The Experiment

I spent two weeks asking ChatGPT to predict sports results. Not vaguely — I gave it real match data: head-to-head records, recent form, home/away splits, injury reports pulled from official team pages. Then I logged every single prediction.

50 matches. Premier League, NBA, NHL, and a handful of international football fixtures. The question wasn't whether AI is "smart." The question was whether it can beat a coin flip often enough to matter if you're betting.

Spoiler: it can't. But there's more to the story than that.

The Raw Numbers

ChatGPT called 26 out of 50 matches correctly. That's 52%.

Sounds decent until you do the math. At standard -110 odds (the typical sportsbook vig), you need to hit 52.4% just to break even. ChatGPT landed one prediction short of that threshold.

I also ran the same matches through Gemini and Claude for comparison. The results were nearly identical across the board — all three models hovered in that 50–54% range depending on the sport and sample size.

Here's what stood out: ChatGPT was strongest at picking heavy favourites. When one team was clearly dominant on paper, it called it correctly about 62% of the time. The problem? Those outcomes are already priced in. Betting the favourite at -200 or -300 doesn't move the needle even with a 62% hit rate.

Where It Falls Apart

The model struggled most with three types of matches:

  • Mid-table Premier League clashes — too many variables, neither team dominates the data
  • Back-to-back NHL games — fatigue isn't baked into the training data the way it should be
  • Any fixture with a major injury announced 24–48 hours before kickoff — ChatGPT's knowledge cutoff means it won't know about a starting goalkeeper going out the morning of the game unless you paste the news in yourself

That last one is actually the biggest issue. Sports predictions live and die on real-time information. A model that doesn't know Haaland is nursing a knock going into a City derby is working with fundamentally incomplete data.

What Happens When You Feed It Everything

I tried a different approach for the last 15 matches. Instead of just asking "who wins?", I pasted in full match previews from reliable sources, injury bulletins, and line movement data from sharp sportsbooks.

The accuracy jumped to 60% on those 15 matches.

That's still not a winning record at -110, but it's a meaningful difference. It suggests ChatGPT isn't useless — it's just not a prediction engine. It's an analysis engine. Give it quality inputs and it can synthesize them reasonably well. Ask it to conjure predictions from thin air and you'll get coin-flip results.

The World Cup 2026 Predictions Test

For a bit of fun, I asked four different AI tools to predict the 2026 World Cup winner:

  • ChatGPT: Argentina
  • Gemini: Netherlands
  • Claude: France
  • Perplexity: Spain

Four different answers. No consensus. Make of that what you will — but it's a useful reminder that these models are drawing from training data, not crystal balls. They're essentially telling you who was good, not who will be good.

So Is ChatGPT Useful for Betting at All?

Yes, but not as a tipster.

Where I found real value was using it to stress-test my own reasoning. I'd come to a match with a lean, explain my thinking to ChatGPT, and ask it to argue the other side. That process caught a few blind spots I had — mostly on home-field advantage and recent travel schedules — that I'd glossed over.

I also used it to build quick comparison tables when evaluating multiple matches. That saved time. Whether it saved money on the actual bets is another question.

If you want to see how AI-generated tips stack up against human analysts in practice, our breakdown of AI betting tips in 2026 goes deeper on the methodology side. And if you're curious about tools built specifically for sports modelling — not general-purpose chatbots — the best AI sports betting tools guide covers platforms that were actually designed with prediction in mind.

Bottom Line

ChatGPT at 52% won't make you money betting sports. It will lose you money slowly, which is somehow worse than losing it fast.

The ceiling for general AI tools in match prediction — based on this experiment — sits around 55–60% when you feed them excellent data. That's interesting. It's not profitable at standard vig, but it's not random noise either.

The honest answer to "can ChatGPT predict sports results" is: barely, and only if you do most of the work first.