Build-in-public · AI Picks

AI vs Human: Who Tips Better at WC 2026?

Before WC26 we ask the honest question: does an AI model beat the average tipper? What we learned from Euro 2024.

Leo Brunnhofer·May 25, 2026·7 min
Leo BrunnhoferFounder · built BetTillDone 2016–2018 (119 players, 76% activation)May 25, 2026Updated: XGitHub

The honest answer: it depends

If you tip 200 matches of a tournament, an AI model beats the average tipper by about 4–7 points (in Kicktipp Classic points). That's not "100% hit rate" — that's more consistent across the whole season.

The human tipper has hot streaks and cold streaks. The model tips the average. Over 200 matches, the average wins.

What we measured at Euro 2024

We ran Euro 2024 in a test round with 47 players and Claude-3.5-Sonnet as an additional "tipper".

  • AI Picks: 142 points, rank 14 of 48
  • Median tipper: 128 points
  • Top tipper: 187 points (a human who nailed 3 surprises exactly)

The AI almost never hit anything exactly right — and almost never hit anything exactly wrong. It tips "Spain 2-1 Italy" when everyone else tips "Spain 2-0". It collects fewer zero-point rounds.

What the AI does well

  1. 1.Consistency: never tired, never drunk, never angry about the lineup
  2. 2.Market integration: reads Polymarket/Pinnacle odds live
  3. 3.Form weighting: considers the last 5 games, not just "gut feel"
  4. 4.Expected Goals (xG): knows underlying stats, not just the table

What the AI does poorly

  1. 1.Local factors: "Salzburg just thrashed Frankfurt 4-0 in the cup semi, everyone is in top form" — only partially captured
  2. 2.Last-24h injury news: if Müller drops out 2 hours before kickoff, AI doesn't know (unless we wire in live news)
  3. 3.Surprise coaching changes: Aguirre takes over Mexico — AI needs 5 matches to grasp the new style
  4. 4.Emotion: "Morocco at a WC in North Africa's reach" — that's a component statistics don't capture

How tiptilldone uses the AI

In Premium (€19 for all of WC26) you get for every match:

  • AI Pick with reasoning (why exactly 2-1?)
  • Market consensus as sanity check (what is Polymarket saying?)
  • xG-based probability for each outcome
  • Confidence score of the AI (sometimes it says itself: "I'm uncertain")

We don't believe the AI does your work for you. It's an additional argument in your group's tipping discourse. You still have to decide whether to follow the AI or tip against it.

The honest recommendation

If you've never seriously tipped — follow the AI for the first 10 matches. You'll be top 25%.

If you're a tipping veteran — use the AI as a counter-argument. Sometimes the AI is right and you have gut feel. The matches where you disagree are where you'll learn most.

Tip live with AI Picks at tiptilldone.com.

Share

WC26 Tipping Tricks

Tipping tricks top tippers use — by email

Home-field bias, contrarian timing, group vs knockout psychology — via email, free, no spam.

1 email/week max · unsubscribe anytime · no spam, promised