How well can machine-generated texts be identified and can language models be trained to avoid identification?

TL;DR

This paper evaluates the effectiveness of machine learning classifiers in detecting AI-generated texts and explores how generative models can be trained to evade such detection, revealing significant challenges in distinguishing human from machine texts.

Contribution

It introduces a reinforcement learning method to refine generative models, enabling them to evade current detection classifiers with high success.

Findings

Shallow classifiers achieve 0.6-0.8 accuracy in detection.

Transformer classifiers have over 0.9 accuracy.

Refined models can evade detection with accuracy below 0.15.

Abstract

With the rise of generative pre-trained transformer models such as GPT-3, GPT-NeoX, or OPT, distinguishing human-generated texts from machine-generated ones has become important. We refined five separate language models to generate synthetic tweets, uncovering that shallow learning classification algorithms, like Naive Bayes, achieve detection accuracy between 0.6 and 0.8. Shallow learning classifiers differ from human-based detection, especially when using higher temperature values during text generation, resulting in a lower detection rate. Humans prioritize linguistic acceptability, which tends to be higher at lower temperature values. In contrast, transformer-based classifiers have an accuracy of 0.9 and above. We found that using a reinforcement learning approach to refine our generative models can successfully evade BERT-based classifiers with a detection accuracy of 0.15 or less.