The Fragility of Reasoning: A Comparative Analysis of LLM Performance in GSM-Symbolic and Social Manipulation via Irrelevant Data-Induced Sophisms

This paper explores the parallel fragilities of artificial and human reasoning when exposed to irrelevant or misleading information.Drawing from Apple’s GSM-Symbolic benchmark (2024) and psychological research on cognitive biases, it compares how large language models (LLMs) and humans experience reasoning degradation — up to 65% in symbolic tasks and up to 60% in social contexts influenced by sophisms and information overload. The study proposes an analogy between pattern-matching errors in AI and cognitive distortions in human decision-making, framing both as vulnerabilities in asymmetric information environments.Its goal is to foster reflection on data integrity, cognitive resilience, and the importance of transparent, open scientific practice. All references are publicly available academic sources, and the paper is intended purely for research and educational purposes within the Fast Science – Open Research initiative.