Using AI to predict the outcome of World Cup games
First of all, as a man of a certain age, can I just say how much I love the Teletext vibe of this site? It's comparing different AI models versus straight FIFA rankings, versus the perspective of a single human, in terms of predicting the outcome of World Cup 2026 football matches.
At the time of writing, we've started the knockout stage, hence the screenshot at the top. The result of the group stage is here.
What does all this show? I had doubts whether any of this was statistically significant, so I asked my Little Robot Friend:
Not really, no.
- Claude Sonnet 4.6: 47/72 correct outcomes (65%)
- Claude Fable 5: 46/72 (64%)
- Gemini 3.1 Pro: 46/72 (64%)
- Ranking Bot: 45/72 (63%)
- Gemini 3.5 Flash: 42/72 (58%)
- GPT-5.5: 40/72 (56%)
- Human 1.0: 40/72 (56%)
Treating top versus bottom (65% vs 56%) as a simple two-proportion comparison over 72 matches gives z ≈ 1.2, p ≈ 0.23. That's nowhere near conventional significance thresholds. The gaps between the top three or four models are smaller still and essentially indistinguishable from noise.
TL;DR: just look at the FIFA rankings. In all probability, Argentina, Spain, or France will win this year's World Cup. But then we knew that before a ball was kicked.