Current AI systems have achieved notable progress in specific tasks such as image interpretation and natural language processing, yet they remain confined to narrow applications and do not exhibit the broad, adaptable intelligence seen in humans. The authors of a 2021 arXiv paper contend that to overcome this limitation, AI research should look to human cognitive mechanisms, particularly the distinction Daniel Kahneman made between fast, intuitive thinking and slow, deliberate reasoning.
They propose a multi-agent architecture in which incoming problems are routed to either system 1 agents that act quickly by relying solely on past experience, or system 2 agents that are engaged when deeper reasoning is required to seek optimal solutions beyond what system 1 can provide. Both types of agents are supported by two complementary models: a world model that encodes domain knowledge about the environment, and a self model that records the system’s past actions and evaluates its own skills and limitations.
While the framework offers a clear structure for integrating metacognitive processes into AI, the paper presents it primarily as a conceptual idea without empirical results from implementation. The authors note that realizing such a system would require substantial technical work to design accurate world and self models, and to calibrate when to switch between fast and slow processing—a challenge that mirrors metacognitive difficulties in human cognition.