Towards the effective and responsible use of imperfect NLP-generated learning content in STEM education

Artificial intelligence (AI) has been a building block to provide adaptive instruction and learning support at scale since the 1970s. In the past decades, researchers have mainly relied on knowledge-based AI to construct a domain model for learner-facing interactive personalized instruction systems, i.e., encoding task domain knowledge in symbols, logic, and rules. More recently, the use of data-driven AI to extract domain knowledge from data has gained attention due to its potential to better represent domains with open or changing worlds, reduce expert-authoring costs, and broaden the scope of user interaction. However, the learning content created based on a data-driven AI domain model is much more likely to contain inaccurate or incomplete information. This has raised concerns about its use, especially since it has become more accessible after the public launch of ChatGPT and similar tools. Should we deploy a system that uses imperfect AI-generated learning content, given its potential harm? If so, how should we design and deploy it effectively and responsibly? To provide insights into these questions, I systematically investigated how adult learners perceive, interact with, and get impacted by imperfect AI-generated content in STEM learning. My dissertation focuses on two types of learning content created with Natural Language Processing (NLP) techniques: (1) formative correctness feedback for short-answer questions and (2) natural language responses to learner-initiated interactive help-seeking. Using a socio-technical lens and a mixed-methods approach, I contributed actionable recommendations on learner support, system design, and system deployment to help diverse learners gain the most from imperfect AI-generated learning content. The dissertation further demonstrates the need to use caution when deploying such imperfect content and provides guidelines for conducting impact assessments before deployment.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Li, Wenting (Tiffany)

Contributors dc:contributor

• Karahalios, Karrie
• Sundaram, Hari
• Zilles, Craig
• Kulkarni, Chinmay

• Artificial Intelligence
• Natural Language Processing
• Personalized Learning
• Autograder
• Pedagogical Chatbot
• Conversational Agent
• Error Management
• Differential Impact
• Fairness
• Hallucination
• Reliance
• Stem Learning
• Large Language Model

Statement dc:rights

• Copyright 2025 Wenting (Tiffany) Li

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130101

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130101