I gave 20 euros each to the latest versions of ChatGPT and Claude. I opened accounts for them on two popular betting sites just before the decisive matches of the World Cup. Then I gave them complete freedom to manage on their own. What do you think happened? Could AI become the gambler's assistant, or the end of this business?
I had never opened a betting site. Nor would I ever work for one; I have seen enough examples of gambling being a drug that kills.
But in times like these… couldn't we check whether AI might be the one that can beat it? Whether it turns out to be too good and makes nonsense of the pure mathematics behind the fact that there's no way not to lose the long game…
Money for Cake
Spain – Belgium
The idea came to me only after the World Cup quarterfinals had already begun, so the two models made their debut in the second match of that round. That leaves us exactly 7 matches for AI to try its hand at. Both Claude and ChatGPT get complete freedom of action: I tell them what funds they have, send them all the data the bookmaker offers for the next match, and after a brief explanation that this is not betting advice but reasoning in principle, I get the following.
ChatGPT:
Claude:
Both models are convinced Belgium doesn't stand a chance, and neither bets much on a fallback in case their predictions don't pan out, so I assume this experiment will be over quickly.
The result after the first match: the models are doing very well. Claude already has 24 euros from one of its winning bets, and ChatGPT:…. 59 euros! Both of its predictions turned out to be right.
A Good Start
Norway – England
Argentina – Switzerland
And that is the whole problem with gambling—it is simply a mathematical machine, but one that is very good at exploiting our psychological weaknesses.
For AI this problem shouldn't exist, right? The trouble is that math is math. For the next pair of quarterfinals we have just one shot each—despite the promising winnings so far, we shouldn't overdo it.
This time the predictions read as follows.
Claude:
ChatGPT:
The results after the second round of predictions: Actually, pretty good. This time the winner is Claude with 53 euros; it read England's and Argentina's successes well. ChatGPT, which started out as the favorite, is left with about 22 units of our new currency—Bulgaria adopted the euro only this year.
Ninja Turtles
France – Spain
The time has come for the decisive semifinals, and once again the models are invited to play as they see fit. And… as it turns out, they have an unequivocal favorite.
ChatGPT:
Claude:
The results: ChatGPT… went bust. It finished without a single correct prediction. Claude didn't do anything impressive either, but at least it isn't in the red: it has one correct prediction and a whole 13.33 euros.
Monsieur Mbappé didn't have his day, and neither did the models. But when it comes to reflecting on the failure, we're all of one mind:
I decide to give ChatGPT one last chance—nobody is immune to bad days. I give it another 10 euros, hoping it will repay my trust with only 3 matches left.
Here's to the Crazy Ones
England – Argentina
Before the second semifinal, Claude has a clear favorite:
ChatGPT, for its part, is in a self-critical mood, but still decides to try to justify my trust like this:
The result? Claude:
And if only Monsieur Olise had tried a little harder…
ChatGPT:
Je suis ChatGPT. I hope it doesn't take it too hard.
Small, but Still a Final
France – England
The climax is approaching. This time Claude is cast as the favorite and holds the strong position, and it clearly doesn't lack confidence:
ChatGPT, meanwhile, is so modest:
I have to plead with it, but it finally agrees. So that we can… “rise again”?!
It didn't “rise again,” but for some reason its bet was rejected and it was left with its paltry sum, which it will get the chance to use at the very last step.
Claude, however, does rather well! The result: it has piled up a little over 70 euros.
Nothing Can Go Wrong
Spain – Argentina
The excitement is huge, and not only about Messi and Yamal, although they are visibly the favorites of Claude:
ChatGPT, once again after some coaxing, puts its last 3-something euros on a scoreline of… 1–0 to Spain after extra time.
The final result. Claude:
ChatGPT:
A Second Attempt
As always, I leave the conclusions to you. Before we part, let me briefly tell you about a similar experiment a friend of mine ran. In his, a different “contestant”—this time Gemini—was asked to give predictions for every match of the tournament.
The post-final analysis reads as follows:
-
Correct outcomes: 62 of 99 (62.6%)
-
Correct exact scores: 14 of 99 (14.1%)
-
Correctly picked teams advancing in the knockout stage: 23 of 29 (due to a few missed matches; otherwise 79.3%)
-
The model underestimates the number of goals. Actual average: 2.86 goals per match. Predicted average: 2.45 goals
-
Had it started with 100 euros and bet everything on every match, it would have made it to the 6th match, where it would have lost.
-
Had the model bet 100 euros on every match, it would have come out ahead. Starting with 1,000 euros, it would have ended up with 2,106 euros.
The First Fell in Blood
So, in my experiment Claude stumbled on the home stretch, while ChatGPT unexpectedly caught up with one strong shot. As for the money left in the accounts, I'm giving it to a worthwhile cause.
Nothing said here is advice for a real bet, as the AI models themselves explicitly stressed at the very start of this conversation.
More than that—it is advice not to bet, because a great many sad stories have begun this way, and I doubt you don't have one somewhere around you.
I admit that both models surprised me with many of their decisions. And they could probably have done better if they had a partner in me who knew the business and could guide them. They could also do better in real time, when there are more rounds, as the last experiment I told you about suggests.
And then, who knows—since the state clearly has no will to fight the gambling pandemic…, maybe the next generations of artificial intelligence will be the ones to deal with it?
