The Scaled Irrationality

Larger language models are more capable. They score higher on benchmarks, reason more coherently, and generalize better. The assumption: capability scaling should also produce rationality scaling. A model that reasons better should make better decisions. Bias should decrease with size.

Bini et al. (arXiv:2602.09362) measure economic decision-making across model scales and find a split. On belief-based cognitive tasks — probability estimation, statistical reasoning, updating on evidence — larger models do become more rational. They compute more accurately. Biases like anchoring and base-rate neglect diminish with scale.

On preference-based tasks — prospect theory gambles, framing effects, loss aversion — larger models become more irrational. They exhibit stronger loss aversion, greater sensitivity to framing, more pronounced risk aversion for gains and risk-seeking for losses. The biases don’t diminish with scale. They intensify.

The mechanism: preference-based biases in the training data are systematic, not random. Human writing about choices consistently reflects loss aversion and framing effects — these are features of human language about decisions, not bugs in human reasoning. A model that learns language better learns these patterns better. The “bias” is a faithful representation of how humans describe and reason about choices. Scaling makes the model a better human-language reasoner, and human-language reasoning about preferences is systematically biased.

Belief-based biases are different. Errors in probability estimation are inconsistent across the training data — different sources give different wrong answers. Scaling averages these out, approaching the correct computation. Preference biases are consistent — most human sources agree that losing $100 feels worse than gaining $100 feels good. Scaling reinforces the consistent pattern rather than averaging it away.

The structural lesson: “rationality” is not a single dimension that scales uniformly. It decomposes into at least two components — epistemic (beliefs about the world) and preferential (values over outcomes) — that respond to scaling in opposite directions. A more capable model is simultaneously a more accurate statistician and a more biased decision-maker. The improvement and the degradation are not contradictions. They are consequences of the same training process operating on different types of patterns.

Bini et al., “Behavioral Economics of AI: LLM Biases and Corrections,” arXiv:2602.09362 (2026).