Efekt zakotvení u velkých jazykových modelů v kontextu manažerského rozhodování

This bachelor’s thesis examines the anchoring effect in large language models in the context of managerial decision-making. The aim of the thesis is to empirically verify whether LLM outputs in quantitative managerial estimates are influenced by numerical values included in the input prompt. The experiment was conducted using six managerial scenarios focused on cost estimation, project duration estimation, and risk estimation. In total, 1,200 responses were collected from four language models. The strength of the anchoring effect was evaluated using the Anchoring Index, and the statistical significance of the differences was tested using a t-test. The results show that the anchoring effect appeared across all types of tasks and all tested models. At the same time, the Ignore Anchor Hint strategy was found to significantly reduce Anchoring Index values; however, it cannot be considered a universal solution. The thesis highlights the need for critical handling of numerical values in prompts and provides recommendations for the more responsible use of LLMs in managerial practice.