Unsolicited reflections on uncertainty in artificial intelligence-based decision support systems

In this article, I address my concerns regarding the implications of uncertainty in decision support systems, with a particular focus on its effects on AI-based decision support systems (AI-DSS). The genesis of this article stems from the reflections shared in a previous piece, where I explore the concept of uncertainty beyond the realm of data quality (here is the link).

Decision Support Systems Overview - A decision support system (DSS) is a software system designed to collect, organize, and analyze company data to provide valuable information for decision-making or to support management and planning activities. Essentially, a DSS aids individuals tasked with making strategic decisions in complex contexts where determining the optimal choice or decision-making strategy is challenging. Their reliability hinges on the algorithms used to process data and the quality of the dataset utilized, constituting a simplified model of reality derived from the data available to the decision-maker.

AI-based Decision Support Systems - In recent years, DSS software systems have evolved with the integration of artificial intelligence (AI) to enhance the reliability of the representation model of reality upon which calculations are based. AI autonomously constructs the representation model of the analyzed universe (AI model) solely from the dataset provided by analysts.

The Importance of Dataset Selection - An AI develops its representation model based solely on the dataset designed by analysts. However, since the real world surpasses our ability to synthesize, analysts strive to strike a balance between capturing enough dimensions to represent macro dynamics while avoiding complexity that hampers result verification.

Emerging Doubts - Despite meticulous dataset design, it remains one of many possible representations of the real-world environment. As one form of uncertainty is linked to environmental complexity, doubts arise regarding whether the dataset itself is immune from environmental uncertainty. This concern transcends bias and addresses a potentially impactful yet less tangible aspect.

Consideration of Dimensions - Doubts arise concerning the selection of dimensions within the dataset and the uncertainty surrounding their influence on the AI model and processing outcomes. Unexpected interactions or interdependencies among dimensions could affect processing results, even those deemed marginal or independent.

Artificial Intelligence Development Process - AI systems operate through neural networks trained for specific tasks, utilizing deep learning. These networks employ layered structures where each layer contributes to final processing, with the ability to learn and solve complex problems autonomously. However, the nonlinear data processing within neural networks renders their processing opaque, resembling black boxes.

Certainty of Results - The primary limitation of AI today lies not in computing power or data availability but in the structure of AI models. Without a comprehensive understanding of the context, caution is warranted when entrusting AI to identify solutions, as it may generate optimal solutions in contexts diverging from reality.

Article source: Linkedin Article by Vincenzo Gioia