Five concrete limits: training data that is incomplete or out of date, the absence of real-time context, the black box effect that makes any reasoning impossible to assess, the overconfidence produced by a figure that looks precise, and the risk of hallucination from a general-purpose language model used outside its remit.
Five limits, and the matching check
None of these limits is corrected by switching model. They come from the data available, from the way a model returns its output, and from the way a figure is read.
This page does not explain what AI does: it lists what trips it up. For the overview — capabilities, uses, vocabulary — see AI and sports predictions. For the question of principle, whether a model can predict a result, see can AI predict a match?.
| Limit | Mechanism | Possible check |
|---|---|---|
| Incomplete or out-of-date data | The model learns on a history that stops at a given date | Ask which period the data covers |
| No real-time context | Line-ups, weather, a last-minute incident are not included | Check the time at which the analysis was produced |
| Black box effect | The output is visible, the reasoning is not | Require the variables used, or otherwise factor that in |
| Overconfidence in a figure | The precision of the display gets confused with the precision of the estimate | Turn the figure back into a frequency: 62% means being wrong 4 times out of 10 |
| Hallucination from a general-purpose model | A plausible text is produced without up-to-date data | Check that the facts cited are sourced and dated |
Limit 1 — incomplete or out-of-date data
A model only learns on what it was given, up to the date it stopped being given any.
The systematic review by Galekwa and co-authors, published as a preprint in 2024, identifies data quality and the ability of models to generalise as two of the main obstacles across the whole field. The problem is not theoretical: a change of coach, a major transfer window or a rule change alters how a team behaves without the history reflecting it.
What is missing most often: secondary competitions, recent data on less well covered teams, and the qualitative variables nobody has taken the trouble to encode.
Limit 2 — the absence of real-time context
The most decisive information often arrives in the hour before kickoff. The official line-up, a last-minute withdrawal, playing conditions.
A model trained and run beforehand does not include them. An analysis produced the day before can therefore be technically correct and already obsolete — something that, on a market, shows up immediately in the odds.
That is in fact the best available test: if the market has moved sharply since the analysis was produced, the analysis is missing something. Reading such a movement is covered on odds movements before a prediction.
Limit 3 — the black box effect
Many models return an output without the path that led to it. You observe the figure, never the reasoning.
The practical consequence is underestimated: with no visible reasoning, it becomes impossible to assess how robust an analysis is other than by its results — that is, by the least reliable criterion there is on an uncertain event.
A model can be right for the wrong reasons, and go on being right until the day the correlation it was relying on stops holding. Without access to the reasoning, that day cannot be anticipated.
Responsible gambling — understanding the limits of a model does not reduce the risk. Sports betting carries a risk of financial loss and presents, according to the Autorité nationale des jeux (the French gambling regulator), the highest individual risk of problem gambling among regulated activities. Set a budget, do not chase your losses and use the limit-setting or self-exclusion tools available. Learn more about responsible gambling.
Limit 4 — overconfidence in a precise figure
A figure displayed to the decimal point suggests a precision the estimate does not have.
"61.7%" and "around 60%" say the same thing, but they do not produce the same effect. The first gives the impression of a calculation mastered down to the last digit; in reality, the uncertainty of the estimate far exceeds the gap between the two wordings.
The useful reflex is to translate any percentage back into a frequency. An estimate at 62% means: being wrong roughly four times out of ten. Put that way, it invites markedly less confidence. It is the same figure a bookmaker's price implies once converted into an implied probability — a model's percentage is not automatically more reliable than the market's.
Limit 5 — hallucination from a general-purpose model
A general-purpose language model produces plausible text, which is not an analysis.
Asked about a match, with no access to up-to-date data, it can confidently state an incorrect line-up, an invented head-to-head record or a statistic that does not exist. Nothing in the form of the answer signals the error: the text flows just as smoothly in both cases.
Two simple checks: verify that the facts cited are dated and attributable to a source, and be wary of any analysis that mentions no limits at all.
Why this page acts as a safeguard for the rest of the cluster
Everything above applies to a human analysis too, with one difference: a figure produced by a machine enjoys a presumption of objectivity it has not earned.
That is why the accuracy benchmarks coming out of research have to be read with care. They describe academic work on predicting results, under given conditions, with specific sports and competitions — not a success rate that transfers to a bet.
What these limits let you conclude — and what they do not
Three levels never to confuse.
- What they let you do — question a tool about its data, its freshness and its exact output, rather than about the results it displays.
- What they let you estimate — the reasonable degree of confidence to place in a given figure, best checked against the market's own estimate, the fair odds once the bookmaker's margin is removed.
- What they never let you conclude — neither that AI is useless, nor that a well-built model ends up predicting a match. Both conclusions are wrong for the same reason: the uncertainty is in the sport, not in the tool.
Why estimating is not predicting →
Back to the AI and sports predictions guide →
Dig into the market
Odds movements are only part of the story. Here are the next topics to read.
Can AI predict a match?
Why estimating a probability is not predicting a result.
Understand the nuanceThe most common mistakes in predictions
Confirmation bias, chasing losses, overconfidence: what distorts an analysis most.
Avoid these mistakesIs a reliable prediction possible?
Why no prediction is guaranteed, and how to measure how solid it is.
Understand reliabilitySports predictions: analyse before you bet
The complete cluster guide: data, odds, reliability, bias and the role of AI.
Read the guideFrequently asked questions
Why can an AI model be wrong without anyone understanding why?
Because many models do not return the reasoning that leads to their output. You observe the result, not the path — which makes it impossible to assess how robust the reasoning is rather than whether an outcome happened to be right.
Is a precise figure such as "62%" more reliable than a rough estimate?
Not necessarily. The precision of the display says nothing about the precision of the estimate. A figure given to the decimal point can rest on partial data.
How do you limit the biases of a prediction generated by AI?
By checking how fresh and how complete the data is, by requiring a probability rather than a named outcome, and by comparing the model's output with the odds market.
Can a general-purpose language model analyse a match?
It can produce a plausible text, which is not the same thing as an analysis. Without access to up-to-date data, it can state inaccurate information with confidence.
Does OddScore provide tips?
No. OddScore compares the odds of several bookmakers, removes the margin and tracks how they evolve. The platform publishes no bet selections and provides no staking advice.
Sources & methodology
This page draws on an academic review devoted to predicting team-sport results with machine learning, on a recent systematic review of the use of machine learning in sports betting, and on the French regulatory framework published by the Autorité nationale des jeux.
- Describe each limit by its mechanism, not by a general warning.
- Separate the limits that come from the data from those that come from the model or from how it is read.
- Present work that has not yet been peer-reviewed as preprints.
- Pair each limit with a concrete check the reader can carry out.
- Bunker, R. & Susnjak, T. (2022). The Application of Machine Learning Techniques for Predicting Match Results in Team Sport: A Review. Journal of Artificial Intelligence Research, 73, 1285-1322.
- Galekwa, R. M., Tshimula, J. M., Tajeuna, E. G. & Kyandoghere, K. (2024). A Systematic Review of Machine Learning in Sports Betting: Techniques, Challenges, and Future Directions. Preprint arXiv 2410.21484 (not peer-reviewed).