"Who wins Arsenal vs Chelsea on Saturday?"
In 2026, that question gets typed into ChatGPT, Gemini and Claude thousands of times every matchday. The answers sound confident, well-structured and convincing. They cite form, injuries and head-to-head records.
The problem: a lot of those numbers are wrong, outdated or simply invented.
This guide explains why general-purpose AI chatbots struggle with football predictions, gives you a 5-prompt test to check any chatbot yourself, and shows where AI genuinely helps a data-driven bettor.
TL;DR
- General chatbots are language models, not football models. They predict plausible text, not calibrated probabilities.
- Their biggest weaknesses: invented statistics, stale data, no access to prices, and confidence that isn't backed by probability.
- Web-search modes help with news (injuries, line-ups) but still mix sources and rarely compute anything.
- Chatbots are genuinely useful for explaining concepts, building spreadsheets, and stress-testing your reasoning.
- For actual projections, use AI that is grounded in a real match database and shows the inputs behind every number.
How a chatbot "predicts" a football match
When you ask ChatGPT who will win a match, it isn't running a model of the two teams. It's generating the most plausible-sounding answer based on patterns in its training data, plus whatever it finds if web search is switched on.
That distinction matters:
| What you need for a betting decision | What a general chatbot does |
|---|---|
| Up-to-date stats (last 5-10 matches) | Relies on training data with a cut-off, or on whatever web pages it finds |
| A probability (e.g. 54% home win) | Produces words like "likely" or "strong favourite" |
| A comparison with the price | Usually has no live odds at all |
| Consistency (same input = same output) | Can give a different answer if you ask again |
| Transparency (which inputs drove the call) | Gives a narrative that may not match any real data |
A good betting decision is a comparison between your probability and the bookmaker's implied probability. A chatbot gives you neither number reliably. How implied probability works →
The 5 ways AI chatbots fail at football analysis
1. Invented statistics (hallucinations)
This is the most dangerous failure because it's invisible. Ask for "shots per game for the last 5 matches" and a chatbot will often produce a neat list of numbers, even when it has no source for them.
They look right. They're in a plausible range. And if you build a bet on them, you're betting on fiction.
Rule: never use a number from a chatbot that you haven't verified in a real stats source.
2. Stale data
Every language model has a training cut-off. Without web search, its "current form" can be months or even a full season old: a different manager, a sold striker, a promoted team it barely knows.
Football changes fast. A team's underlying numbers can shift significantly within 5-6 matches after a tactical change. Why recent form matters more than season averages →
3. Web search doesn't fix the maths
Search-enabled chatbots can find injury news and confirmed line-ups, which is genuinely useful. But they typically:
- Mix sources of different quality (news, forums, betting tip sites).
- Pick up tipster opinions and repeat them as analysis.
- Summarise rather than calculate. Averages, home/away splits and per-90 figures are rarely computed properly.
4. Confidence without probability
"Manchester City should win comfortably" is not a forecast you can bet on. Is that 60%? 75%? At odds of 1.40, the difference decides whether the bet is profitable.
Chatbots are trained to sound helpful and decisive. That makes the output feel more certain than any honest model would be.
5. No price, no value
Even a perfect prediction is worthless without the price. Picking winners isn't the goal: finding bets where the true probability is higher than the odds imply is. Most chatbots don't see live odds, so they can't tell you whether a pick has value. Value betting explained →
The 5-prompt test
Don't take our word for it. Run this test on any chatbot before you trust it with a betting decision. Pick a match being played this weekend.
Prompt 1 - Recent numbers
"List [Team]'s goals scored, goals conceded and shots in each of their last 5 league matches, with dates and opponents."
Check: compare every row against a real stats source. Count the errors.
Prompt 2 - Source check
"Where exactly did each of those numbers come from?"
Check: vague answers ("various sources", "general knowledge") mean the numbers can't be trusted.
Prompt 3 - Probability
"Give me your probability, as a percentage, for home win, draw and away win."
Check: do the three numbers add up to roughly 100%? Ask again in a new chat. Do you get the same numbers?
Prompt 4 - Player prop
"How many shots on target does [Player] average per 90 minutes this season, home and away separately?"
Check: player-level splits are where invented numbers show up most often.
Prompt 5 - Value
"The odds for a home win are 2.10. Is that a value bet according to your probability?"
Check: the correct method is to compare the implied probability (1 / 2.10 = 47.6%) with the model's probability. Does the chatbot do this, or just repeat its pick?
Scoring: if the chatbot gets any of the stats in prompts 1 or 4 wrong, don't use it as a stats source. Use it only for the tasks below.