ConvNLP: Image-based AI Text Detection
Project Overview
Generative AI, especially through Large Language Models (LLMs), presents significant opportunities for enhancing education by facilitating personalized learning experiences, generating content, and evaluating assignments. Despite these advantages, the rise of AI tools raises ethical concerns such as academic dishonesty and plagiarism, as students may become overly dependent on LLMs for their work. To combat these issues, the document proposes an innovative detection method that utilizes image representations of word embeddings alongside a convolutional neural network, known as ZigZag ResNet, designed to identify AI-generated text. This approach aims to bolster academic integrity and trust in student submissions while addressing the limitations of current detection techniques. Overall, while generative AI holds promise for transforming educational practices, it simultaneously challenges traditional notions of authorship and accountability, necessitating new strategies to ensure ethical use.
Key Applications
ZigZag ResNet
Context: Educational institutions combating academic dishonesty and ensuring integrity in student work.
Implementation: Developed a novel convolutional neural network architecture that uses visual representations of text embeddings for detecting AI-generated content.
Outcomes: Achieved an average detection rate of 88.35% across various datasets of AI-generated text, offering computational efficiency and improved generalization over traditional methods.
Challenges: Existing detection tools are often compute-intensive, lack generalization, and can produce false positives, especially against non-native English writers.
Implementation Barriers
Ethical
Concerns regarding academic dishonesty and the potential for misuse of generative AI technologies in completing assignments.
Proposed Solutions: Develop robust methods for detecting AI-generated text to preserve academic integrity.
Technical
The need for vast amounts of data and computing power for training detection models, as well as challenges in existing tools and models struggling to generalize across different datasets and types of AI-generated text.
Proposed Solutions: Utilizing image processing models for text classification, which require fewer resources than traditional NLP models, and developing hybrid models that combine various detection techniques, such as the proposed ZigZag ResNet.
Project Team
Suriya Prakash Jambunathan
Researcher
Ashwath Shankarnarayan
Researcher
Parijat Dube
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Suriya Prakash Jambunathan, Ashwath Shankarnarayan, Parijat Dube
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai