Deciphering AI Decision-Making: A Generalizable Framework of Human-AI Misalignment

Abstract

The field of AI has seen unprecedented advancements, edging close to Artificial General Intelligence (AGI). Latest AIs represented by Large Language Models (LLMs) have shown human-like intelligence, reshaping our understanding of AI cognition. This is particularly evident in AI’s latest capability to make complex decisions with partially observable data and conflicting objectives or values, akin to human decision-making. There is a pressing need for understanding AI decision-making as a cognitive phenomenon instead of a pure engineering system. However, there are no effective tools for decomposing, deconstructing, and measuring the intrinsic and internal processes of AI decision-making as a high-level cognitive process. In addition, vast differences between humans and AI can exist in the execution of key stages of decision-making, posing uncertainties for human-AI alignment. This research aims to develop a metacognitive framework that allows for the precise deconstruction of AI decision-making processes in drone swarm controls, enabling us to benchmark AI against human cognition at each critical stage. Analytical approaches for quantifying and measuring the sources of human-AI misalignment will also be proposed, particularly in drone swarm applications. By understanding where and how AI decision-making diverges from human processes, this research is expected to develop strategies to align AI systems more closely with human values, decision-making patterns, and cognitive processes.

Ähnliche Arbeiten