Artificial Neural Network Text Generation as a New Type of Masking the Authorship

Abstract

Artificial neural network text generation may be used to conceal the authorship of a text. As a result, experts have to be able to tell the difference between natural and generative tests to establish authorship. In this research, human-made texts and texts generated by GPT 4.5 were analyzed using the methods of compositional-semantic, structural-semantic, and grammatical-syntactic analysis. The analysis made it possible to identify and describe some typical flaws in the AI-generated texts caused by violations in their implicative and referential semantics. It revealed some current issues of language modeling and the possibilities of generative language models based on neural network algorithms. The quality of the generated text depended on additional parameters to the prompt. The article describes linguistic features typical of generated texts and provides each case with comparative examples. The prospects for further research lie in an in-depth study of complex AI-generated text structures, e.g., preloaded idiosyncratic texts. The practical result may yield a diagnostic complex of signs based on a list of typical linguistic features of AI-generated texts.

Keywords

• authorship expertise
• neural networks
• text generation
• written speech
• masking of written speech
• language models