can AI forecast at all?

It can, and it is getting better at it. In general forecasting (politics, the economy, sport) the best models are closing in on the best human forecasters and on prediction markets. Researchers at the University of Chicago who compare them with markets in the Prophet Arena project point, however, to a weakness that matters for betting: models take in new information more slowly than the market just before the outcome is decided. In football that is the last hour before kick-off, when the line-ups come out and, as we show in when do odds stop moving, prices still move. And in the LLM-SoccerArena test at the same World Cup, internet access improved the models' forecasts only slightly.

That is the crux. In betting, matching the market is not a success. It is a loss, only a slower one.

Why AI tips will not beat the odds

The odds are already a forecast, and a very good one

The odds are not one trader's opinion. They are a price moved over many days by bookmakers' models, news on line-ups and injuries and the money of bettors who pay for their mistakes. In our database of 104,000 matches from 37 leagues (2015–2026), favourites given 60–70% by the closing odds won 64.7% of matches, against an average price of 64.4%. There were 25.8% draws, where the market expected 25.6%. On average the market is off by fractions of a percentage point. To improve on it you need to know something it does not, and a chatbot knows less, not more.

Artificial intelligence is already in the odds

Big bookmakers and odds suppliers have used machine learning models and live data from every match for years. Professional syndicates bet on the strength of their own models, and their bets move the prices. A chatbot's tip is therefore not up against "a bookmaker who knows nothing about football". It is up against better models, better data and the money of people who make a living from pricing matches. Everything public (form, Elo, xG, the table, the manager's press conference) is in the price before you ask an AI about it.

A language model predicts words, not goals

A language model learned to predict the next words in a text. That is why it writes convincingly, but:

• it does not know a match that has not been played yet, and without internet access it does not even know the latest rounds,
• with internet access it reads the same news the market priced in hours earlier,
• it can state with total confidence an injury that does not exist, or the result of a match that never took place,
• asked twice, it can give two different tips and write an equally confident case for each.

Eloquence is not an edge. A bookmaker does not pay out for a well-written argument, only for a price that is more accurate than its own.

Matching the market is not enough

Let's do the sums. A bet at odds of 2.00 with a bookmaker whose 1X2 margin is 5% implies a fair chance of about 47.6% (50% divided by 1.05). To break even at odds of 2.00, you need to win 1 / 2.00 = 50% of such bets. To make money on this bet, an AI tip would have to know that a team the market gives 47.6% really has at least a 50% chance. That is a gap of 2.4 percentage points, or 5% more than the market's estimate: the margin, and nothing else.

That bar is low, and that is the honest point: the case against AI tips is not that the bar is high, but that the models do not clear even this one. The best models at the World Cup were not better than the market by even a hair. They were slightly worse. How much the margin costs, we work out in the article on how bookmakers make money, and how to strip it out of the odds, we show in the guide on how to calculate fair odds.

"80% accuracy" is not profit

AI tips apps like to boast about their hit rate. A hit rate without the odds says nothing. A "banker" at odds of 1.30 loses roughly one match in four, and you need to land 76.9% of such bets just to break even (1 / 1.30). At 80% you break even at odds of 1.25 (1 / 0.80) and lose money at anything shorter, so the same "80%" can be a profit or a loss depending on the prices behind it. When favourites win 3 matches out of 4, a run of 8 wins in a row turns up in one in ten sets of eight bets (0.75⁸ ≈ 10%) and proves nothing. How to work out the break-even point for any odds, we explain in the guide on keeping a betting record in a spreadsheet.

What works instead of AI: maths and statistics

To be fair: you can win against bookmakers, but almost never by predicting matches better than the market. The best-known example is the 2017 study by Lisandro Kaunitz, Shenjun Zhong and Javier Kreiner. They built no football model at all. They assumed that the average of many bookmakers' odds reflects the true chances well, and bet only where one bookmaker clearly strayed from it. On 10 years of data the method returned about 3.5% in theory, and over five months of betting with real money it returned 8.5%. Then the bookmakers started limiting their stakes.

That is what real edges look like: a value bet worked out from the market rather than against it, price gaps between bookmakers, a faster reaction to news. It is pure arithmetic, and the reward for it is bookmaker limits, not a fortune. We know this from experience: for years we bet value with our own bots, until one bookmaker cut our limit to €0 a bet.

We also tested our own model. Its team strength ratings predict results as well as good club Elo ratings, and yet none of 89 simple betting rules made a profit in a test against Pinnacle prices. We describe it in why a betting model alone is not enough. If an honest statistical model on 35,000 matches does not beat the market, a chatbot guessing from the sports news has no business there.

What AI is useful for

Not for predicting results, but it is useful for a few things:

• it will explain the rules of a market you do not know, such as the Asian handicap,
• it will summarise the terms of a promotion, though you should still check the numbers in the original,
• it will help you write a spreadsheet formula or a simple script for your betting record,
• it will describe a match from data you give it.

The last point is the honest use of a language model in betting: it writes about the numbers it is given, it does not find an edge in them. Our football tips are picked by statistics, not by a chatbot. And we do not claim they beat the market: we show them to make matches easier to understand.

How to spot an AI tips scam

The tricks tip sellers rely on, with AI or without, we describe in the article on tipster scams. The thinking traps that make them easy to believe are in the piece on betting psychology.

Summary

• At the 2026 World Cup none of the four leading language models predicted results better than the odds.
• In a simulated Premier League season every model lost money on average, from 7.9% to 89.6% of its bankroll.
• The odds are a forecast that already contains bookmakers' models, the news and the money of the best-informed bettors.
• Matching the market is not enough: at odds of 2.00 you need to win 50% of bets, while the market gives such a team 47.6%, and no model closed even that gap.
• Real edges come from maths on prices, not from a chatbot, and they end in limits.
• Treat betting as paid entertainment and set a budget before your first bet. You will find the rules in 10 rules to protect your bankroll.

FAQ

Sources

• J. Ding, C. Guo, J. Xu, FIFA World Cup 2026 as a Contamination-Free Benchmark for LLM Forecasting Agents, arXiv, July 2026: four language models and the odds on 104 World Cup matches.
• T. Grady et al. (General Reasoning), KellyBench: A Benchmark for Long-Horizon Sequential Decision Making, arXiv, April 2026: models betting on the 2023/24 Premier League season.
• J. Schröder et al., LLM-SoccerArena: Benchmarking LLMs on Real-World Predictions in Sports, arXiv, July 2026: seven models at the 2026 World Cup, with and without internet access.
• LLM-as-a-Prophet: Understanding Predictive Intelligence with Prophet Arena, University of Chicago: language models against prediction markets.
• Forecasting Research Institute, LLMs Are Closing the Gap on Human Superforecasters: ForecastBench results.
• L. Kaunitz, S. Zhong, J. Kreiner, Beating the bookies with their own numbers, arXiv, 2017: betting from the average of the odds, and the limits bookmakers then imposed.
• M. Dixon, S. Coles, Modelling Association Football Scores and Inefficiencies in the Football Betting Market, Journal of the Royal Statistical Society C, 1997.
• Decrypt, Can AI Beat the Sports Betting Market? 8 of the Top Models Tried: a report on the KellyBench study.