What researchers studied
Researchers in Israel examined how large language models choose between options with different risks and potential rewards. They tested eight models from four providers using hypothetical emergency and disaster-management scenarios. An adaptive binary-search method adjusted the value of a riskier option after earlier responses to locate the point where a model switched its preference. The experiment also compared decisions when expected values were stated explicitly with decisions requiring models to infer those values.
What they found
Six of eight models had median switching points near equal expected value in both presentations. However, individual responses showed deviations from expected-value maximisation. The deviations concentrated in four models: those from two Anthropic models favoured the safer choice, whereas those from two Google models almost always favoured the riskier choice. Explicitly presenting expected values reduced these deviations in all four. One Anthropic model showed a clearer shift between the two presentations than another.
What it means
The study suggests that how a decision is presented can change AI outputs even when the underlying probabilities are equivalent. This matters for organisations considering AI-assisted planning under uncertainty. The adaptive procedure offers a repeatable way to compare systems and identify inconsistencies before deployment. It should be understood as an evaluation tool rather than a guarantee that an AI recommendation will be appropriate.
Important limitation
The scenarios were hypothetical and used a particular probability structure. The researchers did not observe live emergency decisions or measure actual consequences. Model versions, prompts and application settings may change the results. No broad claim about the inherent risk preference of every model from a provider is warranted, and high-stakes decisions still require human judgement.
Source Information
Study: Measuring risk preferences in large language models using adaptive binary search
Authors: Idan-Chaim Cohen, Noya Littor, Aviad Elyashar, Odeya Cohen and Rami Puzis
Journal / Source: Frontiers in Artificial Intelligence
Published: 2 November 2026







