Is AI Instruction Comparable to Human Instruction? Designing a Pedagogical Agent for Complex Task Training

EasyChair Preprint 15730

Abstract

Complex tasks can be difficult to train and difficult to learn. Though train-ing one-on-one with a human instructor is considered the gold standard for learning [1], social agency theory posits that social cues in virtual instructors can elicit the same mental processes in learners that occur with human in-structors [2]. We contend that artificial intelligence (AI)-driven pedagogical agents (PAs) could be superior for analyzing learner data, adapting instruc-tional content, and providing scalable, consistent, on-demand training, par-ticularly when instructors are expensive or unavailable. Ideally, an AI-driven PA would surpass the capability of a human instructor if these features could be fully realized. Because there are key differences in the way that humans and AI present learning content, we must ascertain whether a PA can deliver a quality of learning similar to a human instructor. We compared the knowledge levels of two groups after learning a complex task with either a human instructor or PA. We present nonexperimental preliminary findings that suggest learning from a human instructor and PA are comparable and discuss instructional considerations for PA design approach.

Keyphrases: Artificial Intelligence, Artificial Intelligence AI, Complex Learning, Complex task, Human instructor, Social agency theory, Virtual Instructors, adaptive instructional systems, adaptive training, assistive feedback, corrective feedback, designing effective instructional algorithms, feedback algorithm, gold standard, human instruction, human instructors, human performance, social cues, visual cues

@booklet{EasyChair:15730,
  author    = {Bradford Schroeder and Jason Hochreiter and Sean Thayer and Javier Rivera and Wendi Van Buskirk},
  title     = {Is AI Instruction Comparable to Human Instruction? Designing a Pedagogical Agent for Complex Task Training},
  howpublished = {EasyChair Preprint 15730},
  year      = {EasyChair, 2025}}