The False Positive Problem: A Case Study on GPTZero and AI Detection

Jacob Gonzales

Introduction

In this case study, I explore the high rate of false positives generated by current AI detection software, with a specific focus on GPTZero. After a healthcare professional brought this issue to my attention, I was motivated to look beyond simple bypass methods and investigate the practical usability of these tools.

What is GPTZero?

GPTZero identifies AI-generated text by analyzing its complexity and “perplexity”—essentially looking for the unique statistical fingerprints left by large language models.

The False Positives

During my testing, I found that GPTZero yields highly mixed results. While OpenAI’s AI Classifier tends to be more consistent, it offers less detail about which parts of a text are human-written versus AI-generated.

To test GPTZero’s accuracy, I submitted a 2020 medical paper on severe COVID-19 outcomes, authored by multiple CDC contributors. Despite being written well before the widespread use of generative AI, GPTZero flagged more than 50% of the first paragraph as AI-written. Even when testing the maximum allowed 5,000 characters, the false positive rate remained above 50%.

Testing Medical Literature

To further explore this, I tested the abstracts of 20 different neurology papers. The results were shocking: GPTZero flagged 11 out of the 20 human-written abstracts as having AI-generated portions.

The Control: OpenAI’s AI Classifier

Using OpenAI’s AI Classifier as a baseline, I tested the same 2020 COVID-19 paper. The classifier labeled the text as “unclear,” which, while not entirely accurate, is a safer conclusion than a false positive. Given that the paper was published in 2020—and verified as unchanged via the Internet Archive’s Wayback Machine—we can be 100% certain it was written without AI assistance.

The Problem with AI Detection

The commercialization of inaccurate AI detection software poses significant risks, particularly in education. If a student is wrongfully accused of plagiarism based on a flawed algorithm, the consequences can be severe. At my university, academic dishonesty is taken very seriously, and false accusations can have a lasting impact on a student’s record. Until these tools become more reliable, their use in high-stakes environments should be heavily scrutinized.

Appendix: Tested Papers Data

False Positives (Flagged as AI-generated):

True Negatives (Correctly identified as Human-written):