On Friday, Sept. 11, 2026, students and faculty of Colgate University attended “Exploring AI’s Influence on Higher-Order Reasoning.” Trent Cash, PhD, Lupina Foundation Postdoctoral Scholar in the School of Psychology at the University of Waterloo, spoke about how artificial intelligence (AI) affects cognition, metacognition, and learning.
Cash’s presentation is part of the NASC Colloquium, a weekly series inviting Colgate faculty and guest lecturers to share their expertise in the natural sciences and mathematics, sponsored by the CORE Sciences and the Center for Learning, Teaching, and Research.
Douglas Johnson, dean of academic and curricular affairs and William R. Kenan Jr. Professor of psychological and brain sciences, introduced Cash to the audience. Johnson explained that the goals of a liberal arts education include clear thinking and communication, evaluating evidence, and making judgments to become aware of our limits while challenging ourselves to learn throughout life.
“It is not whether AI is ‘good’ or ‘bad’ for learning. Much depends on what the technology is being used for, when it is being used, and what we are trying to accomplish,” Johnson explained.
The first section of Cash’s presentation revolved around AI and the concept of cognitive offloading — the physical act of displacing thinking to another subject, like using a calculator to do addition instead of solving it in your head or asking a large language model (LLM) how to solve a math set for homework. Cash explained that if a student attempting to learn a language relies on an AI translator for every conversation, they are less likely to learn the language. At the same time, he argued that if a student were to prompt AI to give feedback on their work, doing so would incorporate AI into a learning process that synthesizes information through reflecting, rebutting, and accepting what matters.
Cash then focused on metacognition — the brain’s assessment of situations, like deciding after taking a quiz how concrete one’s understanding of a subject is, based on the cognitive cues of taking the quiz. These specific learning assessments are also known as judgments of learning (JoLs). Humans are less likely to register doubt from LLMs because they do not convey facial cues or other social expressions that show potential doubt. Cash then described one of his studies that focused on participants choosing whether to trust AI chatbots’ judgments of their learning of words. Cash explained that the participants were more likely to take the AI’s JoL into account than not, so the focus is not about whether humans trust LLMs; it’s how much.
Cash’s final section addressed the ideas of truthfulness in LLMs, how they can judge themselves, and how humans judge their responses. Based on the study, Cash explained that current LLMs can give metacognitive judgments as accurately as humans can, and LLMs’ faithfulness — the ability to explain their thought process — has been improving over time.
“His perspective on the concept of faithfulness was interesting to me because understanding in better detail how these programs compute, for lack of a better word, is, in my opinion, fundamental, because literally all they do is compute,” said Malorie Miner ’28, an attendee with an English major and philosophy minor.
In his closing statements, Cash focused on using LLMs as collaborators who work with humans, not for humans. When asked about AI in the classroom, Cash responded that in-person learning is more important than ever because it can serve as laboratories where students can be challenged on how to use AI effectively and serve as spaces where students can get away from AI and exercise their brains.
“There is lots of work to be done by instructors who will have to figure out how to balance their goals with AI and to create assignments and activities that are maximally beneficial in this new world,” said Cash.