Anthony Sicilia
West Virginia University, WV·Computer Science
When someone pushes back on an AI chatbot's answer, the system often cannot tell whether it made a mistake, the user is unsure, or the conversation has moved to a new goal. Over long conversations these misunderstandings pile up, eroding trust and usefulness.
- Fields
Computer Science
- Topics
artificial intelligence
human-AI interaction
trustworthy AI
- Grants
- 1 known grant, through Aug 2029
What they work on
- Develop a Probably Approximately Correct (PAC) learning theory that measures how user disagreement, suggestions and topic shifts change the calibration of an AI system's confidence during a conversation.
- Extend the framework to estimate differences between the AI model's uncertainty and the user's uncertainty during dialogue.
- Use these uncertainty estimates to decide when an AI should clarify, revise or keep its answer, and test the strategies in coding dialogues with novice programmers.
- Evaluate whether confidence estimates and dialogue strategies stay reliable over long conversations and repeated interactions, and in negotiation and social dialogue.
More about the research
This matters as AI tools are used in education, healthcare and research. This collaborative project builds the foundations for AI that tracks how a conversation changes how much it should trust its own confidence. Such a system could then respond by clarifying, correcting itself or asking a follow-up question.
Why it matters
Better-calibrated AI assistants could make human-AI collaboration safer and more transparent in settings such as programming education, healthcare and decision support. There, misplaced confidence or unnoticed misunderstandings can mislead users. The planned open datasets, benchmarks and software could help other developers test whether conversational AI systems handle uncertainty reliably.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Anders Host-Madsen
University of Hawaii at Manoa, HI·Engineering
Derive fundamental performance bounds for machine learning of lossless source coding within the probably approximately correct (PAC) learning framework, and identify which factors control how fast learning converges.
Chirag Shah
University of Washington, WA·Computer Science
Build a proactive conversational agent on an offline hierarchical model-based deep reinforcement learning framework, in which a high level picks conversation strategies and a low level generates responses.
Lu Lin
Pennsylvania State University, PA·Computer Science
Identify critical LLM safety concerns, gaps in current evaluation and infrastructure needs through structured surveys and interviews with experts across computing disciplines.
The school
West Virginia University: 15 labs on PI Pages, $259M a year on research, 201 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.