Short answer: ChatGPT can give you a sensible-sounding view on a match, but on its own it is not a reliable forecaster. It may not know the latest results or injuries, it can state wrong facts confidently, and its percentages are not calibrated. A statistical model fed with checked data does better, and AI is most useful helping that model, not replacing it.
What a general AI chatbot gets right
Large language models such as ChatGPT, Claude or Gemini have read a great deal about football. Asked about a match, they usually reach for the right ingredients: home advantage, recent form, injuries, head-to-head record and the gap in quality between the squads. For big clubs in big leagues, their general sense of who is stronger is often reasonable.
They are also good at reading and summarising. Given a set of sourced articles, a model can pull out which players are injured, who is suspended and what the manager said, much faster than a person.
Where it falls short
- Out-of-date knowledge. A model’s built-in knowledge stops at its training date. Without live data or web search it does not know last weekend’s results, this week’s injuries or a change of manager.
- Confident mistakes. Chatbots can invent details, such as a player who left the club or a suspension that does not exist, and present them as fact.
- Numbers that are not probabilities. When a chatbot says “60% home win”, that figure is not produced by a model fitted to past results. Ask again and you may get a different number.
- No track record. A chat answer is not locked or graded, so there is no way to tell whether its forecasts are any good over hundreds of matches.
How we test AI forecasts in public
We use two AI tools, Claude by Anthropic and Codex by OpenAI, on matches in selected leagues. Each reads a dossier of sourced team news and gives its own probabilities, which we lock before kickoff and grade exactly like our main forecast. In version 1 their forecasts run in shadow: they do not change the published numbers.
On the 11 graded matches where all three have a forecast, the statistical model has an average ranked probability score of 0.1916; Claude’s own forecasts score 0.1926 and Codex’s 0.1936 (lower is better). That is far too few matches to draw a conclusion. The models page shows which of the rules for letting the AI layer count have been met so far.
What to use instead
If you want a forecast you can check, look for three things: probabilities from a model fitted to past results, a public record of every forecast including the misses, and a clear statement of what data the forecast used and when. Our methodology explains how we do each of those, and our comparison of statistical and AI football predictions goes into the trade-offs.
Frequently asked
There is no good public evidence that a general chatbot beats bookmaker prices over a season. Bookmaker odds combine models, money and fast information, and they are hard to beat even for specialist models.
It can help summarise news you give it, but check every claim against a dated source such as the club’s website. That is the rule our own editors follow before anything is published.
