Nima Fatemi
Jackson Burns
Leah Cressler
Dr. Daniel Gruehn (Faculty Mentor)
Project overview
AI Text Detection Study
College students increasingly rely on AI tools like ChatGPT, yet it remains unclear how well they can detect AI-generated writing or how their trust in AI content shifts with confidence and proficiency. We ran a 150-participant behavioral study to test detection accuracy, confidence, and trust patterns when students classify AI vs. human text. Our findings reveal a consistent confidence–accuracy gap, as well as a negative correlation between confidence and trust in AI-written content.
My role:
Lead UX Researcher
Designed and executed the full behavioral study, from framing the research questions to developing measures, running analysis, and synthesizing insights for academic and HAI applications.
Problem:
Students frequently interact with AI-generated text, but their ability to recognize it, their confidence in these judgments, and their trust toward AI content are poorly understood. This creates risks for digital literacy, academic integrity, and over-reliance on AI systems.
Goal:
Measure how accurately students detect AI-generated writing, and identify the psychological factors that influence confidence, AI trust, and miscalibrated judgments.
Responsibilities:
• Designed research protocol, measures, and classification tasks
• Programmed and deployed the online experiment
• Ran statistical analysis and modeled accuracy–confidence–trust relationships
• Synthesized findings into actionable insights for AI literacy and HAI design
• Co-authored poster and presented results at SNCURCS
Research Setup
A single-session, online behavioral study where each participant classified a mix of AI-generated and human-written passages, rated their confidence, their trust in the text, and completed an AI literacy survey.
Research Questions
Can first-year college students reliably distinguish AI-generated text from human-written text in short reading passages?
Does frequency of AI tool use or self-reported AI proficiency improve detection accuracy?
How are people’s confidence in their judgments and their trust in AI-generated text related to actual performance?
Which individual factors (for example AI proficiency, confidence, trust) are most predictive of accurate AI detection?
Study Design
Format
Remote, survey-based experiment completed on participants’ own laptops or phones.
Task type
Within-subject classification task. Every student saw both AI and human passages, so we could compare detection behavior at the individual level.
Independent variables (high level)
Passage type: AI-generated vs human-written
AI usage frequency
AI proficiency (subscales of the ChatGPT Literacy Scale)
Key dependent variables
Classification accuracy
Confidence in each judgment
Trust in the text
Participants
N = 150 first-year undergraduates from introductory psychology courses at NC State.
Age: 18–24 years, mean ≈ 19.7.
Mixed gender and race/ethnicity, roughly representative of the local student body.
The sample intentionally focused on heavy AI adopters in an academic context, since these students are most likely to interact with tools like ChatGPT in their coursework.
Study Design
1- Intro and consent
Participants were briefed that they would evaluate short passages and answer questions about AI usage.
2- Classification task
Each student saw 10 randomized passages:
5 SAT reading passages (human-written)
5 passages generated by ChatGPT from matched prompts
For each passage, they:
Labeled it as “Human” or “AI”
Rated their confidence on a 1–5 Likert scale
3- AI literacy and trust survey
Completed the ChatGPT Literacy Scale (25 items, 5 subscales).
Reported how often they use AI tools in their daily life, and for what purpose.
Rated how much they trust AI-generated text in various settings.
4- Debrief
Participants were debriefed about the purpose of the study and the presence of AI passages.
Measures
Behavioral
Detection accuracy
Percentage of correct classifications across the 10 passages (0–100 percent).
Confidence per judgment
Single-item 5-point rating from “very low” to “very high” for each passage.Trust in the text
Likert-style items on how much they trust that the given text is accurate and represantive of truth.
Self-reports
ChatGPT Literacy Scale (25 items) with five subscales: technical proficiency, critical evaluation, communication proficiency, creative application, and ethical competence. All showed good internal reliability.
AI usage frequency
How often participants use tools like ChatGPT in a typical week.Trust in AI text
Likert-style items on how much they trust AI-generated writing in academic work and study tasks.
My Role
Co-developed the research questions with the faculty mentor, framing the study around detection, confidence, and trust in AI-generated text.
Designed the classification task and wrote the prompts used to generate AI passages so they matched the style and difficulty of human SAT texts.
Selected and configured survey measures, including the ChatGPT Literacy Scale and trust items.
Cleaned and analyzed the data, focusing on accuracy, confidence, proficiency, and trust relationships.
Led the synthesis of findings into a clear story that highlights the confidence–accuracy–trust gap and its implications for digital literacy and assessment design.
Designed the poster layout and presented the work at the State of North Carolina Undergraduate Research and Creativity Symposium.
Results
We analyzed behavioral data from 150 students to see how well they detect AI-generated text, how confident they feel about those judgments, and how much they trust what they read.
Detection Performance
How accurately students could tell AI text from human writing.
Key Patterns
Overall, students were only slightly above chance at telling AI-generated text from human-written passages. AI usage frequency did not improve correctness, and demographic or academic variables (gender, race, major, age, socioeconomic status) showed no clear relationship with performance.
Takeaway
Baseline detection is weak, and no obvious user segment is naturally better at spotting AI text.
Confidence Calibration
How students’ confidence lined up with actual performance.
Key Patterns
Confidence did not predict correctness. Highly confident judgments were often wrong, and low-confidence judgments were sometimes right. Higher AI proficiency increased confidence, but did not improve detection accuracy.
Takeaway
Users can feel very sure and still be wrong, especially as their AI skill grows.
AI Usage and Proficiency
How using AI tools relates to skill and behavior.
Key Patterns
Frequent AI users showed significantly higher AI proficiency scores, but they were not better at detecting AI text than infrequent users.
Takeaway
More hands-on AI experience creates power users, not necessarily better critical judges.
Confidence and Trust in Each Passage
How confidence in a judgment relates to trust in that specific text.
Key Patterns
After each passage, students rated how much they trusted the text to be accurate. Across passages, higher confidence in one’s classification was linked to lower trust in the passage itself. When students felt sure about whether something was AI or human, they tended to trust its content less.
Takeaway
As confidence in “who wrote this” goes up, trust in “is this true” can go down.