Research interests / Human-AI interaction
Designing how people understand and act on intelligent systems.
My work spans software engineering, product design, and AI-driven decision tools. At Verikai and SlateRoom, I encountered the same problem in different contexts: people must act on model outputs without always seeing the evidence or uncertainty behind them. I want to study how interfaces shape that judgment.
Focus areas
- 01
Communicating uncertainty
How should interfaces represent confidence, evidence, and limitations?
- 02
Human-AI decision-making
When do recommendations improve judgment, and when do they distort it?
- 03
Creative collaboration
How can predictive systems support creative decisions while preserving agency?
From practice to research
Where the questions came from
- Model evidenceWhat the system computed
- InterfaceWhat is shown, in what order, with what qualification
- Human decisionWhat the person does next
- Verikai / Applied problem
Making underwriting predictions understandable in a workflow where decisions have consequences.
Verikai case study - SlateRoom / Applied problem
Showing creators the evidence behind each suggested edit, beside the edit itself.
SlateRoom case study
Research approach
How I would study it
Proposed, not yet conducted
I am interested in combining interface prototyping, behavioral studies, and quantitative evaluation to study how evidence presentation affects understanding, trust calibration, and decisions.
Proposed study concept
Proposed A sketch of one study, to show the shape of the question. No part of it has been run, and it reports no results.
Question
Does showing the evidence behind an AI recommendation, alongside a plain statement of its uncertainty, change how well people judge when to follow it?
Possible conditions
- Recommendation only
- Recommendation with its supporting evidence
- Recommendation with evidence and a stated level of uncertainty
Possible measures
- Understanding: can participants say why the system made the recommendation?
- Confidence calibration: does their confidence track whether the recommendation was right?
- Decision quality: how often do they accept good recommendations and reject poor ones?
Confounds and limitations
- Task familiarity and domain expertise
- The base rate of correct recommendations in the task set
- Lab tasks differ from real work, where stakes and time pressure change behavior
Ethics
Informed consent, no deception about whether advice comes from a model, and no real financial or personal stakes for participants.
Preparation
Background for the work
B.A. Computer Science, Mathematics minor · M.A. Interactive Design & Game Development · Executive MBA · Experience in software engineering, product systems, and AI applications.
Contact
Interested in a research conversation?
I'd like to hear from researchers working on human-AI interaction, explainability, and decision support.
carl.brunson@gmail.com