[Paper Review] HowkGPT: Investigating the Detection of ChatGPT-generated University Student Homework through Context-Aware Perplexity Analysis

Christoforos Vasilatos, Manaar Alam+3 more|arXiv (Cornell University)|May 26, 2023

Artificial Intelligence in Healthcare and Education10 citations

TL;DR

HowkGPT uses metadata-driven perplexity thresholds computed with a pretrained GPT-2 model to distinguish ChatGPT-generated university homework from student-written work, with category-specific thresholds enhancing accuracy.

ABSTRACT

As the use of Large Language Models (LLMs) in text generation tasks proliferates, concerns arise over their potential to compromise academic integrity. The education sector currently tussles with distinguishing student-authored homework assignments from AI-generated ones. This paper addresses the challenge by introducing HowkGPT, designed to identify homework assignments generated by AI. HowkGPT is built upon a dataset of academic assignments and accompanying metadata [17] and employs a pretrained LLM to compute perplexity scores for student-authored and ChatGPT-generated responses. These scores then assist in establishing a threshold for discerning the origin of a submitted assignment. Given the specificity and contextual nature of academic work, HowkGPT further refines its analysis by defining category-specific thresholds derived from the metadata, enhancing the precision of the detection. This study emphasizes the critical need for effective strategies to uphold academic integrity amidst the growing influence of LLMs and provides an approach to ensuring fair and accurate grading in educational institutions.

Motivation & Objective

• Motivate academic integrity by distinguishing student-authored homework from AI-generated submissions.
• Leverage a metadata-rich dataset to refine perplexity-based detection for homework tasks.
• Show that category-specific perplexity thresholds outperform a single, dataset-wide threshold.
• Provide a publicly accessible web tool for real-time homework origin assessment.

Proposed method

• Compute perplexity on text using a pretrained GPT-2 model due to lack of access to GPT-3.5/4 in the study.
• Use a moving-window approach to accumulate token-level losses and exponentiate their mean to obtain text perplexity.
• Incorporate dataset metadata with knowledge and cognitive process categories to define category-specific thresholds.
• Evaluate thresholds via ROC/AUC and F1 metrics across different dataset flavors to select optimal perplexity cutoffs.
• Deploy an offline and live web application workflow for threshold computation and origin classification.

Experimental results

Research questions

• RQ1 Can perplexity distinguish student-written from AI-generated homework Texts within a university dataset?
• RQ2 Does incorporating metadata-based text categorization improve detection accuracy over a single threshold?
• RQ3 What are the optimal perplexity thresholds across different question categories and dataset flavors?

Key findings

• Perplexity-based detection accuracy improves when category-specific thresholds are used instead of a single dataset-wide threshold.
• Dataset flavorings (e.g., excluding math/code content) significantly affect the shape of perplexity distributions and threshold performance.
• ROC/AUC and F1 analyses identify different optimal thresholds depending on the chosen metric, reflecting trade-offs between precision and recall.
• The study develops a publicly accessible web platform that performs real-time perplexity-based origin assessment of submissions.