LG AI Research: A Summary of AI-Generated Text Detection Technologies Presented at ACL 2024

As LLMs have dramatically improved text generation capabilities, the importance of AI-generated text detection (MGT) technology is growing to prevent the spread of misinformation and ethical issues.

A representative existing technology, DetectGPT, uses changes in a language model's log-likelihood that occur when text is slightly perturbed. It leverages the fact that AI-generated text exhibits a 'negative curvature' characteristic, where log-likelihood changes sharply upon minor perturbation, and has the advantage of enabling detection without any additional training.

However, recent studies point out that such perturbation-based methods have limitations in that they are vulnerable to adversarial perturbations. Accordingly, at ACL 2024, numerous studies were presented that either address the limitations of existing technologies, introduce new approaches combining classical natural language processing (NLP) insights, or propose new benchmarks for quantitatively measuring the performance of detection models.