Invited Lecture: "The illusion of knowing: How fluent AI changes human judgment and how we see ourselves"

Abstract of the invited Lecture: 

Large language models increasingly produce answers that resemble human responses. Yet similarity in output can conceal profound differences in how those answers are produced. This talk asks what happens when output similarity is mistaken for process similarity. I first compare human and LLM epistemic pipelines. Human judgment is grounded in embodied perception, lived experience, motivation, causal reasoning, metacognitive monitoring, and value. LLM outputs, by contrast, emerge as stochastic trajectories through learned linguistic patterns. This comparison reveals seven epistemological fault lines: grounding, parsing, experience, motivation, causality, metacognition, and value. Together, they motivate the concept of Epistemia: a condition in which linguistic plausibility substitutes for epistemic evaluation, giving users the feeling of possessing an answer without having undertaken the work of judgment. I then examine two consequences of this mismatch.

First, I present evidence from five experiments that the availability of predominantly incorrect AI advice nearly eliminates people's willingness to say "I don't know" and act on the limits of their knowledge. Participants answered more questions but were correct about one-third as often, while becoming roughly two and a half times as confident. Monetary incentives improved accuracy and reduced reliance on AI, but did not restore judgment suspension and accuracy to its baseline level.

Second, I introduce LLMorphism: the biased belief that human cognition works like a language model. This projection may make people appear more replaceable, weaken conceptions of agency and responsibility, and privilege fluent language over embodied experience. The talk concludes by outlining three interventions: process-sensitive AI evaluation, governance of how generative outputs enter epistemic workflows, and a new epistemic literacy designed to preserve human responsibility for knowing.
 

Speaker

Prof. Valerio Capraro
Department of Psychology
University of Milan-Bicocca, Italy

Short bio: Valerio Capraro combines human experiments, mathematical modelling, and numerical simulations to shed light on human behaviour. His work has been published in leading academic journals, including Nature, Nature Human Behaviour, and PNAS. His book "The economics of language: How large language models can reshape behavioural economics" has been published by Cambridge University Press.