2026-10-12T18:53:17Z https://www.scidb.cn/oai

10.57760/sciencedb.psych.01080 2026-08-22T21:02:54Z

2026-08-22 Human-LLM Social Risk Decision Analysis: Data and Code doi:10.57760/sciencedb.psych.01080 en Welcome to the repository for this research project. This repository contains the data, analysis scripts, statistical outputs, and visualization files used in Study 1 and Study 2.Research background and objectivesFraming effects are not determined solely by how outcomes are described but are also influenced by socially meaningful contextual cues. Based on the Ambiguity and Ambivalence (AA) hypothesis, this project examines how decision makers dynamically weight primary social cues, such as group size and kinship constituent, and secondary verbal cues, such as positive versus negative framing. The primary objective was to compare human participants and large language models (LLMs) in terms of their overall risk preferences, framing effects, and sensitivity to changes in cue ambiguity and preference ambivalence. A further objective was to test whether increasing the accessibility of primary social cues would reduce behavioral differences between humans and LLMs. Dataset descriptionThe dataset was collected across two studies using adapted versions of the Asian disease problem. In Study 1, primary social cues were manipulated through four group-size conditions (6, 60, 600, and 6,000 people) and four kinship-constituent conditions within a six-person group (1, 2, 3, or 6 family members). Each scenario was presented using either a positive or negative framing. Nine OpenAI LLMs, ranging from GPT-3.5 to GPT-5.4, were tested through standardized API calls. Each model generated 100 independent valid responses in every scenario-by-framing condition, resulting in 14,400 LLM responses. Human benchmark data were obtained from previously published behavioral studies using equivalent between-subject designs.In Study 2, six decision scenarios were retained, including two group-size conditions and four kinship-constituent conditions. Primary social cues were explicitly highlighted before the decision scenarios to increase their accessibility. Nine LLMs again generated 100 valid responses per scenario-by-framing condition, yielding 10,800 responses. In addition to choosing between a certain and a risky option, both LLMs and human participants rated their emotional valence toward each scenario on a 7-point scale.The human benchmark sample in Study 1 consisted of 768 English-speaking participants, including 384 men, with a mean age of 20.81 years. Of these, 368 participated in the group-size conditions and 400 participated in the kinship-constituent conditions. Study 2 included 200 native Chinese-speaking participants recruited through the Credamo online platform, including 71 men, with a mean age of 30.79 years. All participants in Study 2 provided informed consent and received monetary compensation. Framing effect; Large language model; Decision bias Feng Xiao Yining Mao X.T. (XiaoTian) Wang PUBLIC https://creativecommons.org/licenses/by-nd/4.0 dataset Science Data Bank