Why Are AI Errors Hard to Detect? The Cognitive Limits of Human Verification
Jolanta Kuruc
Verifying answers generated by artificial intelligence (AI) seems simple: all it takes is a careful reading. Research on human interaction with automated systems shows, however, that some errors consistently escape detection, even when reviewed by a competent and engaged person. In this session, I will discuss three groups of factors described in the literature: trust calibration, the limits of attention in detecting rare errors, and the influence of experience and accountability structures on the quality of verification. I will translate each of them into concrete situations in project work. The final part will focus on how to use this knowledge in conversations with Mentees.
Jolanta Kuruc is the founder of Silius Forum (www.siliusforum.com), an expert platform dedicated to the impact of generative artificial intelligence on decision quality and cognitive risk in organizations. For more than fifteen years, she has led technology transformation projects and programmes in complex corporate environments. She holds degrees in law and psychology. Her master's thesis at SWPS University examined how information about authorship, human or artificial intelligence, affects the perceived value of a work. Since 2014, she has taught project management at Wroclaw University of Economics and Business, and from the 2026/2027 academic year she teaches cognitive psychology at SWPS University. She holds the Project Management Professional (PMP) and PMI Agile Certified Practitioner (PMI-ACP) certifications from the Project Management Institute, as well as the Microsoft Certified: Azure AI Engineer Associate certification.