Answers are grounded in this paper and cite their sources.
Understanding human cognitive processes for reasoning is a key step towards developing human-level AI reasoning systems. Syllogistic reasoning — a classical form of deductive reasoning — is a well-established paradigm for investigating human reasoning, known to elicit systematic cognitive biases such as belief bias and atmosphere effects. In this study, we compare the reasoning patterns of humans and three state-of-the-art Large Language Models (LLMs), namely GPT-4, GPT-4o and o4-mini, across multiple syllogism types with belief andatmosphere effects. Using an online questionnaire, we examine the biases in human syllogistic reasoning and compare this with the LLMs responses. We identify biases and differences in accuracy between different LLMs and humans. Belief-inconsistent statements, which contradict generally accepted common knowledge, and the atmosphere effect impact human and LLM answers. Our results show that, although LLMs outperform human participants, they exhibit human-like cognitive effects. We further evaluate the new reasoning model o4-mini, which performs nearly perfectly on the syllogism reasoning tasks and compare logic, human reasoning and LLM reasoning.
No takes yet. Share an insight, caveat, or question.
Krause et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: