Can machines be uncertain?

Researcher Luis Rosa has published a thought-provoking paper titled 'Can machines be uncertain?' on arXiv, investigating whether artificial intelligence systems can genuinely experience states of uncertainty akin to human cognition. The paper adopts a functionalist and behavioral perspective to examine how different AI architectures—including symbolic, connectionist, and hybrid systems—might accommodate uncertainty. Rosa's work arrives at a crucial moment as AI systems like GPT-4 and Claude 3 increasingly make high-stakes decisions, raising fundamental questions about whether they merely simulate understanding or possess genuine cognitive states.

The paper makes several key distinctions, separating epistemic uncertainty (inherent in data or information) from subjective uncertainty (the system's own attitude of being uncertain). It further differentiates between distributed and discrete realizations of subjective uncertainty within AI architectures. A particularly novel contribution is the concept that some uncertainty states represent interrogative attitudes whose content is a question rather than a proposition. This philosophical framework could influence how developers design AI systems to express confidence levels, potentially leading to more transparent and trustworthy AI that better communicates its limitations in critical applications from medical diagnosis to autonomous systems.