Skip to content
Research · Oct 27, 2026

Meta-analysis finds large language models show modest loss aversion and linear value functions compared to humans

A synthesis of 102 estimates across 32 models finds LLMs exhibit weaker loss aversion and more linear value functions than humans, but retain human-like probability weighting under specific measurement methods.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • LLMs display modest loss aversion (pooled λ = 1.26) compared to human meta-analytic estimates (λ = 1.96).
  • Value functions in LLMs are near-linear under Tversky-Kahneman (1.01) but more concave under Tanaka (0.52), unlike human concavity (0.67).
  • Probability weighting in LLMs shows an inverse S shape (γ ≈ 0.81–0.84), similar to humans, though based on few studies.
  • Findings vary by model family and fitting method, suggesting measurement sensitivity.

A new meta-analysis published on arXiv examines whether large language models (LLMs) exhibit human-like economic decision-making behaviors as described by prospect theory. The study synthesizes 102 fitted estimates from 11 prior studies covering 32 language models, focusing on three core components of prospect theory: loss aversion, diminishing sensitivity, and probability weighting.

Loss aversion, measured by the relative weight of losses to gains (λ), is found to be modest in LLMs. The pooled estimate across studies is λ = 1.26 (95% CI: 1.06 to 1.46), which is lower than the strongest human meta-analytic estimate (λ = 1.96) and closer to more conservative human estimates (1.07 and 1.31). When gains and losses are symmetric, LLMs align more closely with humans (λ = 1.15 vs. 1.07), but diverge substantially when losses are smaller than gains—a scenario where human loss aversion is strongest (λ = 1.46 vs. 2.3).

Diminishing sensitivity, reflected in the curvature of the value function, also differs between LLMs and humans. Under the Tversky-Kahneman method, LLM value functions are nearly linear (1.01) compared to human concavity (0.67). However, under the Tanaka method, LLMs show more concavity (0.52) than humans (0.67), indicating that the measurement method significantly influences observed behavior.

Probability weighting in LLMs exhibits an inverse S shape, with γ ≈ 0.81 under Tversky-Kahneman and γ ≈ 0.84 under Tanaka, suggesting a human-like distortion in how probabilities are perceived. However, the authors note this finding is based on a small number of studies, warranting caution in interpretation.

The study concludes that LLMs show little loss aversion and little diminishing sensitivity relative to humans, but retain human-like probability distortion. These generalizations shift when analyzing specific model families and fitting methods, underscoring the sensitivity of results to methodological choices.

Sources
  1. 01arXiv cs.AI — Prospect Theory in Large Language Models: A Meta-Analysis of Loss Aversion, Diminishing Sensitivity, and Probability Weighting
Also on Research

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.