What is football AI? Four families of tools, and what each one cannot do
"Football AI" (or soccer AI) covers four distinct technologies: video analysis, scouting, statistical prediction and database question answering. None of the four replaces the other three. Footlab AI occupies the last two: answering a factual question from 1,561,628 recorded matches, and publishing 1X2 probabilities whose accuracy, measured on every card served, is 54.7 %, on par with the bookmakers.
Accuracy of the served cards and comparison with bookmakers measured on 5 October 2026; the 1X2 base rates (44.3% home wins) measured on 10 August 2026 and checked again on 2 October 2026, unchanged. Coverage volumes are recomputed daily from that same database.
What are the four kinds of football AI?
Four families, separated by what they PRODUCE: an annotated video, a shortlist of players, a probability, or a numeric answer. A video analysis tool cannot produce the output of the other three families, and vice versa.
| Family | Who it is for | Example tool | What this family does not do | Footlab AI |
|---|---|---|---|---|
| Video analysis and tracking | Professional clubs, video analysts | Second Spectrum, PlayVista, TacticAI (DeepMind × Liverpool FC) | Answers no factual question and computes no match probability | No |
| Scouting and recruitment | Sporting directors, agents | SciSports, Comparisonator | Runs on proprietary data sold to clubs, out of reach for the general public | No |
| Statistical prediction | General public | Forebet, NerdyTips | Answers no historical question; almost always tied to odds and affiliate links | Yes; market odds taken into account, never displayed |
| Database question answering | General public, journalists, supporters | StatMuse | Predicts no match; coverage concentrated on a handful of leagues | Yes |
Tools named as examples, observed on 9 August 2026. Footlab AI has no commercial relationship with any of them and earns no commission.
Can an AI answer a factual football question reliably?
Yes, on one condition: the AI must READ a database instead of generating text. A language model used on its own produces a plausible sentence, not a verified figure. Nothing in how it works lets it check whether the goal count it just wrote ever existed.
Footlab AI translates the question into queries on its database, runs them, then answers with the figure and its source. The consequence is checkable by the reader: every answer says where its figure comes from, the Footlab AI database or, when it is not enough, an external search with its links, which is not true of an answer written by a language model alone.
The limit is just as clear: what is not in the database cannot be answered. Detailed match sheets start at the 2008 season; before that, Footlab AI holds honours lists and historical facts, not per-player, per-match statistics.
How accurate is football AI, really?
54.7% of 1X2 outcomes correctly called on the 1,522 Footlab AI cards actually served (measured 5 October 2026), and 55.8% for the bookmakers' favourite on the same matches: Footlab AI is on par with the market, not above it. The right starting point is not 33.3%, the three-way coin toss, but 44.3%: the share of home wins measured over 29,415 matches in the top 20 domestic leagues, seasons 2022-2025.
| Strategy | 1X2 accuracy | What the figure proves |
|---|---|---|
| Three-way coin toss | 33.3% | Nothing: no football information is used at all |
| Call “home win” on every match | 44.3% | Home advantage, known forever and free to reproduce |
| Best constant strategy, league by league | 44.3% | The home win is the most frequent outcome in all 20 leagues tested, without exception |
| Footlab AI cards actually served | 54.7% | That is 9.8 points above “always home” on the same matches (44.9%): this is what the card actually learns about the teams |
| Bookmakers' favourite, same matches | 55.8% | The real benchmark: Footlab AI is at that level, not above it |
| Model alone, European cups (test of 16 July 2026) | 56.7% | Level gaps are wider there, therefore easier to call |
1X2 accuracy = share of fixtures where the outcome shown as most likely actually happened. The served card combines the model with the consensus of published odds; the temporal hold-out, for its part, judges the model alone on matches played after its calibration window, never on the games used to tune it.
What data does a football AI need?
A match history deep enough for regularities to be measurable, and coverage wide enough that the question asked does not fall outside the perimeter. Here is the exact inventory of the Footlab AI database, recomputed daily.
| What is in the database | Volume | Scope |
|---|---|---|
| Recorded matches | 1,561,628 | 2008 → 2027 |
| Players | 493,978 | 171 countries |
| Competitions | 1,247 | 787 leagues, 460 cups |
| Statistical rows | 24,957,225 | events, match sheets, seasons, standings |
Two sources, and only two: API-Football under licence for fixtures, squads and statistics, Wikidata for entity disambiguation and older historical facts. No data is collected by scraping third-party sites.
Football AI or a general chatbot: what is the difference?
A general chatbot answers from memory; a football AI backed by a database answers by reading. The gap shows on questions with a precise figure, a precise date, or an unusual cross-cut: exactly the ones where a language model alone produces a credible, wrong value.
The decisive test fits in one question: “where does that number come from?”. If the tool cannot show the source of the figure it just wrote, the figure is not verifiable.
Frequently asked questions
- Is “football AI” a single technology?
- No. Four distinct families coexist: video analysis and tracking, scouting and recruitment, statistical prediction, and database question answering. They use neither the same data nor the same methods, and none produces the output of the others.
- Can an AI invent a football statistic?
- A language model used on its own can, and regularly does, because it generates the most plausible sequence of words rather than consulting a source. An AI backed by a database guards against it, as long as it reads the figure from the database and states its source next to the answer.
- What is the best accuracy achievable on a football result?
- No public model beats the bookmakers' favourite for long, which calls the right outcome in 52 to 56% of matches depending on the competitions considered. The free benchmark is 44.3%, obtained by always calling the home win, measured over 29,415 matches in the top 20 domestic leagues between 2022 and 2025.
- Does Footlab AI do video analysis?
- No. Footlab AI processes no images and produces no tracking data. The product covers two families out of four: database question answering and statistical prediction.
- Is Footlab AI a sports betting site?
- No. Footlab AI publishes no odds, shows no link to a betting operator and earns no affiliate commission.
Ask Footlab AI a question
A football question, a numeric answer, and the source of that figure.
Open Footlab AISee how accuracy is measured