TL;DR
This paper introduces a novel approach to detect AI-generated text in academic writing by analyzing sentence-level similarities between human and machine responses. By leveraging paraphrasing and cosine similarity, the method offers a more transparent and adaptable plagiarism detection system that moves beyond opaque "black box" models.
What Problem Does This Paper Address?
The rapid advancement and widespread adoption of Large Language Models (LLMs) like ChatGPT have introduced significant challenges to academic integrity. While these models offer powerful capabilities for text generation, they also facilitate plagiarism, making it increasingly difficult for educators to distinguish between human-written and AI-generated content. Existing AI-generated text classifiers often suffer from limited accuracy, high false positive rates, and a lack of transparency, making their predictions difficult for human evaluators to understand and trust. Furthermore, many current tools are specific to certain LLMs and quickly become outdated as new models emerge, necessitating constant retraining.
Methods
The proposed method employs a multi-faceted approach to address the limitations of existing plagiarism detection tools:
- Paraphrasing Model: To account for the diverse ways students might query an LLM, the authors use a T5 language model to generate multiple paraphrased versions of an original question. This ensures the LLM's responses are generated under varied prompts, mimicking real-world student query diversity.
- LLM Integration: These paraphrased questions are then fed into a state-of-the-art LLM (specifically, GPT-3.5-turbo) to generate multiple AI-written answers.
- Sentence-Level Comparison: Each AI-generated answer is broken down into individual sentences. The core of the detection lies in comparing these AI-generated sentences with sentences from a student's response. This is achieved by:
- Embedding Generation: Using
text-embedding-ada-002to create vector representations of sentences. - Cosine Similarity: Calculating the cosine similarity between the embeddings of AI-generated sentences and student response sentences. This quantifies the semantic and syntactic similarity.
- Human-Machine (HM) and Machine-Machine (MM) Comparisons: The paper distinguishes between comparing human-generated text to machine-generated text (HM) and comparing two machine-generated texts (MM).
- Linear Discriminant Analysis (LDA): The cosine similarity scores, along with their corresponding labels (human or AI-generated), are used to train an LDA classifier. This supervised learning model predicts the probability of a sentence being AI-generated.
- Document-Level Analysis: For a broader assessment, the document-level similarity is calculated by averaging the highest cosine similarity scores from sentence-level comparisons within a response. This aggregated score is also used for LDA classification.
Datasets
- Human ChatGPT Comparison Corpus (HC3): This dataset, containing approximately 10,000 questions and corresponding answers from both human experts and ChatGPT, was used for sentence-level and document-level evaluation. It covers various domains like computer science, finance, medicine, and law.
- GPT-wiki-intro dataset: This dataset was used for document-level evaluation and comparison with other models. It includes GPT-2 generated introductions and human-written introductions from Wikipedia articles.
Key Insights and Ideas
- Transparency through Quantifiable Metrics: By breaking down the analysis to the sentence level and using cosine similarity, the method provides quantifiable metrics that are easier for human evaluators to interpret, moving away from opaque "black box" predictions.
- Adaptability to Evolving LLMs: The paraphrasing approach and the focus on NLP techniques rather than specific model architectures make the system inherently adaptable to new LLMs and their upgrades, reducing the need for constant retraining.
- Robustness via Paraphrasing: Generating multiple paraphrased questions for LLM input helps to capture a wider range of potential AI-generated text variations, making the detection more robust.
- Sentence-Level Granularity: The ability to detect plagiarism at the sentence level allows for the identification of subtle instances that might be missed by document-level comparisons alone.
Key Results
The proposed approach achieved 94% accuracy in classifying human and AI-generated answers on the HC3 dataset at the sentence level. The paper also reports strong performance at the document level, with an LDA classifier achieving an accuracy of 0.94 with a best threshold of 0.66. When compared against SOTA models like RoBERTa-QA and OpenAI's text classifier on the GPT-wiki-intro dataset, their model (specifically "Our Model-B" which uses paraphrasing) demonstrated competitive or superior performance in terms of F1 score, precision, and recall.
Limitations and Future Work
- Limitations:
- The paper notes that while their model shows no improvement with paraphrasing on the GPT-wiki-intro dataset, they plan to investigate other approaches to improve model performance.
- The current approach might still suffer from high false positive rates, as acknowledged in the context of existing methods.
- Future Work:
- Collecting additional unbiased datasets for evaluation.
- Exploring the incorporation of different algorithms at the sentence level.
- Employing stylometry techniques to identify individual student writing styles for more personalized plagiarism detection.
- Investigating other potential approaches to improve model performance, such as varying LLM temperature and P values.
Summary updated