Feature-based detection of automated language models: tackling GPT-2, GPT-3 and Grover

Abstract

The recent improvements of language models have drawn much attention to potential cases of use and abuse of automatically generated text. Great effort is put into the development of methods to detect machine generations among human-written text in order to avoid scenarios in which the large-scale generation of text with minimal cost and effort undermines the trust in human interaction and factual information online. While most of the current approaches rely on the availability of expensive language models, we propose a simple feature-based classifier for the detection problem, using carefully crafted features that attempt to model intrinsic differences between human and machine text. Our research contributes to the field in producing a detection method that achieves performance competitive with far more expensive methods, offering an accessible “first line-of-defense” against the abuse of language models. Furthermore, our experiments show that different sampling methods lead to different types of flaws in generated text.