Illusory Pattern Perception in Large Language Models

This dataset was developed to compare human participants and large language models in illusory pattern perception, defined as the tendency to perceive nonexistent regularities, associations, or meaning in random, ambiguous, or insufficiently supported information. The dataset covers four broad categories of experimental tasks involving social judgment, economic decision-making, causal inference, and visual recognition. It consists of both human participant data and large language model data.For the human participant component, 100 participants were recruited through Prolific for each task. The large language model data consist of experimental responses generated by GPT-3.5, GPT-4, GPT-4o, GPT-5, and Doubao. Across all experiments, a total of 20,800 independent queries were conducted. After excluding responses with missing dependent variables or outcomes that could not be validly coded, 20,761 valid responses were retained. These include 1,789 valid responses from the baseline experiments, 2,798 from the temperature-manipulation experiments, 14,394 from the role-playing experiments involving eight demographic personas, and 1,780 from the chain-of-thought prompting experiments. By model, the dataset contains 3,584 valid responses from GPT-3.5, 3,578 from GPT-4, 4,800 from GPT-4o, 4,000 from GPT-5, and 4,799 from Doubao.