Supporting Learners' Use of Imperfect Generative Pedagogical Chatbots: The Role of Chatbot Response Uncertainty and Reduced Verbosity

Paper Title

Supporting Learners' Use of Imperfect Generative Pedagogical Chatbots: The Role of Chatbot Response Uncertainty and Reduced Verbosity

Publication Info

• Topic area: Enhancing the design of generative pedagogical chatbots to improve learning outcomes and error management.
• Keywords: Generative chatbots, pedagogical AI, verbal uncertainty, verbosity, cognitive engagement, error detection, trust calibration, STEM education, Bayesian inference, learning outcomes.

Background and Problem

• Problem / challenge: Learners often overtrust generative chatbots, fail to detect errors, and engage shallowly with chatbot responses due to cognitive overload, reduced mental disturbance, and the prioritization of short-term gains. Current chatbots are designed to provide confident, verbose responses, which exacerbate these issues.
• Significance: Addressing these challenges is critical for leveraging generative chatbots as effective educational tools, ensuring learners achieve deeper cognitive engagement, better error detection, and improved learning outcomes.
• Motivation and related work: Prior studies have explored interaction instructions, reflection scaffolds, and content improvements for chatbots but have not examined how linguistic features like verbal uncertainty and verbosity affect learning. This study fills that gap by investigating these features in a STEM learning context.

Solution

• Proposed approach: Investigate the effects of two chatbot design features—verbal uncertainty and reduced verbosity—on learners’ cognitive engagement, error detection, and learning outcomes.
• Novelty:
  1. Systematic evaluation of verbal uncertainty in generative chatbots to address overtrust and enhance engagement.
  2. Examination of reduced verbosity as a means to improve error detection and reduce cognitive load.
  3. Integration of findings from human-AI interaction and educational psychology to inform chatbot design.
  4. Use of Bayesian causal inference and thematic analysis to assess interventions in an ecologically valid learning environment.
• Procedure and key techniques:
  1. Designed a learning platform where participants used chatbots with varied response styles (uncertain vs. certain, short vs. long) in a 2x2 factorial design.
  2. Conducted experiments with 177 survey participants and 26 interviewees, focusing on introductory statistics topics.
  3. Measured cognitive engagement (use of alternative resources), trust perceptions, error detection, and learning outcomes through pre-tests, post-tests, and surveys.
  4. Analyzed data using Bayesian hierarchical models and thematic analysis of qualitative feedback.

Results

• Concrete findings:
  1. Reduced verbosity improved error correction for factual errors with logical fallacies (FE-w-LF) but did not encourage the use of alternative resources.
  2. Verbal uncertainty reduced reliance on chatbot responses and improved learning outcomes when participants encountered unrelated errors but led to worse outcomes in error-free conditions.
  3. Uncertainty expressions decreased trust perceptions (e.g., willingness to depend, perceived ability) but had mixed effects on perceptions of integrity and benevolence.
• Advantage over baselines:
  • Reduced verbosity decreased cognitive load and improved logical error detection compared to verbose responses.
  • Verbal uncertainty calibrated reliance on chatbot responses, reducing the adoption of incorrect outputs.
• Experiments / evaluation:
  • Conducted in a controlled online learning environment with 262 valid participant-topic pairs.
  • Used Bayesian causal inference for quantitative analysis and thematic analysis for qualitative insights.
  • Evaluated effects across three error types: no error, factual errors with logical fallacies, and factual errors without logical fallacies.
• Limitations and future work:
  • Limited generalizability to other domains, error types, and chatbot designs.
  • Short-term study; longitudinal effects and classroom settings remain unexplored.
  • Future work should refine uncertainty signals, explore alternative designs, and evaluate long-term impacts on learning behaviors.

Summary

This study investigated how verbal uncertainty and reduced verbosity in generative pedagogical chatbots affect learners’ cognitive engagement, error detection, and learning outcomes in STEM education. Reduced verbosity improved error correction for logical fallacies but did not enhance engagement with alternative resources. Verbal uncertainty reduced reliance on chatbot responses and improved learning outcomes when errors were encountered but led to worse outcomes in error-free conditions. These findings highlight the potential of linguistic design adjustments to address challenges in chatbot-assisted learning but emphasize the need for nuanced implementation and further research in diverse contexts.