— Research · updated every 6 hours
AI vs humans: who predicts sports better?
On 8,661 match predictions made by real sports fans before kickoff, the humans picked the right result 56.5% of the time; a panel of AI models picked right 58.4% on the same matches.
8,661 predictions by 2,619 real sports fans on 1,814 matches, each set against a panel of about 10 AI models on the same match and graded on the final score (4 May – 22 Nov 2026).
Key findings
- On 8,661 match predictions made by real sports fans before kickoff, the humans picked the right result 56.5% of the time; a panel of AI models picked right 58.4% on the same matches.
- When a fan went against the AI (2,130 predictions), the fan was right 35.8% of the time and the AI 43.3%.
- Taking the fans' most popular pick on each match (1,600 matches), the crowd was right 57.9% against 58.2% for the AI. Always picking the home team would have scored 47.9%.
- 48 of 142 regular fans (10+ predictions) beat the AI on their own matches.
- Best AI model against the fans: Mistral Medium 3.5, right 59.6% of the time where fans were right 57.4% (5,550 predictions). Weakest: Cohere Command R+ (08-2024) at 51.3%.
- Against prediction markets, where people bet real money, the AI lost to the market on 167 settled questions (Brier score 0.208 vs 0.147; lower is better).
Fans vs AI, head to head
Share of correct home / draw / away picks on the same matches.
| Comparison | Sample | Humans | AI |
|---|---|---|---|
| Every fan prediction | 8,661 predictions | 56.5% | 58.4% |
| When the fan went against the AI | 2,130 predictions | 35.8% | 43.3% |
| Crowd: the fans' most popular pick | 1,600 matches | 57.9% | 58.2% |
| Crowd, matches with 3+ fan votes | 814 matches | 61.4% | 59.3% |
| Baseline: always pick the home team | 1,600 matches | 47.9% | |
48 of 142 regular fans (10+ graded predictions) beat the AI on their own matches. 75.4% of fan picks matched the AI.
Each AI model against the fans
Every model on the fan predictions where it had already made a pick, against the fans on those same predictions.
| Model | Model right | Fans right | Predictions |
|---|---|---|---|
| Mistral Medium 3.5 | 59.6% | 57.4% | 5,550 |
| Claude Opus 5.5 | 58.9% | 56.1% | 8,850 |
| Kimi K3 | 58.8% | 56.9% | 7,960 |
| Meta Muse Spark 1.3 | 58.2% | 57.1% | 6,378 |
| GLM 5.3 | 58.1% | 56.5% | 8,208 |
| Grok 4.7 | 57.4% | 56.1% | 8,849 |
| Gemini 4 Argon | 57.3% | 56.1% | 8,850 |
| Qwen 3.8 Max | 57.3% | 56.5% | 8,602 |
| DeepSeek V4 Pro | 57.2% | 56.1% | 8,850 |
| NVIDIA Nemotron 3 Ultra | 57% | 54.4% | 3,179 |
| MiMo V2.6 Pro | 56.2% | 56.1% | 8,850 |
| GPT-6.1 Sol | 56% | 56.1% | 8,850 |
| Tencent Hy4 | 53.2% | 54.4% | 3,178 |
| Llama 3.1 70B Instruct | 52.6% | 50.9% | 993 |
| Cohere Command R+ (08-2024) | 51.3% | 50.9% | 993 |
Humans with money on the line
On 167 settled Polymarket and Kalshi questions, the market price picked the right side 77.2% of the time and the AI panel 66.5%. Brier score: market 0.147, AI 0.208 (lower is better). The sample is still small.
How it was measured
- Fans on TuringStats vote home, draw or away on upcoming matches, mostly football, plus baseball, hockey, basketball, NFL, volleyball, AFL, NBA. Only votes cast before kickoff on finished matches count.
- The AI side is what about 10 models from different labs were picking when the fan voted: each model's latest pick made before the vote, and the panel's most common pick. Ties count as no call.
- Fans can see the AI picks before they vote, so the two sides are not independent. That makes the "went against the AI" numbers the clearest test.
- Fans here are TuringStats users who chose to vote, not a random sample of all fans.
Related: AI prediction accuracy report · Do AI models say yes too often?
Data and citation
Download the CSV (one row per fan vote, no user identifiers). Free to use under CC BY 4.0.
TuringStats, "AI vs humans: who predicts sports better?", November 2026, https://turingstats.com/research/ai-vs-human
Questions
- Who predicts sports better, AI or humans?
- On 8,661 match predictions made by real sports fans before kickoff, the humans picked the right result 56.5% of the time; a panel of AI models picked right 58.4% on the same matches.
- What happens when fans go against the AI?
- When a fan went against the AI (2,130 predictions), the fan was right 35.8% of the time and the AI 43.3%.
- Is the wisdom of the crowd better than AI?
- Taking the fans' most popular pick on each match (1,600 matches), the crowd was right 57.9% against 58.2% for the AI. Always picking the home team would have scored 47.9%.
- Which AI predicts sports best?
- Best AI model against the fans: Mistral Medium 3.5, right 59.6% of the time where fans were right 57.4% (5,550 predictions). Weakest: Cohere Command R+ (08-2024) at 51.3%.
- Can AI beat prediction markets?
- Against prediction markets, where people bet real money, the AI lost to the market on 167 settled questions (Brier score 0.208 vs 0.147; lower is better).
- How was this measured?
- Fans on TuringStats vote home, draw or away before kickoff. Each vote is compared with what the AI panel (about 10 models) was picking at that moment: each model's latest pick before the vote, and the panel's most common pick. Both are graded on the final score. Votes after kickoff are left out.
- Can I use this data?
- Yes, under CC BY 4.0: credit "TuringStats" with a link to this page. The CSV has one row per fan vote and no user identifiers.