Fast Facts
-
The “gambler’s fallacy” wasn’t the main pattern: after long streaks, all eight small language models became more likely to predict another repeat—up to 60–82% versus the correct 50%.
-
Telling models the coin was fair barely helped. Their long-streak predictions stayed biased toward continuation, and instruction-tuning made little overall difference.
-
What models say can differ from what they predict. One Qwen model verbally favored the gambler’s-fallacy answer but assigned higher probability to the streak continuing.
-
The findings are suggestive, not universal: models partly adapted to streaky or alternating data, but the study tested only small open models and coin flips written as H/T.
A streak can mislead
The Monte Carlo roulette story warns against betting that a streak must end. The exact losses remain uncertain, but the mistake has a name: the gambler’s fallacy. With a fair coin, each flip stays independent. After 10 heads, the chance of another head remains 50%.
Models lean toward continuation
A test of eight small language models found a different pull. Researchers measured next-flip probabilities across 1,200 random sequences, then repeated the test with prompts that explicitly described a fair coin. After streaks of 10 or more, every model assigned a 60% to 82% chance to another repeat. The fair-coin reminder reduced that bias only slightly.
Short streaks produced more mixed results. Some models briefly leaned toward a flip in the opposite direction. Also, models did not always match their words: one model said a streak made a repeat less likely, yet its probabilities favored continuation.
Useful signal, with limits
The test suggests that models often copy patterns in their context rather than reason consistently about chance. Still, a streak can provide evidence when the coin’s fairness is unknown. In that case, expecting repeats may make sense.
The results cover small, open models and coin flips written as letters. Other models, prompts and real-world sequences may behave differently. For people using AI to forecast or simulate events, the takeaway is practical: check a model’s probabilities, not just its explanation.
Expand Your Tech Knowledge
Dive deeper into the world of Cryptocurrency and its impact on global finance.
Explore past and present digital transformations on the Internet Archive.
AITechV1
