This meta‑analysis aggregates 102 fitted estimates from 11 studies covering 32 large language models to examine whether they exhibit the three core components of prospect theory: loss aversion, diminishing sensitivity, and probability weighting.
The pooled loss‑aversion coefficient is $\lambda=1.26$ (95% CI 1.06–1.46), well below the strongest human meta‑analytic estimate $\lambda=1.96$ but comparable to more conservative human values (1.07, 1.31). When gains and losses are symmetric, models show $\lambda=1.15$ versus the human 1.07; however, when losses are smaller than gains, the model $\lambda$ drops to 1.46, far beneath the human 2.3 where loss aversion peaks.
Value‑function curvature depends on the measurement method. Under the Tversky‑Kahneman approach, models are nearly linear with a curvature of 1.01, contrasting with the human value of 0.67. Using the Tanaka method, the model curvature is 0.52, still less concave than the human 0.67, indicating that diminishing sensitivity is method‑dependent.
Probability weighting displays a human‑like inverse‑S shape: $\gamma=0.81$ (TK) and $\gamma=0.84$ (Tanaka), although based on few studies.
In sum, large language models exhibit modest loss aversion and limited diminishing sensitivity, yet retain human‑like probability distortion. These patterns shift subtly across model families and fitting procedures.
Review: The work offers a rigorous quantitative benchmark, highlighting that current LLMs differ markedly from humans in risk perception—especially loss sensitivity—providing valuable guidance for alignment and human‑AI interaction research.