Generative AI Text Detection: Strengths and Weaknesses
School Name
South Carolina Governor's School for Science and Mathematics
Grade Level
12th Grade
Presentation Topic
Computer Science
Presentation Type
Mentored
Abstract
The advancement of AI since 2022 has led to increased usage in all aspects of the world around us. This introduces a new era of online threats due to misinformation. Using AI, bad actors can easily generate vast amounts of believable misinformation which can be used to manipulate public opinion. In this study, we evaluate various AI text detection models, along with circumvention techniques such as DFT fooler, complex paraphrasing, and humanizers which modify AI-generated text to circumvent detectors. We found that even advanced detection models such as GPTZero and ZeroGPT used by top universities were weak when challenged by DFT Fooler or humanizer models. While current detection methods are effective against simple texts, they need much improvement to face the challenges of real-world applications.
Recommended Citation
Bansal, Arpan, "Generative AI Text Detection: Strengths and Weaknesses" (2025). South Carolina Junior Academy of Science. 46.
https://scholarexchange.furman.edu/scjas/2025/all/46
Location
PENNY 216
Start Date
4-5-2025 9:30 AM
Presentation Format
Oral Only
Group Project
No
Generative AI Text Detection: Strengths and Weaknesses
PENNY 216
The advancement of AI since 2022 has led to increased usage in all aspects of the world around us. This introduces a new era of online threats due to misinformation. Using AI, bad actors can easily generate vast amounts of believable misinformation which can be used to manipulate public opinion. In this study, we evaluate various AI text detection models, along with circumvention techniques such as DFT fooler, complex paraphrasing, and humanizers which modify AI-generated text to circumvent detectors. We found that even advanced detection models such as GPTZero and ZeroGPT used by top universities were weak when challenged by DFT Fooler or humanizer models. While current detection methods are effective against simple texts, they need much improvement to face the challenges of real-world applications.
