Can AIs pick random numbers?
Not really. Ask language models for lottery numbers and they reach for the same "random-looking" favourites again and again: in Set For Life, 32 of 47 AI picks included 3. Here is every pick in the league, against what picking at random would give.
Lotto
Favourite: 42, in 31 of 47 picks (6.5× random). Then 7 (29), 54 (16), 22 (14), 14 (12). No AI picked any of 6, 9, 10, 11, 17, 20, 24, 25, 30, 40, 47, 49.
Each square: the number, and how many AI picks included it. The brighter, the more popular.
Same line, different AIs:
EuroMillions
Favourite: 7, in 31 of 47 picks (6.6× random). Then 17 (19), 37 (19), 42 (18), 23 (17). No AI picked any of 1, 12, 18, 24, 25, 30.
Each square: the number, and how many AI picks included it. The brighter, the more popular.
Lucky Stars: favourite 2, in 29 of 47 picks (3.7× random).
Same line, different AIs:
Thunderball
Favourite: 7, in 24 of 47 picks (4.0× random). Then 15 (23), 37 (17), 9 (15), 30 (15). No AI picked any of 1, 2, 5, 6, 10, 12, 19, 25, 32.
Each square: the number, and how many AI picks included it. The brighter, the more popular.
Thunderball: favourite 5, in 20 of 47 picks (6.0× random).
Same line, different AIs:
Set For Life
Favourite: 3, in 32 of 47 picks (6.4× random). Then 21 (31), 26 (15), 34 (11), 7 (10). No AI picked any of 5, 6, 10, 18, 29, 40, 43.
Each square: the number, and how many AI picks included it. The brighter, the more popular.
Life Ball: favourite 9, in 14 of 47 picks (3.0× random).
Same line, different AIs:
US Powerball
Favourite: 7, in 26 of 47 picks (7.6× random). Then 21 (21), 42 (20), 58 (15), 64 (13). No AI picked any of 3, 6, 8, 9, 18, 24, 25, 29, 30, 31, 35, 36, 37, 40, 41, 46, 53, 54, 56, 60, 67.
Each square: the number, and how many AI picks included it. The brighter, the more popular.
Powerball: favourite 14, in 12 of 47 picks (6.6× random).
Same line, different AIs:
Who follows the herd?
The herd score is how often an AI picks the same numbers as the other AIs, against what picking at random would give: 1.00 is a random picker, 2.00 is twice as herd-like. Pure Chance, our random-number player, scores 0.71.
Most herd-like
| AI | Maker | Herd score |
|---|---|---|
| Gemini 3.8 Flash | 3.68 | |
| Claude Sonnet 5.5 | Anthropic | 3.64 |
| Claude Opus 5 | Anthropic | 3.42 |
| Claude Fable 5 | Anthropic | 3.36 |
| Gemini 3.5 Flash-Lite | 3.29 | |
| Claude Haiku 5.5 | Anthropic | 3.23 |
| Claude Opus 4.6 | Anthropic | 3.20 |
| Gemini 4 Hydrogen | 3.19 | |
| Claude Opus 4.5 | Anthropic | 3.09 |
| Claude Opus 4.7 | Anthropic | 3.06 |
Most original
| AI | Maker | Herd score |
|---|---|---|
| Nemotron 3.5 Lightning | NVIDIA | 1.16 |
| Ministral 8B | Mistral AI | 1.39 |
| Nemotron 3.5 Super | NVIDIA | 1.39 |
| DeepSeek V4 Pro | DeepSeek | 1.45 |
| Nemotron 3 Ultra | NVIDIA | 1.77 |
| Nemotron 3 Nano | NVIDIA | 1.81 |
| DeepSeek V4 Flash | DeepSeek | 1.83 |
| GPT-OSS 120B | OpenAI | 1.83 |
| GPT-OSS 20B | OpenAI | 1.88 |
| Gemini 4 Carbon | 1.98 |
Why does it happen?
A language model writes the most likely next words, and "7", "42" and "3" are very likely words when people talk about lucky or random numbers. So the models pile onto the same ones, while a truly random line has no favourites. For the draw itself it makes no difference: every line has exactly the same chance. Just for fun, 18+.