How accurate are AI football predictions?

Verifiable systems typically get 50–55% of three-way (home/draw/away) results right; anything advertising far more is usually measuring something easier. Football is low-scoring and three-outcome, which makes it one of the hardest major sports to call: a single deflection or red card routinely flips a correct analysis into a wrong result.

The reference points are public. Across major European leagues, the home team has historically won about 45% of matches — so "always pick home" already scores around 45% while knowing nothing. The market favourite wins a little over half the time. And the strongest entries in public academic soccer-prediction challenges have landed just above 50% on three-way results. That is the honest playing field: the real contest happens in a band of a few percentage points, not between 50% and 90%.

An accuracy percentage means nothing until you know what zero effort would have scored on the same matches.

Two other things quietly inflate headline numbers. First, the market: a two-outcome pick like "double chance" or over/under is far easier to get right than a three-way result, so an 80% figure on those markets is not comparable to a 52% figure on 1X2. Second, the sample: a hot fortnight is not a record. Any figure quoted without its market, window, and match count is a slogan wearing a percent sign.

Five checks separate a record from an ad — apply them to any tool, including ScoreGPT:

1. Can you see every pick, or only the winners? A real record shows losses in the same place it shows wins.
2. Is the market named? Three-way results, double chance, over/under, and "team X or draw" are different difficulty classes.
3. Is the window defined? "87% accurate" over an unstated period is unfalsifiable. Look for dated, continuous history.
4. Were the picks published before kickoff? Post-hoc selections are not predictions.
5. Is the sample big enough to mean anything? A handful of matches proves nothing either way — in any direction.

If you cannot see the losses, you are not looking at a record — you are looking at an ad.

Honest answer: the leaderboard rotates, and nobody can credibly crown a permanent winner. In ScoreGPT's own live grading of five frontier models — GPT-5.6 (OpenAI), Claude Opus 4.8 (Anthropic), Grok 4.5 (xAI), GLM-5.2 (Z.ai), and Kimi K3 (Moonshot) — the weekly leader changes, and the differences between models sit within the noise you would expect at these sample sizes. Anyone who tells you one AI is the best at football predictions is describing a snapshot, not a fact.

The same applies with more force to any single competition, including the Premier League: per-league samples are a fraction of the total, so single-league rankings are mostly noise. The useful question is not "which AI won last week?" but "which tool shows you the running score at all?" The live independent leaderboard exists so you can watch the comparison instead of trusting a claim.

ScoreGPT does not quote a sustained accuracy figure in its marketing — the live record is the claim. Every prediction from every model is graded in public after full time, wins and losses alike, and the accuracy leaderboard shows each model's recent record over rolling windows, so a hot streak and a cold streak are equally visible.

For a completed, dated case study, the World Cup 2026 AI report card preserves the day-by-day graded record of an entire tournament — every prediction published before kickoff and graded after full time. That is the standard this page argues every prediction tool should meet, and the reason ScoreGPT can afford to be honest about the modest numbers this game actually allows: it is the only kind of number that survives being checked.