International study suggests ways to train large language models in wise reasoning and measure the wisdom of AI
A new study is the first to suggest realistic ways to integrate wisdom into artificial intelligence, to create AI systems that will be more robust, transparent, cooperative and safe.
Researchers from the University of Waterloo led the team, which includes experts in psychology, computer science, and engineering. Their paper proposes ways to train large language models to be wiser, explore new architectures that could support wise reasoning, and suggest benchmarks to measure AI wisdom.
The timing of the work is critical because as AI capabilities race ahead, wisdom isn’t keeping pace, raising safety and reliability concerns.
“Artificial intelligence is getting smarter every day, but one important human skill it lacks is wisdom,” said Dr. Sam Johnson, professor of psychology at Waterloo and co-lead author of the study. “Wisdom isn’t just about knowledge or intelligence. It’s about the mental skills needed to handle life’s challenges, such as making difficult decisions or navigating unpredictable social situations.”
Whereas current AI systems excel at well-defined tasks, they struggle when problems are messy or unclear, because they lack the full range of strategies that humans use to navigate uncertainty, according to the researchers. The reason: these AI systems lack the full toolkit humans use to handle uncertainty. The new approach focuses on teaching AI to think about its own thinking or metacognition — recognizing the limits of its knowledge, adjusting to different contexts, weighing multiple viewpoints, and staying flexible to how situations might unfold.
“Wisdom has seemed too philosophical, too human-centred to formalize for machines,” said Dr. Igor Grossmann, professor of psychology at Waterloo and study co-lead. “But by breaking it down into specific strategies such as intellectual humility, perspective-seeking, and context adaptation, we can create a concrete roadmap for building AI that doesn't just compute, but reasons wisely.”
The researchers propose that wise AI systems could handle novel problems and environments, work more cooperatively in pursuing shared goals, be more explainable to users, and be safer by better aligning goals with human users.
“If the smartest person in the world were a toddler, we still wouldn’t hand them the nuclear codes,” Johnson said. “AI is increasingly resembling a child genius, still needing a healthy dose of wisdom from its human parents.”
Researchers from the University of Waterloo, Université de Montréal, the Max Planck Institutes, Santa Fe Institute, Stanford University, Warwick Business School and Google DeepMind contributed to this work. Next steps include collaborating with industry to develop computational models of human wisdom to guide AI design.
The paper, Imagining and building wise machines: The centrality of AI metacognition, appears in in Trends in Cognitive Sciences.
