From complacency to collapse: how bias propagation undermines decision-making

Purpose: In light of the growing integration of Large Language Models (LLMs) into business operations and decision-making processes, this paper analyses the role of user cognitive biases in the propagation of interpretative errors within Human-AI Interaction processes, with particular reference to research and development (R&D) contexts. Conceptual framework: The paper proposes a conceptual model, the Bias Amplification Loop (BAL), to describe the dynamics through which the user’s initial cognitive framing and the probabilistic mechanisms of LLMs tend to reinforce one another in successive cycles of interaction, with particular reference to the different exploration-exploitation phases typical of R&D processes. Implications: The analysis highlights how the unregulated integration of LLMs can foster cumulative processes of interpretative convergence, a reduction in cognitive diversity, and a progressive narrowing of the exploratory space. If not adequately mitigated, these dynamics risk compromising organisations’ ability to evaluate innovative alternatives and adapt to competitive contexts characterised by high uncertainty. Originality: The BAL framework offers a dynamic and path-dependent perspective on the epistemic vulnerabilities associated with the use of LLMs in organisational decision-making processes, whilst proposing an operational interpretative framework for AI governance in innovation contexts.