Identifying Key Parts of Speech That Differentiate Human and AI-Generated Text Using Eye-Tracking Data
Published:
Project Overview
This project explores the linguistic cues that subconsciously influence how humans perceive text authored by either a human or an AI model such as ChatGPT. By leveraging eye-tracking data, we identify patterns in visual attention and correlate them with specific parts of speech in the text.
Methodology
We used the Tobii Pro eye tracker to record participant gaze behavior as they read various text samples — some written by humans, others generated by AI. From this, we extracted key features including:
- Total Fixation Count
- Fixation Duration
- Scanpath Sequence
These were analyzed using the Velocity Threshold Identification (I-VT) algorithm at two threshold settings: 30°/s and 100°/s, to explore both coarse and fine-grained reading behavior.
Key Findings
- For human-written text, participants primarily fixated on nouns, indicating focus on factual content and subjects.
- For AI-generated text, participants showed more fixation on adjectives, reflecting attention to stylistic or exaggerated elements commonly used by language models to enhance fluency or appeal.
Conclusion
Our study highlights consistent behavioral patterns that reflect cognitive processing differences between human and AI-generated writing. The insights could be valuable for:
- Detecting AI-authored content
- Improving readability of generated text
- Understanding human trust in machine-written communication
This research contributes to the emerging field of explainable AI perception and computational psycholinguistics.